| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "bytes" |
| 5 | "context" |
| 6 | "encoding/json" |
| 7 | "errors" |
| 8 | "fmt" |
| 9 | "strings" |
| 10 | "sync" |
| 11 | "sync/atomic" |
| 12 | "time" |
| 13 | "unicode/utf8" |
| 14 | |
| 15 | "mvdan.cc/sh/v3/syntax" |
| 16 | |
| 17 | "reasonix/internal/ablation" |
| 18 | "reasonix/internal/capability" |
| 19 | "reasonix/internal/checkpoint" |
| 20 | "reasonix/internal/diff" |
| 21 | "reasonix/internal/event" |
| 22 | "reasonix/internal/evidence" |
| 23 | "reasonix/internal/extension/dispatch" |
| 24 | "reasonix/internal/instruction" |
| 25 | "reasonix/internal/jobs" |
| 26 | "reasonix/internal/memory" |
| 27 | "reasonix/internal/nilutil" |
| 28 | "reasonix/internal/permission" |
| 29 | "reasonix/internal/planmode" |
| 30 | "reasonix/internal/provider" |
| 31 | "reasonix/internal/sandbox" |
| 32 | "reasonix/internal/shellparse" |
| 33 | "reasonix/internal/tool" |
| 34 | "reasonix/internal/workspacelease" |
| 35 | ) |
| 36 | |
| 37 | // maxToolOutputBytes caps a single tool result before it goes into the model's |
| 38 | // context. ~32KB is roughly 8K tokens — enough for a full file read or a busy |
| 39 | // grep, while preventing one accidental "read this 5 MB log" from blowing the |
| 40 | // window before the next compaction runs. |
| 41 | const maxToolOutputBytes = 32 * 1024 |
| 42 | |
| 43 | const maxEmptyFinalBlocks = 3 |
| 44 | |
| 45 | // maxStreamRecoveries is the number of body-phase stream retries after the |
| 46 | // initial sampling attempt (Codex-aligned default: 1 + 5 = 6 attempts total). |
| 47 | const maxStreamRecoveries = 5 |
| 48 | const maxSamplingAttempts = maxStreamRecoveries + 1 |
| 49 | const maxExecutorHandoffNudges = 1 |
| 50 | |
| 51 | const defaultReasoningByteLimit = 128 * 1024 |
| 52 | |
| 53 | const finishReasonClientReasoningLimit = "client_reasoning_limit" |
| 54 | |
| 55 | var errReasoningByteLimitExceeded = errors.New("reasoning output exceeded client byte limit") |
| 56 | |
| 57 | // DeliveryRuntimeMarker is the delivery-mode contract block appended to user |
| 58 | // turns (withTurnPreferences). Exported as the single source of truth for the |
| 59 | // byte-exact suffix strip in preview derivation and for cross-package tests; |
| 60 | // its text is cache-frozen — changing it breaks steer replay matching and the |
| 61 | // prefix stability of every live delivery session. |
| 62 | const DeliveryRuntimeMarker = `<delivery-runtime> |
| 63 | This session is in delivery-first mode. Before any state-changing tool call, |
| 64 | establish concrete, verifiable acceptance criteria with todo_write. After the |
| 65 | change, inspect the result, run relevant verification, and sign off each step |
| 66 | with complete_step citing the successful verification command. The host enforces |
| 67 | these gates and will reject mutation or finalization when evidence is missing. |
| 68 | </delivery-runtime>` |
| 69 | |
| 70 | // Renderer redraws the assistant's final-answer text as styled output. It is |
| 71 | // applied only after a turn's text stream completes, so the user sees raw |
| 72 | // markdown stream live, then a single redraw replaces it with formatted |
| 73 | // output. The renderer is intentionally interface-shaped so the agent stays |
| 74 | // independent of the cli's markdown library choice. Consumed by TextSink. |
| 75 | type Renderer interface { |
| 76 | Render(text string) string |
| 77 | } |
| 78 | |
| 79 | // Asker puts structured multiple-choice questions to the user and blocks for the |
| 80 | // answers. The agent consults it for the `ask` tool. It is interface-shaped so |
| 81 | // the agent stays independent of the frontend; a nil asker means no interactive |
| 82 | // user (headless runs), where `ask` returns a "decide for yourself" result. The |
| 83 | // interactive frontends wire the controller in as the Asker. |
| 84 | type Asker interface { |
| 85 | Ask(ctx context.Context, questions []event.AskQuestion) ([]event.AskAnswer, error) |
| 86 | } |
| 87 | |
| 88 | // callContextKey carries the executing tool call's identity into Execute. |
| 89 | type callContextKey struct{} |
| 90 | type parentSessionContextKey struct{} |
| 91 | type subagentDepthContextKey struct{} |
| 92 | type userImagesContextKey struct{} |
| 93 | |
| 94 | // callContext is the per-call context a tool can read. parentID is the call being |
| 95 | // executed and sink is the agent's event sink (the `task` tool uses both to nest |
| 96 | // a sub-agent's events under this call); asker lets the `ask` tool reach the user. |
| 97 | type callContext struct { |
| 98 | parentID string |
| 99 | sink event.Sink |
| 100 | asker Asker |
| 101 | planMode bool |
| 102 | } |
| 103 | |
| 104 | // withCallContext stamps ctx with the executing call's ID, the agent's sink, and |
| 105 | // the asker. executeOne sets this before every Execute; `task` reads it (via |
| 106 | // CallContext) to nest sub-agent events, and `ask` reads the asker to prompt. |
| 107 | // The plan-mode flag is mirrored onto the leaf planmode key so tools that must |
| 108 | // not import this package (for example internal/tool/builtin) can still read it. |
| 109 | func withCallContext(ctx context.Context, parentID string, sink event.Sink, asker Asker, planMode bool) context.Context { |
| 110 | ctx = planmode.WithActive(ctx, planMode) |
| 111 | return context.WithValue(ctx, callContextKey{}, callContext{parentID: parentID, sink: sink, asker: asker, planMode: planMode}) |
| 112 | } |
| 113 | |
| 114 | // WithToolCallContext stamps ctx as a host-initiated top-level tool call. |
| 115 | // Normal model-selected tools receive this context from executeOne; controller |
| 116 | // entry points that deliberately invoke the same tool machinery (for example a |
| 117 | // user typing /<subagent-skill>) use this exported wrapper so nested sub-agent |
| 118 | // activity still reaches the parent event stream and plan-mode policy remains |
| 119 | // visible to the invoked runner. |
| 120 | func WithToolCallContext(ctx context.Context, parentID string, sink event.Sink, asker Asker, planMode bool) context.Context { |
| 121 | return withCallContext(ctx, parentID, sink, asker, planMode) |
| 122 | } |
| 123 | |
| 124 | // CallContext returns the executing call's ID, the agent's sink, and the asker, |
| 125 | // if the context was set by an agent's executeOne. ok is false for a plain |
| 126 | // context (headless tool tests, calls made outside the run loop). |
| 127 | func CallContext(ctx context.Context) (parentID string, sink event.Sink, asker Asker, ok bool) { |
| 128 | cc, ok := ctx.Value(callContextKey{}).(callContext) |
| 129 | if !ok { |
| 130 | return "", nil, nil, false |
| 131 | } |
| 132 | return cc.parentID, cc.sink, cc.asker, true |
| 133 | } |
| 134 | |
| 135 | // PlanModeFromContext reports whether the tool call is executing during the |
| 136 | // plan-first workflow. Tools may use it for phase-specific behavior, but it is |
| 137 | // not a permission or read-only boundary. |
| 138 | func PlanModeFromContext(ctx context.Context) bool { |
| 139 | cc, ok := ctx.Value(callContextKey{}).(callContext) |
| 140 | return ok && cc.planMode |
| 141 | } |
| 142 | |
| 143 | // WithParentSession stamps the active parent session ID onto a turn context so |
| 144 | // persisted sub-agents can record and enforce their owning conversation. |
| 145 | func WithParentSession(ctx context.Context, parentSession string) context.Context { |
| 146 | return context.WithValue(ctx, parentSessionContextKey{}, strings.TrimSpace(parentSession)) |
| 147 | } |
| 148 | |
| 149 | // ParentSession returns the active parent session ID carried by a turn context. |
| 150 | func ParentSession(ctx context.Context) string { |
| 151 | parentSession, _ := ctx.Value(parentSessionContextKey{}).(string) |
| 152 | return strings.TrimSpace(parentSession) |
| 153 | } |
| 154 | |
| 155 | // WithSubagentDepth carries the current subagent depth through nested tool calls. |
| 156 | // The root agent runs at depth 0; each spawned subagent increments by one. |
| 157 | func WithSubagentDepth(ctx context.Context, depth int) context.Context { |
| 158 | if depth < 0 { |
| 159 | depth = 0 |
| 160 | } |
| 161 | return context.WithValue(ctx, subagentDepthContextKey{}, depth) |
| 162 | } |
| 163 | |
| 164 | // SubagentDepth returns the current subagent depth carried by a turn context. |
| 165 | func SubagentDepth(ctx context.Context) int { |
| 166 | depth, _ := ctx.Value(subagentDepthContextKey{}).(int) |
| 167 | if depth < 0 { |
| 168 | return 0 |
| 169 | } |
| 170 | return depth |
| 171 | } |
| 172 | |
| 173 | // WithUserImages carries the data URLs of images the user attached to this turn, |
| 174 | // resolved by the controller (which owns attachments) since the agent must not |
| 175 | // depend on it. Run embeds them on the user message; the provider sends them only |
| 176 | // when the model is vision-capable. |
| 177 | func WithUserImages(ctx context.Context, images []string) context.Context { |
| 178 | return context.WithValue(ctx, userImagesContextKey{}, images) |
| 179 | } |
| 180 | |
| 181 | func userImages(ctx context.Context) []string { |
| 182 | images, _ := ctx.Value(userImagesContextKey{}).([]string) |
| 183 | return images |
| 184 | } |
| 185 | |
| 186 | // Gate decides, per tool call, whether it may run. The agent consults it at |
| 187 | // execute time after any explicit planning-phase opt-out. It is interface-shaped so the agent |
| 188 | // stays independent of the permission package and of how "ask" is resolved |
| 189 | // (silently in headless runs, interactively in the chat TUI). A nil gate means |
| 190 | // no gating — every call runs, preserving behaviour for callers that don't wire |
| 191 | // one in. reason is fed back to the model when allow is false; a non-nil err |
| 192 | // (e.g. ctx cancelled awaiting approval) is treated as a block for that call. |
| 193 | type Gate interface { |
| 194 | Check(ctx context.Context, toolName string, args json.RawMessage, readOnly bool) (allow bool, reason string, err error) |
| 195 | } |
| 196 | |
| 197 | // ExplicitDenyGate exposes the only global permission decision that applies to |
| 198 | // an already-authorized MCP server. Installing or approving a server is the |
| 199 | // user's authorization boundary; ordinary ask/fallback posture must not add a |
| 200 | // second per-call prompt, while explicit deny rules remain authoritative. |
| 201 | type ExplicitDenyGate interface { |
| 202 | ExplicitlyDenies(toolName string, args json.RawMessage) bool |
| 203 | } |
| 204 | |
| 205 | const PlanModeReadOnlyCommandApprovalTool = "plan_mode_read_only_command" |
| 206 | |
| 207 | // PlanModeReadOnlyTrustRequest describes a bash command that is safe enough to |
| 208 | // ask the user to accept as read-only during planning. Command is the concrete |
| 209 | // attempted command and Prefix is the reusable prefix to trust. |
| 210 | type PlanModeReadOnlyTrustRequest struct { |
| 211 | ToolName string |
| 212 | Command string |
| 213 | Prefix string |
| 214 | Args json.RawMessage |
| 215 | } |
| 216 | |
| 217 | // PlanModeReadOnlyTrustGate is the legacy Plan bash trust bridge. It remains in |
| 218 | // the internal API for controller compatibility, but ordinary Plan execution no |
| 219 | // longer invokes it; bash calls use the normal permission gate. |
| 220 | type PlanModeReadOnlyTrustGate interface { |
| 221 | CheckPlanModeReadOnlyTrust(ctx context.Context, req PlanModeReadOnlyTrustRequest) (allow bool, reason string, err error) |
| 222 | } |
| 223 | |
| 224 | const DefaultMaxSubagentDepth = 2 |
| 225 | |
| 226 | // NormalizeMaxSubagentDepth applies the public config contract: values below 1 |
| 227 | // preserve the old single-delegation boundary. |
| 228 | func NormalizeMaxSubagentDepth(depth int) int { |
| 229 | if depth < 1 { |
| 230 | return 1 |
| 231 | } |
| 232 | return depth |
| 233 | } |
| 234 | |
| 235 | // ToolHooks fires user-configured shell hooks around each tool call. PreToolUse |
| 236 | // runs before the call and may block it (block=true; message is the reason fed |
| 237 | // back to the model); PostToolUse runs after and only surfaces output to the |
| 238 | // user (it can't block). It is interface-shaped so the agent stays independent |
| 239 | // of the hook package — a nil hooks field disables hook firing entirely. |
| 240 | type ToolHooks interface { |
| 241 | PreToolUse(ctx context.Context, name string, args json.RawMessage) (block bool, message string) |
| 242 | PostToolUse(ctx context.Context, name string, args json.RawMessage, result string) |
| 243 | PostToolUseFailure(ctx context.Context, name string, args json.RawMessage, result string, err error) |
| 244 | // PostLLMCall fires after each model turn completes (streaming finishes) |
| 245 | // but before reasoning_content is stored. It returns the (possibly |
| 246 | // translated) reasoning string — the original when no hook is configured. |
| 247 | // HasPostLLMCall reports whether such a hook exists, so the agent keeps |
| 248 | // streaming reasoning live when none is wired up. |
| 249 | PostLLMCall(ctx context.Context, reasoning string, turn int) string |
| 250 | HasPostLLMCall() bool |
| 251 | // SubagentStop fires when a `task` sub-agent finishes (foreground). PreCompact |
| 252 | // fires just before a compaction pass and returns extra summary guidance (its |
| 253 | // hooks' stdout) to fold into the summary prompt; "" when no hook contributes. |
| 254 | SubagentStop(ctx context.Context, last string) |
| 255 | PreCompact(ctx context.Context, trigger string) string |
| 256 | } |
| 257 | |
| 258 | // Agent drives a single task: a Provider, a tool Registry, and a Session wired |
| 259 | // into the main loop. |
| 260 | type Agent struct { |
| 261 | prov provider.Provider |
| 262 | tools *tool.Registry |
| 263 | session *Session |
| 264 | sessMu sync.Mutex // guards the session pointer for external Session()/SetSession |
| 265 | maxSteps int |
| 266 | maxStepsKey string |
| 267 | reasoningByteLimit int |
| 268 | maxOutputTokens int |
| 269 | // executorHandoffGuard is enabled by Coordinator for the executor agent. The |
| 270 | // per-turn marker check in Run keeps ordinary single-model turns unaffected. |
| 271 | executorHandoffGuard bool |
| 272 | temperature float64 |
| 273 | pricing *provider.Pricing |
| 274 | usageSource string |
| 275 | modelRef string |
| 276 | responseLanguage atomic.Value // string: auto|zh|en |
| 277 | reasoningLanguage atomic.Value // string: auto|zh|en |
| 278 | |
| 279 | // sink receives the turn's typed event stream (reasoning/text deltas, tool |
| 280 | // dispatch/results, usage, notices). The agent no longer formats output |
| 281 | // itself — a frontend's Sink decides how to render. Never nil; New defaults |
| 282 | // it to event.Discard. |
| 283 | sink event.Sink |
| 284 | |
| 285 | // lastUsage caches the most recent per-turn telemetry the provider reported so |
| 286 | // the CLI can expose a context gauge without re-scraping the usage line. The |
| 287 | // run loop writes it while a frontend's status line reads it, so it is atomic. |
| 288 | lastUsage atomic.Pointer[provider.Usage] |
| 289 | |
| 290 | // sessCacheHit/sessCacheMiss accumulate cache tokens across every API call |
| 291 | // this session, so frontends can show the aggregate hit-rate (Σhit/Σ(hit+miss)) |
| 292 | // — a steadier, cost-oriented number than the single-turn rate. They are NOT |
| 293 | // reset on compaction (compaction only rewrites session.Messages), so the |
| 294 | // aggregate never craters when the prefix is summarized away. Atomic: the run |
| 295 | // loop accumulates them while the status line reads them. |
| 296 | sessCacheHit atomic.Int64 |
| 297 | sessCacheMiss atomic.Int64 |
| 298 | |
| 299 | // lastPrefixShape records the previous provider request's cacheable prefix |
| 300 | // so usage events can explain prefix churn on the next request. |
| 301 | lastPrefixShape PrefixShape |
| 302 | haveLastPrefixShape bool |
| 303 | |
| 304 | // warnedMissingToolCallReasoning marks one active missing-reasoning incident |
| 305 | // within this agent. The legacy name is retained because the persisted state |
| 306 | // predates silent recovery; it now gates one automatic retry rather than a |
| 307 | // user-visible warning. A healthy tool-call turn clears it. Loop-owned; |
| 308 | // reset by SetSession. |
| 309 | warnedMissingToolCallReasoning bool |
| 310 | // missingReasoningWarnStateChecked avoids a file transaction on every |
| 311 | // healthy tool-call turn. It resets with the session so a new Agent can |
| 312 | // continue or confirm an incident persisted by an earlier process. |
| 313 | missingReasoningWarnStateChecked bool |
| 314 | // missingReasoningHealthyStreak provides the same three-turn anti-flapping |
| 315 | // policy when no cross-process state directory is configured. |
| 316 | missingReasoningHealthyStreak int |
| 317 | // missingReasoningWarnPendingResolveAt keeps a healthy observation retryable |
| 318 | // when its state write fails. The next missing turn retries that watermark |
| 319 | // before consulting the persisted incident and otherwise fails visible. |
| 320 | missingReasoningWarnPendingResolveAt time.Time |
| 321 | |
| 322 | // missingReasoningWarnState rate-limits recovery retries across sessions and |
| 323 | // processes by an opaque provider-configuration fingerprint (#7059). The |
| 324 | // legacy type/file names preserve the on-disk v2 contract. nil (no dir in |
| 325 | // Options) keeps in-memory active-incident gating only. |
| 326 | missingReasoningWarnState *missingReasoningWarnState |
| 327 | |
| 328 | // planMode enables planning workflow instructions and explicit phase opt-outs. |
| 329 | // It does not replace the permission or sandbox boundary. The system prompt and |
| 330 | // tool list never change with the toggle, preserving the provider-cache prefix. |
| 331 | planMode atomic.Bool |
| 332 | |
| 333 | // readOnlyExecution is a construction-time defense for planner/research |
| 334 | // agents. Unlike planMode it is not a collaboration toggle: it remains on |
| 335 | // for the agent's lifetime and validates proxy calls after resolution. |
| 336 | readOnlyExecution bool |
| 337 | |
| 338 | // mutationDependencyBarrier is set for the remainder of a provider tool |
| 339 | // batch after any mutating call fails or is blocked. executeOne re-checks |
| 340 | // it after proxy resolution so use_capability cannot bypass the barrier by |
| 341 | // advertising schema-level ReadOnly()==true. Parallel read-only segments |
| 342 | // never set it. Cleared at the start of each executeBatch. |
| 343 | mutationDependencyBarrier atomic.Bool |
| 344 | |
| 345 | // plannerMCPExecution relaxes the strict read-only MCP boundary for the |
| 346 | // two-model Planner only: authorized, non-destructive MCP targets may run |
| 347 | // through use_capability even without readOnlyHint. Ordinary writers, bash, |
| 348 | // and destructive MCP stay blocked. Strict read-only sub-agents leave this |
| 349 | // false and still require readOnlyHint. |
| 350 | plannerMCPExecution bool |
| 351 | |
| 352 | // gate, when non-nil, is the per-call permission gate for both standard and |
| 353 | // Plan workflows. nil disables gating entirely. |
| 354 | gate Gate |
| 355 | |
| 356 | // extensions, when non-nil, is the frozen Extension Protocol v1 dispatcher |
| 357 | // for this controller generation. The run loop consults it at the |
| 358 | // agent-side intercept points (see extensions.go); nil means no v1 runtime |
| 359 | // packages are installed and every point passes through byte-identically. |
| 360 | extensions *dispatch.Dispatcher |
| 361 | |
| 362 | // recoveryGate, when non-nil, is the Auto Guard boundary for Auto mode. |
| 363 | // Shared by root and sub-agents for the same controller task. nil disables |
| 364 | // recovery checks (Ask/YOLO, headless without wiring, or feature off). |
| 365 | recoveryGate RecoveryGate |
| 366 | // recoveryAgentID labels this agent on recovery cards (empty = root). |
| 367 | recoveryAgentID string |
| 368 | // recoveryTaskID isolates recovery state across concurrent top-level tasks. |
| 369 | // Empty shares the root task bucket. |
| 370 | recoveryTaskID string |
| 371 | // recoveryTaskSummary is the bounded task text for this Agent.Run. It lets a |
| 372 | // shared recovery gate review sub-agent mutations against the child task, |
| 373 | // rather than the root controller transcript. |
| 374 | recoveryTaskSummary string |
| 375 | // recoveryRunSeq gives ordinary (non-goal) runs a collision-free host scope. |
| 376 | // Goal runs use their stable delivery scope instead. |
| 377 | recoveryRunSeq atomic.Uint64 |
| 378 | |
| 379 | // planModeReadOnlyTrust is retained for legacy controller wiring. The main |
| 380 | // Plan execution path no longer consults it. |
| 381 | planModeReadOnlyTrust PlanModeReadOnlyTrustGate |
| 382 | |
| 383 | // sandboxEscapeApprover, when non-nil, can ask the user whether one shell |
| 384 | // command may rerun unconfined after the OS sandbox failed to start. |
| 385 | sandboxEscapeApprover sandbox.EscapeApprover |
| 386 | |
| 387 | // configWriteApprover, when non-nil, can ask the user whether a file tool |
| 388 | // may write a Reasonix-managed config file outside the workspace roots. |
| 389 | configWriteApprover tool.ConfigWriteApprover |
| 390 | |
| 391 | // hooks, when non-nil, fires PreToolUse / PostToolUse shell hooks around each |
| 392 | // tool call. nil disables hook firing. |
| 393 | hooks ToolHooks |
| 394 | |
| 395 | // asker, when non-nil, lets the `ask` tool put questions to the user. nil in |
| 396 | // headless runs (no interactive user). Set via SetAsker. |
| 397 | asker Asker |
| 398 | |
| 399 | // onPreEdit, when non-nil, is called with a writer tool's previewed change |
| 400 | // just before it runs — the seam the checkpoint store uses to snapshot a |
| 401 | // file's pre-edit content. Only fires for non-ReadOnly tools that implement |
| 402 | // tool.Previewer (so bash, whose targets are unknowable, is never tracked). |
| 403 | // Set via SetPreEditHook. Prefer mutationObserver when both are set. |
| 404 | onPreEdit func(diff.Change) |
| 405 | |
| 406 | // mutationObserver is the host-side unified file mutation observer. It |
| 407 | // captures preimages before tools run and after-fingerprints regardless of |
| 408 | // success/failure. Passed through Options to sub-agents; never changes |
| 409 | // provider-visible tool schemas or prompts. |
| 410 | mutationObserver *checkpoint.MutationObserver |
| 411 | |
| 412 | // jobs, when non-nil, is the session's background-job manager. executeOne |
| 413 | // stamps it onto each tool call's context so the background tools (bash |
| 414 | // run_in_background, task run_in_background, bash_output/kill_shell/wait) can |
| 415 | // reach it. nil leaves those tools to degrade gracefully. |
| 416 | jobs *jobs.Manager |
| 417 | |
| 418 | // writeScheduler coordinates parent-agent writes against background |
| 419 | // subagent write claims. Set on the parent executor only (subagentDepth 0); |
| 420 | // reservation is taken around Execute so late-loaded MCP/Economy tools are |
| 421 | // covered without registry wrapping. Provider-visible schemas are unchanged. |
| 422 | writeScheduler *SubagentScheduler |
| 423 | // writeWorkspaceRoot is the workspace used to normalize parent write |
| 424 | // reservations when writeScheduler is set. |
| 425 | writeWorkspaceRoot string |
| 426 | |
| 427 | // workspaceLease is shared by every writer-capable agent in one Delivery |
| 428 | // session. It is acquired lazily on the first mutation and held through the |
| 429 | // final participating run/background job so verification remains isolated. |
| 430 | workspaceLease *workspacelease.Owner |
| 431 | |
| 432 | // steerQueue holds mid-turn user messages queued while the agent is |
| 433 | // running. Each is consumed once per loop iteration, persisted to the |
| 434 | // session for history replay, and sent to the model as guidance (not a |
| 435 | // new task). Cache miss for the next API call is unavoidable but limited |
| 436 | // to one call — the prefix stays stable otherwise. |
| 437 | steerMu sync.Mutex |
| 438 | steerQueue []string |
| 439 | steerConsumed bool |
| 440 | // steerRunActive is true while Run is executing. Steer only queues while |
| 441 | // it is set; once the turn's exit flush has drained the queue, later |
| 442 | // steers are rejected so the caller can deliver them as a regular turn |
| 443 | // instead of leaving them in a queue no loop will ever consume. |
| 444 | steerRunActive bool |
| 445 | |
| 446 | // evidence is a per-user-turn ledger of host-observed tool receipts. It lets |
| 447 | // complete_step validate that cited evidence happened before the claim. |
| 448 | evidence *evidence.Ledger |
| 449 | |
| 450 | // todoState is the host's canonical task list: the latest successful |
| 451 | // todo_write with completions applied by complete_step. Unlike the per-turn |
| 452 | // ledger it survives turn boundaries and compaction (it never rides in the |
| 453 | // prompt), so the final-answer gate still sees an unfinished plan a later |
| 454 | // turn would otherwise hide. Rebuilt from the session in SetSession. |
| 455 | todoMu sync.Mutex |
| 456 | todoState []evidence.TodoItem |
| 457 | |
| 458 | // hostAdvanceSeq guarantees unique tool IDs across turns: every |
| 459 | // emitTodoState call increments it so the frontend always sees a fresh |
| 460 | // dispatch even when the same panel index is signed off in different turns. |
| 461 | hostAdvanceSeq atomic.Int64 |
| 462 | |
| 463 | // projectChecks are structured project instructions that complete_step can |
| 464 | // verify against same-turn bash receipts after a write-backed completion. |
| 465 | projectChecks []instruction.VerifyCheck |
| 466 | |
| 467 | // deliveryProfile enables the runtime-enforced delivery contract. The stable |
| 468 | // profile prompt explains intent; these fields are host state and never enter |
| 469 | // the provider-cached prefix. deliveryCriteriaEstablished resets per user turn |
| 470 | // but may inherit an unfinished canonical task list on continuation. |
| 471 | deliveryProfile bool |
| 472 | deliveryCriteriaEstablished bool |
| 473 | deliveryTaskExpected bool |
| 474 | deliveryMutationExpected bool |
| 475 | deliveryPersistentExpected bool |
| 476 | deliveryScopeID string |
| 477 | deliveryScopeActive bool |
| 478 | deliveryCheckpoint evidence.DeliveryCheckpoint |
| 479 | |
| 480 | // ablation names the subsystems a benchmark arm switched off. The zero value |
| 481 | // is the control arm. |
| 482 | ablation ablation.Set |
| 483 | |
| 484 | // classifierTaskText is the host-trusted task text for delivery intent |
| 485 | // classification, set by sub-agent spawners whose Run input carries host |
| 486 | // framing. Empty means classify the raw input verbatim. |
| 487 | classifierTaskText string |
| 488 | |
| 489 | // preserveEvidenceOnce makes the next Run keep the turn evidence ledger |
| 490 | // instead of resetting it. RunSubAgentWithSession sets it before a |
| 491 | // review_report completion nudge so the retry can cite the read receipts |
| 492 | // the subagent already earned; consumed (cleared) by that Run. |
| 493 | preserveEvidenceOnce bool |
| 494 | // deliveryRecoveryPending is armed only when this agent exhausts final |
| 495 | // readiness. An explicit host recovery action can consume it to preserve the |
| 496 | // failed turn's receipts once; an ordinary user turn still resets evidence. |
| 497 | deliveryRecoveryPending bool |
| 498 | // readinessRecovered marks a run that started with evidence preserved from |
| 499 | // (or a pending recovery of) a prior readiness failure, so the final allowed |
| 500 | // audit can report Recovered=true. Set per turn in beginRunTurn. |
| 501 | readinessRecovered bool |
| 502 | |
| 503 | // capabilityLedger tracks require/prefer outcomes for this user turn only. |
| 504 | // Never serialized into prompts or session state. |
| 505 | capabilityLedger *capability.Ledger |
| 506 | // capabilityAudit accumulates non-persisted routing/proxy counters. |
| 507 | capabilityAudit *capability.Audit |
| 508 | // lastCapabilityGate tracks prefer-reminder state across final-answer retries. |
| 509 | capabilityPreferReminded bool |
| 510 | // capabilityRequireMissSeen / capabilityPreferMissSeen remember that the |
| 511 | // final gate reported a miss earlier this turn, so a later clean gate is |
| 512 | // audited as a recovery. Reset per turn in SeedCapabilityRoute. |
| 513 | capabilityRequireMissSeen bool |
| 514 | capabilityPreferMissSeen bool |
| 515 | // pendingReviewWarnings are warn-level findings to surface in the final summary. |
| 516 | pendingReviewWarnings []string |
| 517 | |
| 518 | // memQueue, when non-nil, lets the remember/forget tools fold a turn-tail note |
| 519 | // about a just-made memory change into the next turn, so it applies this |
| 520 | // session without touching the cache-stable prefix. Set via SetMemoryQueue. |
| 521 | memQueue memory.Queue |
| 522 | |
| 523 | // subagentDepth tracks the current agent's nesting depth. maxSubagentDepth |
| 524 | // caps delegation; when reached, recursive agent/skill tools are excluded. |
| 525 | subagentDepth int |
| 526 | maxSubagentDepth int |
| 527 | |
| 528 | // Context management: when a turn's prompt nears contextWindow, the older |
| 529 | // middle of the session is summarized away, keeping a token-bounded recent |
| 530 | // tail verbatim (recentKeep is the message floor) and archiving the originals |
| 531 | // under archiveDir. compactStuck latches when compaction can't get the prompt |
| 532 | // under the window (consecutiveCompacts crosses the limit), so auto-compaction |
| 533 | // pauses instead of looping. softCompactNoticed gates the one-shot soft-ratio |
| 534 | // notice so it fires once per approach, not every turn. |
| 535 | contextWindow int |
| 536 | softCompactRatio float64 |
| 537 | toolResultSnipRatio float64 |
| 538 | compactRatio float64 |
| 539 | compactForceRatio float64 |
| 540 | softCompactNoticed bool |
| 541 | recentKeep int |
| 542 | archiveDir string |
| 543 | keepPolicy KeepPolicy |
| 544 | compactStuck bool |
| 545 | consecutiveCompacts int |
| 546 | // activeTurnCreatedAt identifies the real/synthetic user message that began |
| 547 | // the currently running turn. Compaction may rewrite older history while a |
| 548 | // tool loop is active, but it must keep this message and everything after it |
| 549 | // verbatim so cancellation/crash recovery can retain completed tool pairs. |
| 550 | activeTurnCreatedAt atomic.Int64 |
| 551 | |
| 552 | // stormSig / stormCount track a run of turns that keep failing or getting |
| 553 | // blocked the same way so the loop can break a death-spiral. The signature is |
| 554 | // each call's (tool, error/blocker) in order, NOT (tool, args): a stuck model |
| 555 | // reliably reworks the arguments cosmetically (a re-worded essay, a reordered |
| 556 | // object, a different shell command) while the host returns the same refusal or |
| 557 | // failure every time — keying on args misses the loop entirely. Because errors |
| 558 | // that embed their subject (e.g. "file not found: /x") differ per target, |
| 559 | // genuine varied probing does not collapse to one signature. Reset whenever a |
| 560 | // turn does anything else (a different failure/block shape, or any success). |
| 561 | // See applyStormBreaker. |
| 562 | stormSig string |
| 563 | stormCount int |
| 564 | |
| 565 | // blockedTurnStreak counts consecutive turns in which every tool call was |
| 566 | // blocked by the host (permission, plan mode, hook, or loop guard). |
| 567 | // stormSig catches a model fixated on one call shape; this catches a model |
| 568 | // rotating between blocked shapes — alternating tools, reordering a batch, |
| 569 | // or blockers whose text varies per attempt — which is zero progress all |
| 570 | // the same. Reset by any turn containing a non-blocked outcome and at the |
| 571 | // start of each user turn. See applyStormBreaker. |
| 572 | blockedTurnStreak int |
| 573 | |
| 574 | // loopGuardArmed / loopGuardReceiptMark let final readiness stand down |
| 575 | // after a loop guard fired this user turn: once the host has told the model |
| 576 | // to stop retrying and report the blocker, demanding the receipts that the |
| 577 | // blocker prevents would restart the loop the guard just broke. The mark is |
| 578 | // the evidence-ledger receipt count from just before the guarded batch, so |
| 579 | // real progress — a successful write or command receipt landing after it — |
| 580 | // revokes the pass, while the bookkeeping the guard itself recommends |
| 581 | // (ask, todo_write, complete_step) keeps it. Host state, not message text: |
| 582 | // tool output that merely quotes "[loop guard]" must not unlock readiness. |
| 583 | // Reset at the start of each user turn. See loopGuardAllowsFinal. |
| 584 | loopGuardArmed bool |
| 585 | loopGuardReceiptMark int |
| 586 | |
| 587 | // repeatSuccessCounts tracks write-like tool calls that have already |
| 588 | // succeeded in this user turn. This catches the complementary loop shape to |
| 589 | // stormSig: a model keeps doing the same successful write, so there is no |
| 590 | // error for the failure-only storm breaker to see. |
| 591 | repeatSuccessCounts map[string]int |
| 592 | |
| 593 | // repeatFailureCounts tracks semantically identical write-like calls that |
| 594 | // keep failing with the same failure class. Unlike stormSig, successful |
| 595 | // reads do not blindly clear this state: re-reading a file and then |
| 596 | // resending the same stale anchor is still zero progress. Stale-anchor |
| 597 | // records also survive target mutations until Preview proves the anchor is |
| 598 | // applicable again. Ordinary turns reset the map at Run start; Goal |
| 599 | // continuations retain it while their stable delivery scope is unchanged. |
| 600 | repeatFailureCounts map[string]repeatFailureRecord |
| 601 | repeatFailureScope string |
| 602 | } |
| 603 | |
| 604 | type repeatFailureRecord struct { |
| 605 | count int |
| 606 | errClass string |
| 607 | paths []string |
| 608 | stateRecheck bool |
| 609 | } |
| 610 | |
| 611 | // KeepPolicy is a bitmask controlling which messages are preserved beyond the |
| 612 | // recent tail during compaction. |
| 613 | type KeepPolicy int |
| 614 | |
| 615 | const ( |
| 616 | KeepErrors KeepPolicy = 1 << iota |
| 617 | KeepUserMarked |
| 618 | ) |
| 619 | |
| 620 | // SetPlanMode toggles the plan-first workflow flag. Ordinary calls still use |
| 621 | // Permissions/Sandbox; only explicit phase opt-outs are refused. The system |
| 622 | // prompt and tool schemas stay untouched, while the caller supplies the |
| 623 | // model-facing Marker in a user turn. |
| 624 | func (a *Agent) SetPlanMode(v bool) { a.planMode.Store(v) } |
| 625 | |
| 626 | // SetTools replaces the agent's tool registry. The next API call picks up the |
| 627 | // new tool schema; tools already cached in the provider prefix are unaffected |
| 628 | // until the prefix is invalidated. Safe to call between turns. |
| 629 | func (a *Agent) SetTools(tools *tool.Registry) { |
| 630 | if a == nil { |
| 631 | return |
| 632 | } |
| 633 | a.tools = tools |
| 634 | } |
| 635 | |
| 636 | // SetReasoningLanguage updates the visible reasoning language preference for |
| 637 | // subsequent user-role messages emitted by this agent. |
| 638 | func (a *Agent) SetReasoningLanguage(lang string) { |
| 639 | if a == nil { |
| 640 | return |
| 641 | } |
| 642 | a.reasoningLanguage.Store(NormalizeReasoningLanguage(lang)) |
| 643 | } |
| 644 | |
| 645 | // SetResponseLanguage updates the final-answer language preference for |
| 646 | // subsequent user-role messages emitted by this agent. |
| 647 | func (a *Agent) SetResponseLanguage(lang string) { |
| 648 | if a == nil { |
| 649 | return |
| 650 | } |
| 651 | a.responseLanguage.Store(NormalizeResponseLanguage(lang)) |
| 652 | } |
| 653 | |
| 654 | // SetGate installs the per-call permission gate. Used by interactive CLI sessions to swap the |
| 655 | // headless gate built in setup for an interactive one that prompts the user; |
| 656 | // nil disables gating. Safe to call before the run loop starts. |
| 657 | func (a *Agent) SetGate(g Gate) { |
| 658 | if nilutil.IsNil(g) { |
| 659 | g = nil |
| 660 | } |
| 661 | a.gate = g |
| 662 | } |
| 663 | |
| 664 | // SetExtensions installs the extension dispatcher after construction. Boot |
| 665 | // uses it because sidecars — and therefore the dispatcher — only exist after |
| 666 | // snapshot assembly, which runs after the agent is built. Safe to call before |
| 667 | // the run loop starts; nil disables interception. |
| 668 | func (a *Agent) SetExtensions(d *dispatch.Dispatcher) { |
| 669 | if a == nil { |
| 670 | return |
| 671 | } |
| 672 | a.extensions = d |
| 673 | } |
| 674 | |
| 675 | // SetRecoveryGate installs Auto Guard. Safe to call before the run loop starts; |
| 676 | // nil disables its checks. |
| 677 | func (a *Agent) SetRecoveryGate(g RecoveryGate) { |
| 678 | if a == nil { |
| 679 | return |
| 680 | } |
| 681 | if nilutil.IsNil(g) { |
| 682 | g = nil |
| 683 | } |
| 684 | a.recoveryGate = g |
| 685 | } |
| 686 | |
| 687 | // SetRecoveryIdentity sets the agent/task labels used on recovery cards. |
| 688 | func (a *Agent) SetRecoveryIdentity(agentID, taskID string) { |
| 689 | if a == nil { |
| 690 | return |
| 691 | } |
| 692 | a.recoveryAgentID = strings.TrimSpace(agentID) |
| 693 | a.recoveryTaskID = strings.TrimSpace(taskID) |
| 694 | } |
| 695 | |
| 696 | // RecoveryGate returns the attached Auto Guard (may be nil). |
| 697 | func (a *Agent) RecoveryGate() RecoveryGate { |
| 698 | if a == nil { |
| 699 | return nil |
| 700 | } |
| 701 | return a.recoveryGate |
| 702 | } |
| 703 | |
| 704 | // SetPlanModeReadOnlyTrustGate retains the legacy confirmation bridge for old |
| 705 | // controller/session data. Main Plan execution no longer calls it. |
| 706 | func (a *Agent) SetPlanModeReadOnlyTrustGate(g PlanModeReadOnlyTrustGate) { |
| 707 | if nilutil.IsNil(g) { |
| 708 | g = nil |
| 709 | } |
| 710 | a.planModeReadOnlyTrust = g |
| 711 | } |
| 712 | |
| 713 | // SetSandboxEscapeApprover installs the optional one-shot approval path used by |
| 714 | // the bash tool when an enforced OS sandbox fails to start. |
| 715 | func (a *Agent) SetSandboxEscapeApprover(g sandbox.EscapeApprover) { |
| 716 | if nilutil.IsNil(g) { |
| 717 | g = nil |
| 718 | } |
| 719 | a.sandboxEscapeApprover = g |
| 720 | } |
| 721 | |
| 722 | // SetConfigWriteApprover installs the optional per-write approval path used by |
| 723 | // the file tools when a target is a Reasonix-managed config file outside the |
| 724 | // workspace write roots. |
| 725 | func (a *Agent) SetConfigWriteApprover(g tool.ConfigWriteApprover) { |
| 726 | if nilutil.IsNil(g) { |
| 727 | g = nil |
| 728 | } |
| 729 | a.configWriteApprover = g |
| 730 | } |
| 731 | |
| 732 | func (a *Agent) withTurnPreferences(input string) string { |
| 733 | if a == nil { |
| 734 | return input |
| 735 | } |
| 736 | responseLang := "auto" |
| 737 | if v := a.responseLanguage.Load(); v != nil { |
| 738 | if s, ok := v.(string); ok { |
| 739 | responseLang = s |
| 740 | } |
| 741 | } |
| 742 | input = WithResponseLanguage(input, responseLang) |
| 743 | |
| 744 | lang := "auto" |
| 745 | if v := a.reasoningLanguage.Load(); v != nil { |
| 746 | if s, ok := v.(string); ok { |
| 747 | lang = s |
| 748 | } |
| 749 | } |
| 750 | input = WithReasoningLanguage(input, lang) |
| 751 | if a.deliveryProfile && !strings.Contains(input, "<delivery-runtime>") { |
| 752 | input = strings.TrimSpace(input) + "\n\n" + DeliveryRuntimeMarker |
| 753 | } |
| 754 | return input |
| 755 | } |
| 756 | |
| 757 | // SetAsker installs the asker the `ask` tool uses to question the user. |
| 758 | // Interactive frontends wire one in; headless runs leave it nil. |
| 759 | func (a *Agent) SetAsker(as Asker) { a.asker = as } |
| 760 | |
| 761 | // SetMemoryQueue installs the sink the remember/forget tools use to apply a |
| 762 | // memory change in the current session. The controller wires itself in. |
| 763 | func (a *Agent) SetMemoryQueue(q memory.Queue) { a.memQueue = q } |
| 764 | |
| 765 | // SetPreEditHook installs the pre-edit snapshot hook (see onPreEdit). The |
| 766 | // controller wires it to its per-session checkpoint store; nil disables capture. |
| 767 | // Prefer SetMutationObserver for v2 capture (before+after fingerprints). |
| 768 | func (a *Agent) SetPreEditHook(fn func(diff.Change)) { a.onPreEdit = fn } |
| 769 | |
| 770 | // SetMutationObserver installs the unified mutation observer. When set, it |
| 771 | // supersedes onPreEdit for capture and also records after-mutation fingerprints. |
| 772 | // When a task tool is already registered it inherits the observer for sub-agents. |
| 773 | func (a *Agent) SetMutationObserver(obs *checkpoint.MutationObserver) { |
| 774 | a.mutationObserver = obs |
| 775 | if a.tools == nil || obs == nil { |
| 776 | return |
| 777 | } |
| 778 | if t, ok := a.tools.Get("task"); ok { |
| 779 | if task, ok := t.(*TaskTool); ok { |
| 780 | task.WithMutationObserver(obs) |
| 781 | } |
| 782 | } |
| 783 | } |
| 784 | |
| 785 | // MutationObserver returns the installed observer (may be nil). |
| 786 | func (a *Agent) MutationObserver() *checkpoint.MutationObserver { |
| 787 | if a == nil { |
| 788 | return nil |
| 789 | } |
| 790 | return a.mutationObserver |
| 791 | } |
| 792 | |
| 793 | // Session returns the agent's current conversation, useful for persistence |
| 794 | // hooks that need to read the message log between turns. sessMu serialises this |
| 795 | // pointer read against SetSession, so a frontend (serve's concurrent /history and |
| 796 | // /new handlers) can't race the swap. The run loop touches a.session directly and |
| 797 | // only swaps it via SetSession while idle, so its reads need no lock. |
| 798 | func (a *Agent) Session() *Session { |
| 799 | a.sessMu.Lock() |
| 800 | defer a.sessMu.Unlock() |
| 801 | return a.session |
| 802 | } |
| 803 | |
| 804 | // SetSession replaces the agent's conversation wholesale. Used by |
| 805 | // `reasonix --resume` to load a saved JSONL transcript before the first turn, |
| 806 | // so the model picks up exactly where it left off. Callers serialise it against a |
| 807 | // running turn (it only fires while idle); sessMu guards the pointer swap itself. |
| 808 | func (a *Agent) SetSession(s *Session) { |
| 809 | a.sessMu.Lock() |
| 810 | a.session = s |
| 811 | a.sessMu.Unlock() |
| 812 | a.sessCacheHit.Store(0) |
| 813 | a.sessCacheMiss.Store(0) |
| 814 | a.warnedMissingToolCallReasoning = false |
| 815 | a.missingReasoningWarnStateChecked = false |
| 816 | a.missingReasoningHealthyStreak = 0 |
| 817 | a.repeatFailureCounts = nil |
| 818 | a.repeatFailureScope = "" |
| 819 | if s != nil { |
| 820 | a.rebuildTodoState(s.Snapshot()) |
| 821 | } |
| 822 | } |
| 823 | |
| 824 | // LastUsage returns the most recent per-turn token telemetry the provider |
| 825 | // reported (nil if no turn has run yet). The TUI uses it to show a context |
| 826 | // gauge alongside the prompt; the actual cache decisions still live inside |
| 827 | // maybeCompact. |
| 828 | func (a *Agent) LastUsage() *provider.Usage { return a.lastUsage.Load() } |
| 829 | |
| 830 | // SessionCache returns the cumulative cache hit/miss prompt tokens across every |
| 831 | // API call this session — the basis for the status line's aggregate hit-rate. |
| 832 | func (a *Agent) SessionCache() (hit, miss int) { |
| 833 | return int(a.sessCacheHit.Load()), int(a.sessCacheMiss.Load()) |
| 834 | } |
| 835 | |
| 836 | // ContextWindow returns the configured context-window size in tokens. 0 |
| 837 | // means compaction is disabled for this agent. |
| 838 | func (a *Agent) ContextWindow() int { return a.contextWindow } |
| 839 | |
| 840 | // mid-turn steer marker. |
| 841 | // MidTurnSteerPrefix marks user messages that were injected mid-turn as |
| 842 | // guidance (via Steer). The model sees them as instructions; frontends |
| 843 | // display them as a notice, not a regular user bubble. |
| 844 | const MidTurnSteerPrefix = "[Mid-turn steer queued by the user. Do not treat this as a new task; use it only as additional guidance for the current task after completing the current step.]" |
| 845 | |
| 846 | func midTurnSteerMessage(text string) string { |
| 847 | return MidTurnSteerPrefix + "\n" + text |
| 848 | } |
| 849 | |
| 850 | // SteerText checks whether content is a mid-turn steer message and, if so, |
| 851 | // returns the original user text without the wrapper prefix. The returned |
| 852 | // text preserves the user's exact input — it only strips the prefix and the |
| 853 | // "\n" separator that midTurnSteerMessage inserts between the prefix and the |
| 854 | // user text; it does not trim spaces so the history replay matches the live |
| 855 | // Steer event rendering character-for-character. |
| 856 | // |
| 857 | // Steers are persisted through withTurnPreferences, which can prepend |
| 858 | // transient language blocks (for Chinese text even in auto mode) and append |
| 859 | // the delivery-runtime marker. Both are transport framing, not steer text: |
| 860 | // leading blocks are skipped before matching the prefix and a trailing |
| 861 | // marker is cut from the returned text, so replay recognizes steers |
| 862 | // regardless of the session's language and profile settings. |
| 863 | func SteerText(content string) (string, bool) { |
| 864 | s := content |
| 865 | for { |
| 866 | if after, found := strings.CutPrefix(s, MidTurnSteerPrefix); found { |
| 867 | // Strip only the "\n" separator, preserving the user's original text. |
| 868 | after = strings.TrimPrefix(after, "\n") |
| 869 | if trimmed, cut := strings.CutSuffix(after, "\n\n"+DeliveryRuntimeMarker); cut { |
| 870 | after = trimmed |
| 871 | } |
| 872 | return after, true |
| 873 | } |
| 874 | next, ok := trimLeadingSteerWrapper(s) |
| 875 | if !ok { |
| 876 | return "", false |
| 877 | } |
| 878 | s = next |
| 879 | } |
| 880 | } |
| 881 | |
| 882 | // trimLeadingSteerWrapper removes one leading transient preference block that |
| 883 | // withTurnPreferences may have placed ahead of the steer prefix. It reports |
| 884 | // false when content does not start with such a block. |
| 885 | func trimLeadingSteerWrapper(content string) (string, bool) { |
| 886 | s := strings.TrimLeft(content, " \t\r\n") |
| 887 | for _, tag := range []string{"response-language", "reasoning-language"} { |
| 888 | if !strings.HasPrefix(s, "<"+tag+">") { |
| 889 | continue |
| 890 | } |
| 891 | if rest, ok := trimLeadingTransientBlock(s, tag); ok { |
| 892 | return rest, true |
| 893 | } |
| 894 | } |
| 895 | return content, false |
| 896 | } |
| 897 | |
| 898 | // Steer queues a message for mid-turn injection. It reports whether an active |
| 899 | // turn accepted the text; on false nothing was queued and the caller must |
| 900 | // deliver it another way (typically as a new turn). Without the active check, |
| 901 | // a steer landing in the window between the turn's exit flush and the |
| 902 | // controller observing running=false would sit in the queue unconsumed and |
| 903 | // unpersisted — invisible to both the model and history. |
| 904 | func (a *Agent) Steer(text string) bool { |
| 905 | a.steerMu.Lock() |
| 906 | defer a.steerMu.Unlock() |
| 907 | if !a.steerRunActive { |
| 908 | return false |
| 909 | } |
| 910 | a.steerQueue = append(a.steerQueue, text) |
| 911 | a.steerConsumed = false |
| 912 | return true |
| 913 | } |
| 914 | |
| 915 | // SteerConsumed returns true when the steer queue became empty after the last consume. |
| 916 | func (a *Agent) SteerConsumed() bool { |
| 917 | a.steerMu.Lock() |
| 918 | defer a.steerMu.Unlock() |
| 919 | return a.steerConsumed |
| 920 | } |
| 921 | |
| 922 | func (a *Agent) consumeSteer() (string, bool) { |
| 923 | a.steerMu.Lock() |
| 924 | defer a.steerMu.Unlock() |
| 925 | if len(a.steerQueue) == 0 { |
| 926 | return "", false |
| 927 | } |
| 928 | t := a.steerQueue[0] |
| 929 | a.steerQueue = a.steerQueue[1:] |
| 930 | a.steerConsumed = len(a.steerQueue) == 0 |
| 931 | return t, true |
| 932 | } |
| 933 | |
| 934 | // closeSteerIntakeIfIdle atomically closes the normal-completion race between |
| 935 | // the final queue check and Run returning. A steer accepted before this check |
| 936 | // keeps the loop alive; one arriving after it is rejected so the host can keep |
| 937 | // the user's draft and retry it as a regular follow-up. |
| 938 | func (a *Agent) closeSteerIntakeIfIdle() bool { |
| 939 | a.steerMu.Lock() |
| 940 | defer a.steerMu.Unlock() |
| 941 | if len(a.steerQueue) > 0 { |
| 942 | return false |
| 943 | } |
| 944 | a.steerRunActive = false |
| 945 | return true |
| 946 | } |
| 947 | |
| 948 | // flushSteerQueue ends the turn's steer intake. Guidance that arrived too late |
| 949 | // to be consumed is persisted for transcript visibility but marked local-only: |
| 950 | // replaying it to the model on the next unrelated user turn can execute a stale |
| 951 | // historical task (#7045). An explicit warning keeps the transcript honest |
| 952 | // without presenting the text as successfully applied guidance (#6238). |
| 953 | func (a *Agent) flushSteerQueue() { |
| 954 | a.steerMu.Lock() |
| 955 | pending := a.steerQueue |
| 956 | a.steerQueue = nil |
| 957 | if len(pending) > 0 { |
| 958 | a.steerConsumed = true |
| 959 | } |
| 960 | a.steerRunActive = false |
| 961 | a.steerMu.Unlock() |
| 962 | for _, text := range pending { |
| 963 | a.RecordUnappliedSteer(text) |
| 964 | } |
| 965 | } |
| 966 | |
| 967 | // UnappliedSteerNotice returns the durable warning shown for guidance that was |
| 968 | // accepted during an abnormal turn exit but never reached a provider request. |
| 969 | func UnappliedSteerNotice(text string) string { |
| 970 | return "Guidance was not applied because the turn ended before it could be processed. Send it again if it is still needed:\n" + text |
| 971 | } |
| 972 | |
| 973 | // RecordUnappliedSteer stores guidance that could not affect its intended |
| 974 | // in-flight turn. The orphan-tool sentinel makes older readers drop the record |
| 975 | // during wire normalization, while current readers use LocalOnly to exclude it |
| 976 | // before every provider request. |
| 977 | func (a *Agent) RecordUnappliedSteer(text string) { |
| 978 | if a == nil || a.session == nil { |
| 979 | return |
| 980 | } |
| 981 | a.session.Add(provider.Message{ |
| 982 | Role: provider.RoleTool, |
| 983 | Content: a.withTurnPreferences(midTurnSteerMessage(text)), |
| 984 | ToolCallID: provider.LocalOnlyToolID, |
| 985 | Name: provider.LocalOnlyToolName, |
| 986 | LocalOnly: true, |
| 987 | }) |
| 988 | a.sink.Emit(event.Event{ |
| 989 | Kind: event.Notice, |
| 990 | Level: event.LevelWarn, |
| 991 | Code: event.NoticeCodeUnappliedSteer, |
| 992 | Text: UnappliedSteerNotice(text), |
| 993 | }) |
| 994 | } |
| 995 | |
| 996 | func (a *Agent) steerQueueLen() int { |
| 997 | a.steerMu.Lock() |
| 998 | defer a.steerMu.Unlock() |
| 999 | return len(a.steerQueue) |
| 1000 | } |
| 1001 | |
| 1002 | // CompactRatio returns the fraction of the window at which auto-compaction |
| 1003 | // fires (e.g. 0.8). The status line uses it to show headroom to the next compact. |
| 1004 | func (a *Agent) CompactRatio() float64 { return a.compactRatio } |
| 1005 | |
| 1006 | // CompactNow runs one compaction pass immediately, regardless of the |
| 1007 | // usage-ratio threshold maybeCompact normally honours. Used by the chat |
| 1008 | // TUI's `/compact` command so the user can reset the prefix before it |
| 1009 | // naturally fills up. |
| 1010 | func (a *Agent) CompactNow(ctx context.Context, instructions string) error { |
| 1011 | return a.compact(ctx, "manual", instructions, true) |
| 1012 | } |
| 1013 | |
| 1014 | // Options configures an Agent. |
| 1015 | type Options struct { |
| 1016 | MaxSteps int |
| 1017 | // MaxStepsKey names the explicit runtime control shown when the MaxSteps guard |
| 1018 | // is hit. Empty defaults to the generic max_steps tool/runtime parameter. |
| 1019 | MaxStepsKey string |
| 1020 | // ReasoningByteLimit bounds a single stream's hidden reasoning bytes. Zero |
| 1021 | // uses the default guard; a negative value disables only this client guard. |
| 1022 | // Provider output budgets are a separate protocol/model capability. |
| 1023 | ReasoningByteLimit int |
| 1024 | // MaxOutputTokens overrides the provider's configured/default total output |
| 1025 | // budget. Zero delegates to the provider; a negative value asks optional |
| 1026 | // protocols to omit the budget (Anthropic still requires max_tokens). |
| 1027 | MaxOutputTokens int |
| 1028 | Temperature float64 |
| 1029 | Pricing *provider.Pricing // optional, for per-turn cost display |
| 1030 | UsageSource string // optional billable usage source; default executor |
| 1031 | // ModelRef names the canonical "provider/model" ref backing this agent's |
| 1032 | // provider instance. It is attached to emitted Usage events so downstream |
| 1033 | // usage accounting can attribute tokens to the exact model. |
| 1034 | ModelRef string |
| 1035 | |
| 1036 | // Gate is the per-call permission gate. nil disables gating. |
| 1037 | Gate Gate |
| 1038 | |
| 1039 | // ReadOnlyExecution enables a permanent host-side read-only boundary for |
| 1040 | // planner and research agents. It is intentionally independent of Plan mode |
| 1041 | // so a stale collaboration flag cannot authorize a dynamic writer target. |
| 1042 | ReadOnlyExecution bool |
| 1043 | |
| 1044 | // PlannerMCPExecution enables Planner-trusted MCP through use_capability: |
| 1045 | // authorized, non-destructive tools may run without readOnlyHint. Only |
| 1046 | // NewPlannerAgent sets this; strict read-only sub-agents must not. |
| 1047 | PlannerMCPExecution bool |
| 1048 | |
| 1049 | // PlanModeReadOnlyTrustGate is retained for legacy controller compatibility. |
| 1050 | // The main Plan execution path no longer invokes it. |
| 1051 | PlanModeReadOnlyTrustGate PlanModeReadOnlyTrustGate |
| 1052 | |
| 1053 | // SandboxEscapeApprover confirms a one-shot unconfined shell rerun after an |
| 1054 | // enforced OS sandbox fails. nil keeps fail-closed behavior. |
| 1055 | SandboxEscapeApprover sandbox.EscapeApprover |
| 1056 | |
| 1057 | // ConfigWriteApprover confirms file-tool writes to Reasonix-managed config |
| 1058 | // files outside the workspace roots. nil keeps fail-closed behavior. |
| 1059 | ConfigWriteApprover tool.ConfigWriteApprover |
| 1060 | |
| 1061 | // Context management. ContextWindow <= 0 disables compaction. Ratios and |
| 1062 | // RecentKeep fall back to defaults when unset. |
| 1063 | ContextWindow int |
| 1064 | SoftCompactRatio float64 |
| 1065 | ToolResultSnipRatio float64 |
| 1066 | CompactRatio float64 |
| 1067 | CompactForceRatio float64 |
| 1068 | RecentKeep int |
| 1069 | ArchiveDir string |
| 1070 | KeepPolicy KeepPolicy |
| 1071 | |
| 1072 | // Hooks fires PreToolUse / PostToolUse shell hooks around tool calls. nil |
| 1073 | // disables hook firing. |
| 1074 | Hooks ToolHooks |
| 1075 | |
| 1076 | // MissingReasoningWarnStateDir, when non-empty, points at the shared |
| 1077 | // directory where missing tool-call thinking recovery retries are gated by |
| 1078 | // opaque provider-configuration fingerprint (#7059). The field name is kept |
| 1079 | // for source compatibility. Boot always supplies it; direct construction |
| 1080 | // with an empty value keeps in-memory gating. |
| 1081 | MissingReasoningWarnStateDir string |
| 1082 | |
| 1083 | // Jobs is the session's background-job manager (nil disables background tools). |
| 1084 | Jobs *jobs.Manager |
| 1085 | |
| 1086 | // WriteScheduler is the session-scoped subagent concurrency/write-claim |
| 1087 | // controller. When set on the parent executor, write-capable tools reserve |
| 1088 | // paths for the duration of Execute so background writers cannot TOCTOU |
| 1089 | // race parent writes. Subagents leave this nil (or depth > 0 skips it). |
| 1090 | WriteScheduler *SubagentScheduler |
| 1091 | // WriteWorkspaceRoot normalizes parent write reservations. |
| 1092 | WriteWorkspaceRoot string |
| 1093 | |
| 1094 | // WorkspaceLease serializes Delivery mutations across sessions that target |
| 1095 | // the same workspace. nil preserves source compatibility for direct Agent |
| 1096 | // construction; boot always supplies it for Delivery sessions. |
| 1097 | WorkspaceLease *workspacelease.Owner |
| 1098 | |
| 1099 | // ProjectChecks are host-observable structured checks extracted during boot. |
| 1100 | ProjectChecks []instruction.VerifyCheck |
| 1101 | |
| 1102 | // DeliveryProfile enforces acceptance criteria before mutations and requires |
| 1103 | // post-change review, verification, and evidence-backed sign-off before a |
| 1104 | // final answer. It changes host control flow, not tool schemas. |
| 1105 | DeliveryProfile bool |
| 1106 | |
| 1107 | // Ablation switches subsystems off for a benchmark arm. The zero value runs |
| 1108 | // everything, so ordinary callers leave it unset. |
| 1109 | Ablation ablation.Set |
| 1110 | |
| 1111 | // ClassifierTaskText, when non-empty, is the pristine task text delivery |
| 1112 | // intent classification should judge instead of the raw Run input. Sub-agent |
| 1113 | // spawners set it before prepending host framing (subagent/workspace context, |
| 1114 | // review contracts) so framing verbs cannot arm expectations and user input |
| 1115 | // dressed up as framing cannot disarm them. |
| 1116 | ClassifierTaskText string |
| 1117 | |
| 1118 | // CapabilityLedger is the optional turn-scoped capability route ledger for |
| 1119 | // Delivery require/prefer gates. Nil disables capability gates. |
| 1120 | CapabilityLedger *capability.Ledger |
| 1121 | // CapabilityAudit is the optional non-persisted metrics sink for routing. |
| 1122 | CapabilityAudit *capability.Audit |
| 1123 | |
| 1124 | // RequireReviewReportKind, when non-empty, makes RunSubAgentWithSession fail |
| 1125 | // unless the subagent recorded a successful review_report of this kind — |
| 1126 | // review/security subagents must return typed, host-verifiable reports. |
| 1127 | RequireReviewReportKind evidence.ReviewKind |
| 1128 | |
| 1129 | // ReasoningLanguage controls visible reasoning language preference as transient |
| 1130 | // user-turn context. Empty/auto injects nothing. |
| 1131 | ReasoningLanguage string |
| 1132 | |
| 1133 | // ResponseLanguage controls final-answer language preference as transient |
| 1134 | // user-turn context. Empty/auto keeps the stable same-as-user policy. |
| 1135 | ResponseLanguage string |
| 1136 | |
| 1137 | // PlanModeReadOnlyCommands is retained for old config/controller data. Main |
| 1138 | // Plan execution classifies bash through Permissions instead. |
| 1139 | PlanModeReadOnlyCommands []string |
| 1140 | |
| 1141 | // RecoveryGate is the optional Auto Guard boundary. It checks deterministic |
| 1142 | // high-risk mutations and failure recovery before permission approval and |
| 1143 | // write-lock acquisition. |
| 1144 | RecoveryGate RecoveryGate |
| 1145 | // RecoveryAgentID labels this agent on recovery cards (empty = root). |
| 1146 | RecoveryAgentID string |
| 1147 | // RecoveryTaskID isolates recovery state for this agent (empty = root task). |
| 1148 | RecoveryTaskID string |
| 1149 | |
| 1150 | // SubagentDepth is the current nesting depth for this agent. Root sessions are |
| 1151 | // depth 0; child subagents are depth 1. MaxSubagentDepth caps delegation. |
| 1152 | SubagentDepth int |
| 1153 | MaxSubagentDepth int |
| 1154 | |
| 1155 | // Extensions is the frozen extension dispatcher for this agent's controller |
| 1156 | // generation (Extension Protocol v1). Nil means no v1 runtime packages are |
| 1157 | // installed; the run loop then passes every intercept point through |
| 1158 | // byte-identically. Boot installs it with SetExtensions once sidecars are |
| 1159 | // live (they start after the agent is constructed). |
| 1160 | Extensions *dispatch.Dispatcher |
| 1161 | |
| 1162 | // MutationObserver is the host-side file mutation observer shared with |
| 1163 | // (or cloned for) sub-agents. nil disables v2 capture. Does not affect |
| 1164 | // provider-visible tool schemas or prompts. |
| 1165 | MutationObserver *checkpoint.MutationObserver |
| 1166 | } |
| 1167 | |
| 1168 | // New constructs an Agent. MaxSteps <= 0 means no cap — the run loop continues |
| 1169 | // until the model gives a final answer, the context is cancelled, or the |
| 1170 | // provider errors (compaction keeps the context bounded). A nil sink is replaced |
| 1171 | // with event.Discard so the agent can always emit unconditionally. |
| 1172 | func New(prov provider.Provider, tools *tool.Registry, session *Session, opts Options, sink event.Sink) *Agent { |
| 1173 | if opts.SoftCompactRatio <= 0 { |
| 1174 | opts.SoftCompactRatio = defaultSoftCompactRatio |
| 1175 | } |
| 1176 | if opts.ToolResultSnipRatio <= 0 { |
| 1177 | opts.ToolResultSnipRatio = defaultToolResultSnipRatio |
| 1178 | } |
| 1179 | if opts.CompactRatio <= 0 { |
| 1180 | opts.CompactRatio = defaultCompactRatio |
| 1181 | } |
| 1182 | if opts.ToolResultSnipRatio >= opts.CompactRatio { |
| 1183 | opts.ToolResultSnipRatio = opts.CompactRatio |
| 1184 | } |
| 1185 | if opts.CompactForceRatio <= 0 { |
| 1186 | opts.CompactForceRatio = defaultCompactForceRatio |
| 1187 | } |
| 1188 | if opts.RecentKeep <= 0 { |
| 1189 | opts.RecentKeep = minRecentKeep |
| 1190 | } |
| 1191 | if nilutil.IsNil(sink) { |
| 1192 | sink = event.Discard |
| 1193 | } |
| 1194 | gate := opts.Gate |
| 1195 | if nilutil.IsNil(gate) { |
| 1196 | gate = nil |
| 1197 | } |
| 1198 | planModeReadOnlyTrust := opts.PlanModeReadOnlyTrustGate |
| 1199 | if nilutil.IsNil(planModeReadOnlyTrust) { |
| 1200 | planModeReadOnlyTrust = nil |
| 1201 | } |
| 1202 | sandboxEscapeApprover := opts.SandboxEscapeApprover |
| 1203 | if nilutil.IsNil(sandboxEscapeApprover) { |
| 1204 | sandboxEscapeApprover = nil |
| 1205 | } |
| 1206 | configWriteApprover := opts.ConfigWriteApprover |
| 1207 | if nilutil.IsNil(configWriteApprover) { |
| 1208 | configWriteApprover = nil |
| 1209 | } |
| 1210 | hooks := opts.Hooks |
| 1211 | if nilutil.IsNil(hooks) { |
| 1212 | hooks = nil |
| 1213 | } |
| 1214 | maxStepsKey := opts.MaxStepsKey |
| 1215 | if strings.TrimSpace(maxStepsKey) == "" { |
| 1216 | maxStepsKey = "max_steps" |
| 1217 | } |
| 1218 | maxSubagentDepth := opts.MaxSubagentDepth |
| 1219 | if maxSubagentDepth == 0 { |
| 1220 | maxSubagentDepth = DefaultMaxSubagentDepth |
| 1221 | } else { |
| 1222 | maxSubagentDepth = NormalizeMaxSubagentDepth(maxSubagentDepth) |
| 1223 | } |
| 1224 | subagentDepth := opts.SubagentDepth |
| 1225 | if subagentDepth < 0 { |
| 1226 | subagentDepth = 0 |
| 1227 | } |
| 1228 | reasoningByteLimit := opts.ReasoningByteLimit |
| 1229 | if reasoningByteLimit == 0 { |
| 1230 | reasoningByteLimit = defaultReasoningByteLimit |
| 1231 | } |
| 1232 | a := &Agent{ |
| 1233 | prov: prov, |
| 1234 | tools: tools, |
| 1235 | session: session, |
| 1236 | maxSteps: opts.MaxSteps, |
| 1237 | maxStepsKey: maxStepsKey, |
| 1238 | reasoningByteLimit: reasoningByteLimit, |
| 1239 | maxOutputTokens: opts.MaxOutputTokens, |
| 1240 | temperature: opts.Temperature, |
| 1241 | pricing: opts.Pricing, |
| 1242 | usageSource: usageSourceOrDefault(opts.UsageSource, event.UsageSourceExecutor), |
| 1243 | modelRef: strings.TrimSpace(opts.ModelRef), |
| 1244 | sink: sink, |
| 1245 | gate: gate, |
| 1246 | extensions: opts.Extensions, |
| 1247 | recoveryGate: opts.RecoveryGate, |
| 1248 | recoveryAgentID: strings.TrimSpace(opts.RecoveryAgentID), |
| 1249 | recoveryTaskID: strings.TrimSpace(opts.RecoveryTaskID), |
| 1250 | readOnlyExecution: opts.ReadOnlyExecution, |
| 1251 | plannerMCPExecution: opts.PlannerMCPExecution, |
| 1252 | planModeReadOnlyTrust: planModeReadOnlyTrust, |
| 1253 | sandboxEscapeApprover: sandboxEscapeApprover, |
| 1254 | configWriteApprover: configWriteApprover, |
| 1255 | hooks: hooks, |
| 1256 | jobs: opts.Jobs, |
| 1257 | writeScheduler: opts.WriteScheduler, |
| 1258 | writeWorkspaceRoot: strings.TrimSpace(opts.WriteWorkspaceRoot), |
| 1259 | workspaceLease: opts.WorkspaceLease, |
| 1260 | missingReasoningWarnState: missingReasoningWarnStateFor(opts.MissingReasoningWarnStateDir), |
| 1261 | evidence: evidence.NewLedger(), |
| 1262 | projectChecks: append([]instruction.VerifyCheck(nil), opts.ProjectChecks...), |
| 1263 | deliveryProfile: opts.DeliveryProfile, |
| 1264 | ablation: opts.Ablation, |
| 1265 | classifierTaskText: opts.ClassifierTaskText, |
| 1266 | capabilityLedger: opts.CapabilityLedger, |
| 1267 | capabilityAudit: opts.CapabilityAudit, |
| 1268 | contextWindow: opts.ContextWindow, |
| 1269 | softCompactRatio: opts.SoftCompactRatio, |
| 1270 | toolResultSnipRatio: opts.ToolResultSnipRatio, |
| 1271 | compactRatio: opts.CompactRatio, |
| 1272 | compactForceRatio: opts.CompactForceRatio, |
| 1273 | recentKeep: opts.RecentKeep, |
| 1274 | archiveDir: opts.ArchiveDir, |
| 1275 | keepPolicy: opts.KeepPolicy, |
| 1276 | subagentDepth: subagentDepth, |
| 1277 | maxSubagentDepth: maxSubagentDepth, |
| 1278 | mutationObserver: opts.MutationObserver, |
| 1279 | } |
| 1280 | a.SetResponseLanguage(opts.ResponseLanguage) |
| 1281 | a.SetReasoningLanguage(opts.ReasoningLanguage) |
| 1282 | return a |
| 1283 | } |
| 1284 | |
| 1285 | func usageSourceOrDefault(source, fallback string) string { |
| 1286 | source = strings.TrimSpace(source) |
| 1287 | if source != "" { |
| 1288 | return source |
| 1289 | } |
| 1290 | return fallback |
| 1291 | } |
| 1292 | |
| 1293 | // missingReasoningWarnStateFor returns nil when no state dir is configured, so |
| 1294 | // direct Agent construction keeps the historical once-per-session notice scope. |
| 1295 | func missingReasoningWarnStateFor(dir string) *missingReasoningWarnState { |
| 1296 | if strings.TrimSpace(dir) == "" { |
| 1297 | return nil |
| 1298 | } |
| 1299 | return newMissingReasoningWarnState(dir) |
| 1300 | } |
| 1301 | |
| 1302 | // reserveParentWrite holds write claims for the duration of a parent-agent |
| 1303 | // write tool call. Returns a no-op release when reservation is not needed |
| 1304 | // (subagent, read-only, no scheduler, or non-write tool). |
| 1305 | func (a *Agent) reserveParentWrite(runTool tool.Tool, args json.RawMessage, readOnly bool) (release func(), err error) { |
| 1306 | noop := func() {} |
| 1307 | if a == nil || a.writeScheduler == nil || a.subagentDepth > 0 || readOnly || runTool == nil { |
| 1308 | return noop, nil |
| 1309 | } |
| 1310 | name := runTool.Name() |
| 1311 | if !parentWriteGuardTarget(name) { |
| 1312 | return noop, nil |
| 1313 | } |
| 1314 | claim, err := parentWriteReservation(a.writeWorkspaceRoot, name, args) |
| 1315 | if err != nil { |
| 1316 | return noop, err |
| 1317 | } |
| 1318 | return a.writeScheduler.ReserveParentWrite(claim) |
| 1319 | } |
| 1320 | |
| 1321 | // Run appends the user input and drives the tool loop until the model returns a |
| 1322 | // final answer (no tool calls), the context is cancelled, or the provider errors. |
| 1323 | // With maxSteps <= 0 the loop is unbounded — the natural termination is the model |
| 1324 | // finishing, and the real safety bounds are user cancellation and compaction, not |
| 1325 | // a round count. A positive maxSteps imposes an optional hard guard, surfaced as |
| 1326 | // a resumable notice when hit. |
| 1327 | // Run is the agent lifecycle entry point: lifecycle setup, turn initialization, |
| 1328 | // tool-round loop, and deferred cleanup. Turn policy lives in beginRunTurn / |
| 1329 | // runToolLoop / handleFinalResponse / handleToolRound so the state machine stays |
| 1330 | // readable without changing provider-visible behavior or lock ownership. |
| 1331 | func (a *Agent) Run(ctx context.Context, input string) (runErr error) { |
| 1332 | runMaxSteps := a.maxSteps |
| 1333 | runMaxStepsKey := a.maxStepsKey |
| 1334 | runLimitHostOwned := false |
| 1335 | if limit, ok := runStepLimitFromContext(ctx); ok { |
| 1336 | runMaxSteps = limit.steps |
| 1337 | runLimitHostOwned = true |
| 1338 | if limit.key != "" { |
| 1339 | runMaxStepsKey = limit.key |
| 1340 | } |
| 1341 | } |
| 1342 | a.recoveryRunSeq.Add(1) |
| 1343 | if a.deliveryProfile && a.workspaceLease != nil { |
| 1344 | a.workspaceLease.BeginRun() |
| 1345 | defer a.workspaceLease.EndRun() |
| 1346 | } |
| 1347 | turnStartedAt := time.Now() |
| 1348 | workDurationMs := func() int64 { |
| 1349 | if elapsed := time.Since(turnStartedAt).Milliseconds(); elapsed > 0 { |
| 1350 | return elapsed |
| 1351 | } |
| 1352 | return 1 |
| 1353 | } |
| 1354 | defer a.flushSteerQueue() |
| 1355 | a.steerMu.Lock() |
| 1356 | a.steerConsumed = false |
| 1357 | a.steerRunActive = true |
| 1358 | a.steerMu.Unlock() |
| 1359 | |
| 1360 | // Commit background-job evidence leases only after this turn delivers. |
| 1361 | // wait/bash_output merge a finished background writer's receipts into the |
| 1362 | // ledger provisionally; if the turn reaches a final answer (runErr == nil) |
| 1363 | // the delivery gates have verified and reviewed those mutations, so the |
| 1364 | // job's evidence can be permanently drained. A failed or cancelled turn |
| 1365 | // leaves the lease uncommitted so the next turn re-collects it. |
| 1366 | defer func() { |
| 1367 | if runErr != nil || a.evidence == nil || a.jobs == nil { |
| 1368 | return |
| 1369 | } |
| 1370 | for _, lease := range a.evidence.BackgroundLeases() { |
| 1371 | a.jobs.CommitEvidenceForSession(lease.Session, lease.JobID) |
| 1372 | } |
| 1373 | }() |
| 1374 | if _, scoped := DeliveryExecutionScopeFromContext(ctx); scoped { |
| 1375 | defer func() { a.updateDeliveryCheckpoint(runErr) }() |
| 1376 | } |
| 1377 | defer a.activeTurnCreatedAt.Store(0) |
| 1378 | |
| 1379 | // agent.before_start: an extension may abort the run before the user turn |
| 1380 | // is appended. The redacted reason surfaces like a normal run error. |
| 1381 | if err := a.interceptAgentStart(ctx); err != nil { |
| 1382 | return err |
| 1383 | } |
| 1384 | |
| 1385 | _, state := a.beginRunTurn(ctx, input) |
| 1386 | state.runMaxSteps = runMaxSteps |
| 1387 | state.runMaxStepsKey = runMaxStepsKey |
| 1388 | state.runLimitHostOwned = runLimitHostOwned |
| 1389 | state.workDurationMs = workDurationMs |
| 1390 | return a.runToolLoop(ctx, state) |
| 1391 | } |
| 1392 | |
| 1393 | // observeMissingToolCallReasoning classifies a thinking-mode tool-call turn and |
| 1394 | // claims the single silent retry allowed for its active compatibility incident. |
| 1395 | // DeepSeek requires provider-issued thinking content to be replayed, so a |
| 1396 | // missing value is retried once before tools execute. Persistent broken rounds |
| 1397 | // use the existing exact-configuration cooldown; a healthy round resolves the |
| 1398 | // incident after three consecutive healthy turns and re-arms a future isolated |
| 1399 | // regression (#6259, #7059). |
| 1400 | func (a *Agent) observeMissingToolCallReasoning(calls []provider.ToolCall, reasoning string) (missing, shouldRetry bool) { |
| 1401 | if len(calls) == 0 || !provider.WarnOnMissingToolCallReasoning(a.prov) { |
| 1402 | return false, false |
| 1403 | } |
| 1404 | fingerprint := provider.MissingToolCallReasoningWarningFingerprint(a.prov) |
| 1405 | observedAt := time.Now() |
| 1406 | if strings.TrimSpace(reasoning) != "" { |
| 1407 | if a.missingReasoningWarnState == nil { |
| 1408 | if a.warnedMissingToolCallReasoning { |
| 1409 | a.missingReasoningHealthyStreak++ |
| 1410 | if a.missingReasoningHealthyStreak >= missingReasoningHealthyResolveStreak { |
| 1411 | a.warnedMissingToolCallReasoning = false |
| 1412 | a.missingReasoningHealthyStreak = 0 |
| 1413 | } |
| 1414 | } |
| 1415 | return false, false |
| 1416 | } |
| 1417 | shouldResolve := !a.missingReasoningWarnStateChecked || a.warnedMissingToolCallReasoning |
| 1418 | if shouldResolve { |
| 1419 | result := missingReasoningResolveResult{Recorded: true, Resolved: true} |
| 1420 | if pending := a.missingReasoningWarnPendingResolveAt; !pending.IsZero() { |
| 1421 | result = a.missingReasoningWarnState.resolveAt(fingerprint, pending) |
| 1422 | if result.Recorded { |
| 1423 | a.missingReasoningWarnPendingResolveAt = time.Time{} |
| 1424 | } |
| 1425 | } |
| 1426 | if result.Recorded { |
| 1427 | result = a.missingReasoningWarnState.resolveAt(fingerprint, observedAt) |
| 1428 | } |
| 1429 | if !result.Recorded { |
| 1430 | if observedAt.After(a.missingReasoningWarnPendingResolveAt) { |
| 1431 | a.missingReasoningWarnPendingResolveAt = observedAt |
| 1432 | } |
| 1433 | a.warnedMissingToolCallReasoning = true |
| 1434 | a.missingReasoningWarnStateChecked = false |
| 1435 | } else if result.Resolved { |
| 1436 | a.warnedMissingToolCallReasoning = false |
| 1437 | a.missingReasoningWarnStateChecked = true |
| 1438 | } else { |
| 1439 | a.warnedMissingToolCallReasoning = true |
| 1440 | a.missingReasoningWarnStateChecked = false |
| 1441 | } |
| 1442 | } |
| 1443 | return false, false |
| 1444 | } |
| 1445 | a.missingReasoningHealthyStreak = 0 |
| 1446 | if s := a.missingReasoningWarnState; s != nil { |
| 1447 | stateReady := true |
| 1448 | alreadyActive := a.warnedMissingToolCallReasoning |
| 1449 | if pending := a.missingReasoningWarnPendingResolveAt; !pending.IsZero() { |
| 1450 | result := s.resolveAt(fingerprint, pending) |
| 1451 | stateReady = result.Recorded |
| 1452 | if result.Recorded { |
| 1453 | a.missingReasoningWarnPendingResolveAt = time.Time{} |
| 1454 | if result.Resolved { |
| 1455 | alreadyActive = false |
| 1456 | a.warnedMissingToolCallReasoning = false |
| 1457 | } |
| 1458 | } |
| 1459 | } |
| 1460 | claimed := stateReady && s.claimAt(fingerprint, observedAt) |
| 1461 | if !claimed || alreadyActive { |
| 1462 | // This exact configuration already attempted recovery for the active |
| 1463 | // incident, so keep the empty-key fallback without doubling requests. |
| 1464 | a.warnedMissingToolCallReasoning = true |
| 1465 | a.missingReasoningWarnStateChecked = true |
| 1466 | return true, false |
| 1467 | } |
| 1468 | if !stateReady { |
| 1469 | a.missingReasoningWarnStateChecked = false |
| 1470 | } |
| 1471 | } else if a.warnedMissingToolCallReasoning { |
| 1472 | return true, false |
| 1473 | } |
| 1474 | a.warnedMissingToolCallReasoning = true |
| 1475 | if a.missingReasoningWarnPendingResolveAt.IsZero() { |
| 1476 | a.missingReasoningWarnStateChecked = true |
| 1477 | } |
| 1478 | return true, true |
| 1479 | } |
| 1480 | |
| 1481 | // maxStepsPause is the deliberate stop when a positive tool-call budget runs |
| 1482 | // out: the session already holds the completed work and the user is asked to |
| 1483 | // continue. It is a control-flow signal, not a provider failure. Coordinator |
| 1484 | // treats planner research budgets specially: ordinary plan-and-execute work |
| 1485 | // falls back to the executor, while explicit execution boundaries fail closed. |
| 1486 | type maxStepsPause struct { |
| 1487 | steps int |
| 1488 | key string |
| 1489 | } |
| 1490 | |
| 1491 | func (e *maxStepsPause) Error() string { |
| 1492 | return fmt.Sprintf("paused after %d tool-call rounds (%s) — the work so far is saved; send another message to continue, or set %s higher or to 0 for no limit", e.steps, e.key, e.key) |
| 1493 | } |
| 1494 | |
| 1495 | type todoStallPause struct { |
| 1496 | rounds int |
| 1497 | } |
| 1498 | |
| 1499 | func (e *todoStallPause) Error() string { |
| 1500 | return fmt.Sprintf("paused after %d tool-call rounds without advancing the current todo — the work so far is saved; inspect the blocker or send another message to continue", e.rounds) |
| 1501 | } |
| 1502 | |
| 1503 | func isToolLoopPause(err error) bool { |
| 1504 | var maxPause *maxStepsPause |
| 1505 | var stallPause *todoStallPause |
| 1506 | return errors.As(err, &maxPause) || errors.As(err, &stallPause) |
| 1507 | } |
| 1508 | |
| 1509 | // ReadinessResult is the host-consumable outcome of the Delivery final-answer |
| 1510 | // readiness check. The Controller reads it after each goal turn; plain turns |
| 1511 | // receive the same outcome as a FinalReadinessError. |
| 1512 | type ReadinessResult struct { |
| 1513 | // Ready is true when no missing requirement remains. |
| 1514 | Ready bool |
| 1515 | // Missing lists stable category ids of the missing requirements |
| 1516 | // (project_check, todo, criteria, verification, review, signoff, action, |
| 1517 | // mutation, capability). Empty when Ready. |
| 1518 | Missing []string |
| 1519 | // Reason is the user-facing summary of what is still missing. |
| 1520 | Reason string |
| 1521 | // ProgressKey is the host-verifiable progress signature of the current |
| 1522 | // evidence state. Identical ProgressKey across consecutive goal turns |
| 1523 | // means no host-observable progress was made. |
| 1524 | ProgressKey string |
| 1525 | } |
| 1526 | |
| 1527 | // ReadinessResult returns the current final-readiness outcome for the host. |
| 1528 | func (a *Agent) ReadinessResult() ReadinessResult { |
| 1529 | check := a.finalReadinessCheckFor() |
| 1530 | if check.reason == "" { |
| 1531 | return ReadinessResult{Ready: true, ProgressKey: check.progressSignature()} |
| 1532 | } |
| 1533 | return ReadinessResult{ |
| 1534 | Ready: false, |
| 1535 | Missing: check.missingIDs(), |
| 1536 | Reason: check.reason, |
| 1537 | ProgressKey: check.progressSignature(), |
| 1538 | } |
| 1539 | } |
| 1540 | |
| 1541 | // HostProgressSignature returns a compact signature of host-observable progress |
| 1542 | // across the current delivery scope: successful writes, commands, todo writes, |
| 1543 | // signoffs, and reviews. Identical signatures across consecutive goal turns |
| 1544 | // mean no host-verifiable progress was made — reads, reworded answers, and |
| 1545 | // repeated continue reasons never reset the stall counter. |
| 1546 | func (a *Agent) HostProgressSignature() string { |
| 1547 | if a == nil || a.evidence == nil { |
| 1548 | return "" |
| 1549 | } |
| 1550 | s := a.evidence.ReceiptProgressSummary() |
| 1551 | return fmt.Sprintf("w=%d;c=%d;t=%d;s=%d;r=%d", s.Writes, s.Commands, s.Todos, s.Signoffs, s.Reviews) |
| 1552 | } |
| 1553 | |
| 1554 | type finalReadinessCheck struct { |
| 1555 | applies bool |
| 1556 | reason string |
| 1557 | missingProjectChecks int |
| 1558 | incompleteTodos int |
| 1559 | missingAcceptanceCriteria int |
| 1560 | missingVerification int |
| 1561 | missingReview int |
| 1562 | missingSignoff int |
| 1563 | missingActionEvidence int |
| 1564 | missingMutation int |
| 1565 | missingCapabilities int |
| 1566 | } |
| 1567 | |
| 1568 | func (c finalReadinessCheck) progressSignature() string { |
| 1569 | return fmt.Sprintf("%d/%d/%d/%d/%d/%d/%d/%d/%d/%d\x00%s", |
| 1570 | c.missingProjectChecks, |
| 1571 | c.incompleteTodos, |
| 1572 | c.missingAcceptanceCriteria, |
| 1573 | c.missingVerification, |
| 1574 | c.missingReview, |
| 1575 | c.missingSignoff, |
| 1576 | c.missingActionEvidence, |
| 1577 | c.missingMutation, |
| 1578 | c.missingCapabilities, |
| 1579 | boolInt(c.applies), |
| 1580 | c.reason, |
| 1581 | ) |
| 1582 | } |
| 1583 | |
| 1584 | func (c finalReadinessCheck) missingIDs() []string { |
| 1585 | missing := make([]string, 0, 9) |
| 1586 | add := func(id string, count int) { |
| 1587 | if count > 0 { |
| 1588 | missing = append(missing, id) |
| 1589 | } |
| 1590 | } |
| 1591 | add("project_check", c.missingProjectChecks) |
| 1592 | add("todo", c.incompleteTodos) |
| 1593 | add("criteria", c.missingAcceptanceCriteria) |
| 1594 | add("verification", c.missingVerification) |
| 1595 | add("review", c.missingReview) |
| 1596 | add("signoff", c.missingSignoff) |
| 1597 | add("action", c.missingActionEvidence) |
| 1598 | add("mutation", c.missingMutation) |
| 1599 | add("capability", c.missingCapabilities) |
| 1600 | return missing |
| 1601 | } |
| 1602 | |
| 1603 | func boolInt(v bool) int { |
| 1604 | if v { |
| 1605 | return 1 |
| 1606 | } |
| 1607 | return 0 |
| 1608 | } |
| 1609 | |
| 1610 | func (c finalReadinessCheck) audit(result evidence.ReadinessAuditResult, recovered bool) evidence.ReadinessAudit { |
| 1611 | return evidence.ReadinessAudit{ |
| 1612 | Result: result, |
| 1613 | Recovered: recovered, |
| 1614 | MissingProjectChecks: c.missingProjectChecks, |
| 1615 | IncompleteTodos: c.incompleteTodos, |
| 1616 | CommandMismatchMissing: c.missingProjectChecks, |
| 1617 | MissingAcceptanceCriteria: c.missingAcceptanceCriteria, |
| 1618 | MissingVerification: c.missingVerification, |
| 1619 | MissingReview: c.missingReview, |
| 1620 | MissingSignoff: c.missingSignoff, |
| 1621 | MissingActionEvidence: c.missingActionEvidence, |
| 1622 | MissingMutation: c.missingMutation, |
| 1623 | MissingCapabilities: c.missingCapabilities, |
| 1624 | } |
| 1625 | } |
| 1626 | |
| 1627 | func (a *Agent) finalReadinessCheckFor() finalReadinessCheck { |
| 1628 | if a.evidence == nil || a.ablation.Off(ablation.Evidence) { |
| 1629 | return finalReadinessCheck{} |
| 1630 | } |
| 1631 | var missing []string |
| 1632 | out := finalReadinessCheck{} |
| 1633 | // Planning returns a proposal to the controller, which owns the approval gate |
| 1634 | // and starts a fresh execution turn after Plan is disabled. Delivery completion |
| 1635 | // requirements, including required capabilities, wait for that execution turn: |
| 1636 | // forcing them here could make a writer requirement contradict the plan-first |
| 1637 | // workflow. This is a workflow boundary only; model-initiated tool calls above |
| 1638 | // still use the normal Permissions/Sandbox path. |
| 1639 | if a.planMode.Load() { |
| 1640 | return out |
| 1641 | } |
| 1642 | { |
| 1643 | incomplete, hasTodos := a.evidence.IncompleteLatestTodos() |
| 1644 | if !hasTodos && a.evidence.HasAnySuccessfulReceipt() { |
| 1645 | incomplete, hasTodos = a.incompleteCanonicalTodos() |
| 1646 | } |
| 1647 | if hasTodos && len(incomplete) > 0 && a.evidence.HasSuccessfulTodoProgressReceipt() { |
| 1648 | out.applies = true |
| 1649 | out.incompleteTodos = len(incomplete) |
| 1650 | missing = append(missing, finalReadinessIncompleteTodos(incomplete)) |
| 1651 | } |
| 1652 | } |
| 1653 | writer, hasWriter := a.evidence.LatestSuccessfulWriterIndex() |
| 1654 | deliveryMutation := false |
| 1655 | deliveryVerificationOnly := false |
| 1656 | checkpoint := a.deliveryCheckpoint |
| 1657 | checkpointApplies := a.deliveryScopeActive && checkpoint.ScopeID == a.deliveryScopeID |
| 1658 | if a.deliveryProfile { |
| 1659 | if mutation, ok := a.evidence.LatestSuccessfulMutationIndex(); ok { |
| 1660 | writer, hasWriter = mutation, true |
| 1661 | deliveryMutation = true |
| 1662 | } else if checkpointApplies && checkpoint.PendingMutation { |
| 1663 | // The mutation happened before a controller rebuild/restart. Treat it as |
| 1664 | // the baseline so this run can satisfy verification/review/sign-off |
| 1665 | // without manufacturing another write. |
| 1666 | writer, hasWriter = -1, true |
| 1667 | deliveryMutation = true |
| 1668 | } else if checkpointApplies && checkpoint.MutationObserved { |
| 1669 | deliveryMutation = true |
| 1670 | } |
| 1671 | workObserved := a.evidence.HasSuccessfulWorkReceipt() || (checkpointApplies && checkpoint.WorkObserved) |
| 1672 | if a.deliveryTaskExpected && !a.deliveryPersistentExpected && !workObserved { |
| 1673 | out.missingActionEvidence++ |
| 1674 | missing = append(missing, "perform host-observable work for this technical task before answering") |
| 1675 | } |
| 1676 | if a.deliveryPersistentExpected && !a.evidence.HasSuccessfulToolReceipt("remember") { |
| 1677 | out.missingMutation++ |
| 1678 | missing = append(missing, "save the requested durable memory with the remember tool before answering") |
| 1679 | } |
| 1680 | if a.deliveryMutationExpected && !deliveryMutation { |
| 1681 | out.missingMutation++ |
| 1682 | missing = append(missing, "the request requires a state change, but no successful mutation was observed") |
| 1683 | } |
| 1684 | if !hasWriter && a.evidence.HasSuccessfulVerificationCommand() { |
| 1685 | writer, hasWriter = -1, true |
| 1686 | deliveryVerificationOnly = true |
| 1687 | } |
| 1688 | // Required/preferred capability gates apply before the no-writer fast |
| 1689 | // path below: a user-required Skill/MCP must not be skippable by |
| 1690 | // answering from ordinary reads alone. |
| 1691 | if msg := a.capabilityGateFailure(); msg != "" { |
| 1692 | out.applies = true |
| 1693 | out.missingCapabilities++ |
| 1694 | missing = append(missing, msg) |
| 1695 | } |
| 1696 | if a.deliveryPersistentExpected && !a.deliveryMutationExpected && !a.evidence.HasSuccessfulMutationOtherThan("remember") { |
| 1697 | // A durable-memory-only request has its own concrete receipt contract. |
| 1698 | // It must not inherit code-delivery todo/test/diff/review ceremonies; |
| 1699 | // any unrelated mutation falls through to the full contract below. |
| 1700 | out.applies = true |
| 1701 | if len(missing) > 0 { |
| 1702 | out.reason = strings.Join(missing, "; ") |
| 1703 | } |
| 1704 | return out |
| 1705 | } |
| 1706 | } |
| 1707 | if !hasWriter { |
| 1708 | if len(missing) > 0 { |
| 1709 | if a.loopGuardAllowsFinal() { |
| 1710 | return out |
| 1711 | } |
| 1712 | out.reason = strings.Join(missing, "; ") |
| 1713 | } |
| 1714 | return out |
| 1715 | } |
| 1716 | hasProjectChecks := len(a.projectChecks) > 0 |
| 1717 | hasTodoReceipt := a.evidence.HasSuccessfulTodoWrite() |
| 1718 | if !a.deliveryProfile && !hasProjectChecks && !hasTodoReceipt && len(missing) == 0 { |
| 1719 | return finalReadinessCheck{} |
| 1720 | } |
| 1721 | out.applies = true |
| 1722 | if a.deliveryProfile { |
| 1723 | criteriaEstablished := a.deliveryCriteriaEstablished || (checkpointApplies && checkpoint.CriteriaEstablished) |
| 1724 | if !criteriaEstablished { |
| 1725 | out.missingAcceptanceCriteria++ |
| 1726 | missing = append(missing, "establish concrete acceptance criteria with todo_write before changing state") |
| 1727 | } |
| 1728 | hasCompleteStep := a.evidence.HasSuccessfulCompleteStepAfter(writer) |
| 1729 | if !hasCompleteStep { |
| 1730 | out.missingSignoff++ |
| 1731 | missing = append(missing, "call complete_step after the latest mutation") |
| 1732 | } |
| 1733 | if !a.evidence.HasSuccessfulDeliverySignoffAfter(writer) { |
| 1734 | out.missingVerification++ |
| 1735 | missing = append(missing, "run relevant verification after the latest mutation and cite that successful command in complete_step") |
| 1736 | } |
| 1737 | if deliveryMutation && !a.evidence.HasSuccessfulReviewAfter(writer) { |
| 1738 | out.missingReview++ |
| 1739 | missing = append(missing, "inspect the changed result after the latest mutation (read the touched file or run git diff/status)") |
| 1740 | } |
| 1741 | if msg := a.deliveryReviewGateFailure(); msg != "" { |
| 1742 | out.missingReview++ |
| 1743 | missing = append(missing, msg) |
| 1744 | } |
| 1745 | // The capability gate already ran before the no-writer fast path above. |
| 1746 | } |
| 1747 | for _, check := range a.projectChecks { |
| 1748 | if deliveryVerificationOnly { |
| 1749 | break |
| 1750 | } |
| 1751 | command := strings.TrimSpace(check.Command) |
| 1752 | if command == "" { |
| 1753 | continue |
| 1754 | } |
| 1755 | if !a.evidence.HasSuccessfulCommandAfter(command, writer) { |
| 1756 | out.missingProjectChecks++ |
| 1757 | missing = append(missing, fmt.Sprintf("run %q from %s after the latest write", command, finalReadinessCheckSource(check))) |
| 1758 | } |
| 1759 | } |
| 1760 | |
| 1761 | if len(missing) == 0 { |
| 1762 | return out |
| 1763 | } |
| 1764 | if a.loopGuardAllowsFinal() { |
| 1765 | return out |
| 1766 | } |
| 1767 | out.reason = strings.Join(missing, "; ") |
| 1768 | return out |
| 1769 | } |
| 1770 | |
| 1771 | // DeliveryCheckpoint returns the compact Goal-scoped delivery state. It is safe |
| 1772 | // to persist next to the Goal sidecar because it contains no raw arguments. |
| 1773 | func (a *Agent) DeliveryCheckpoint() evidence.DeliveryCheckpoint { |
| 1774 | return a.deliveryCheckpoint |
| 1775 | } |
| 1776 | |
| 1777 | // RestoreDeliveryCheckpoint seeds a rebuilt controller before its next Goal |
| 1778 | // run. A mismatched/empty scope is ignored conservatively. |
| 1779 | func (a *Agent) RestoreDeliveryCheckpoint(checkpoint evidence.DeliveryCheckpoint) { |
| 1780 | checkpoint.ScopeID = strings.TrimSpace(checkpoint.ScopeID) |
| 1781 | if checkpoint.ScopeID == "" { |
| 1782 | return |
| 1783 | } |
| 1784 | a.deliveryCheckpoint = checkpoint |
| 1785 | a.deliveryScopeID = checkpoint.ScopeID |
| 1786 | } |
| 1787 | |
| 1788 | // PrepareDeliveryRecovery preserves the exhausted turn's evidence for exactly |
| 1789 | // one explicit continuation. It returns false when there is no matching |
| 1790 | // readiness failure, so normal follow-up turns cannot inherit stale mutations. |
| 1791 | func (a *Agent) PrepareDeliveryRecovery() bool { |
| 1792 | if !a.deliveryProfile || !a.deliveryRecoveryPending { |
| 1793 | return false |
| 1794 | } |
| 1795 | a.preserveEvidenceOnce = true |
| 1796 | a.deliveryRecoveryPending = false |
| 1797 | return true |
| 1798 | } |
| 1799 | |
| 1800 | func (a *Agent) updateDeliveryCheckpoint(runErr error) { |
| 1801 | if !a.deliveryScopeActive || a.deliveryScopeID == "" || a.evidence == nil { |
| 1802 | return |
| 1803 | } |
| 1804 | cp := a.deliveryCheckpoint |
| 1805 | if cp.ScopeID != a.deliveryScopeID { |
| 1806 | cp = evidence.DeliveryCheckpoint{ScopeID: a.deliveryScopeID} |
| 1807 | } |
| 1808 | cp.CriteriaEstablished = cp.CriteriaEstablished || a.deliveryCriteriaEstablished || a.evidence.HasSuccessfulTodoWrite() |
| 1809 | cp.WorkObserved = cp.WorkObserved || a.evidence.HasSuccessfulWorkReceipt() |
| 1810 | persistentOnlyReady := a.deliveryPersistentExpected && !a.deliveryMutationExpected && |
| 1811 | a.evidence.HasSuccessfulToolReceipt("remember") && !a.evidence.HasSuccessfulMutationOtherThan("remember") |
| 1812 | if _, ok := a.evidence.LatestSuccessfulMutationIndex(); ok && !persistentOnlyReady { |
| 1813 | cp.MutationObserved = true |
| 1814 | cp.PendingMutation = true |
| 1815 | } |
| 1816 | if persistentOnlyReady { |
| 1817 | cp.MutationObserved = true |
| 1818 | } |
| 1819 | if runErr == nil && cp.PendingMutation && a.deliveryMutationCheckpointReady() { |
| 1820 | cp.PendingMutation = false |
| 1821 | } |
| 1822 | a.deliveryCheckpoint = cp |
| 1823 | } |
| 1824 | |
| 1825 | func (a *Agent) deliveryMutationCheckpointReady() bool { |
| 1826 | if a.evidence == nil || !a.deliveryCriteriaEstablished { |
| 1827 | return false |
| 1828 | } |
| 1829 | mutation, ok := a.evidence.LatestSuccessfulMutationIndex() |
| 1830 | if !ok { |
| 1831 | mutation = -1 |
| 1832 | } |
| 1833 | return a.evidence.HasSuccessfulCompleteStepAfter(mutation) && |
| 1834 | a.evidence.HasSuccessfulDeliverySignoffAfter(mutation) && |
| 1835 | a.evidence.HasSuccessfulReviewAfter(mutation) && |
| 1836 | a.deliveryReviewGateFailure() == "" |
| 1837 | } |
| 1838 | |
| 1839 | // armLoopGuardPass records that a loop guard fired this user turn. |
| 1840 | // receiptMark is the evidence-ledger receipt count from just before the |
| 1841 | // guarded batch ran, so a successful write or command receipt recorded after |
| 1842 | // it counts as real progress and revokes the pass (see loopGuardAllowsFinal). |
| 1843 | func (a *Agent) armLoopGuardPass(receiptMark int) { |
| 1844 | a.loopGuardArmed = true |
| 1845 | a.loopGuardReceiptMark = receiptMark |
| 1846 | } |
| 1847 | |
| 1848 | // loopGuardAllowsFinal reports whether final readiness should stand down: a |
| 1849 | // loop guard fired this user turn and no host-observable progress — a |
| 1850 | // successful write or command receipt — has landed since. In that state the |
| 1851 | // missing receipts are exactly what the blocker prevents, so demanding them |
| 1852 | // would restart the retry loop the guard just broke; the model must be free to |
| 1853 | // report the blocker instead. The bookkeeping the guard recommends (ask, |
| 1854 | // todo_write, complete_step) produces neither write nor command receipts, so |
| 1855 | // it keeps the pass; real progress revokes it because receipts are obtainable |
| 1856 | // again and readiness should resume enforcing them. |
| 1857 | func (a *Agent) loopGuardAllowsFinal() bool { |
| 1858 | if a == nil || !a.loopGuardArmed { |
| 1859 | return false |
| 1860 | } |
| 1861 | if a.evidence == nil { |
| 1862 | return true |
| 1863 | } |
| 1864 | return !a.evidence.HasWriteOrCommandSince(a.loopGuardReceiptMark) |
| 1865 | } |
| 1866 | |
| 1867 | func finalReadinessIncompleteTodos(items []evidence.TodoStepMatch) string { |
| 1868 | parts := make([]string, 0, len(items)) |
| 1869 | for _, item := range items { |
| 1870 | label := strings.TrimSpace(item.Content) |
| 1871 | if label == "" { |
| 1872 | label = fmt.Sprintf("todo %d", item.Index) |
| 1873 | } |
| 1874 | parts = append(parts, fmt.Sprintf("%s: %s", label, item.Status)) |
| 1875 | } |
| 1876 | return "latest successful todo_write still has incomplete items: " + strings.Join(parts, ", ") |
| 1877 | } |
| 1878 | |
| 1879 | func (a *Agent) setTodoState(todos []evidence.TodoItem) { |
| 1880 | a.todoMu.Lock() |
| 1881 | a.todoState = evidence.NormalizeSerialTodos(todos) |
| 1882 | a.todoMu.Unlock() |
| 1883 | } |
| 1884 | |
| 1885 | // SeedTodoState initializes the canonical task list from a host-generated |
| 1886 | // starter list, such as an approved plan. A new host seed replaces stale state |
| 1887 | // from earlier work so complete_step matches the plan the UI just displayed. |
| 1888 | func (a *Agent) SeedTodoState(todos []evidence.TodoItem) { |
| 1889 | if len(todos) == 0 { |
| 1890 | return |
| 1891 | } |
| 1892 | a.setTodoState(todos) |
| 1893 | } |
| 1894 | |
| 1895 | // ReplaceTodoState mirrors a host-generated todo list into the canonical state. |
| 1896 | // It is used when the host, rather than the model, owns the full state transition. |
| 1897 | func (a *Agent) ReplaceTodoState(todos []evidence.TodoItem) { |
| 1898 | a.setTodoState(todos) |
| 1899 | a.recordTodoState(a.CanonicalTodoState()) |
| 1900 | } |
| 1901 | |
| 1902 | // CanonicalTodoState returns a copy of the host-reconstructed task list. |
| 1903 | func (a *Agent) CanonicalTodoState() []evidence.TodoItem { |
| 1904 | a.todoMu.Lock() |
| 1905 | defer a.todoMu.Unlock() |
| 1906 | return append([]evidence.TodoItem(nil), a.todoState...) |
| 1907 | } |
| 1908 | |
| 1909 | func (a *Agent) incompleteCanonicalTodos() ([]evidence.TodoStepMatch, bool) { |
| 1910 | a.todoMu.Lock() |
| 1911 | defer a.todoMu.Unlock() |
| 1912 | if len(a.todoState) == 0 { |
| 1913 | return nil, false |
| 1914 | } |
| 1915 | return evidence.IncompleteTodos(a.todoState), true |
| 1916 | } |
| 1917 | |
| 1918 | func (a *Agent) hasIncompleteCanonicalCriteria() bool { |
| 1919 | a.todoMu.Lock() |
| 1920 | defer a.todoMu.Unlock() |
| 1921 | return len(a.todoState) > 0 && len(evidence.IncompleteTodos(a.todoState)) > 0 |
| 1922 | } |
| 1923 | |
| 1924 | func (a *Agent) hasActiveCanonicalTodo() bool { |
| 1925 | a.todoMu.Lock() |
| 1926 | defer a.todoMu.Unlock() |
| 1927 | for _, todo := range a.todoState { |
| 1928 | if canonicalTodoStatus(todo.Status) == "in_progress" { |
| 1929 | return true |
| 1930 | } |
| 1931 | } |
| 1932 | return false |
| 1933 | } |
| 1934 | |
| 1935 | func (a *Agent) canonicalTodoProgress() (int, bool) { |
| 1936 | a.todoMu.Lock() |
| 1937 | defer a.todoMu.Unlock() |
| 1938 | completed := 0 |
| 1939 | incomplete := false |
| 1940 | for _, todo := range a.todoState { |
| 1941 | status := canonicalTodoStatus(todo.Status) |
| 1942 | if status == "completed" { |
| 1943 | completed++ |
| 1944 | } else { |
| 1945 | incomplete = true |
| 1946 | } |
| 1947 | } |
| 1948 | return completed, incomplete |
| 1949 | } |
| 1950 | |
| 1951 | // registryHasWriterTools reports whether any registered tool can mutate state. |
| 1952 | // A strictly read-only registry (read_only_task / read_only_skill subagents) |
| 1953 | // can never satisfy a "state change required" delivery expectation, so that |
| 1954 | // expectation must not be armed for it. |
| 1955 | func registryHasWriterTools(reg *tool.Registry) bool { |
| 1956 | if reg == nil { |
| 1957 | return false |
| 1958 | } |
| 1959 | for _, name := range reg.Names() { |
| 1960 | if t, ok := reg.Get(name); ok && !t.ReadOnly() { |
| 1961 | return true |
| 1962 | } |
| 1963 | } |
| 1964 | return false |
| 1965 | } |
| 1966 | |
| 1967 | type deliveryTaskIntent uint8 |
| 1968 | |
| 1969 | const ( |
| 1970 | deliveryIntentConversation deliveryTaskIntent = iota |
| 1971 | deliveryIntentAdvisory |
| 1972 | deliveryIntentObservableRead |
| 1973 | deliveryIntentMutation |
| 1974 | deliveryIntentPersistentAction |
| 1975 | ) |
| 1976 | |
| 1977 | func classifyDeliveryTaskIntent(input string) deliveryTaskIntent { |
| 1978 | switch { |
| 1979 | case deliveryTaskHasMutationIntent(input): |
| 1980 | return deliveryIntentMutation |
| 1981 | case deliveryTaskNeedsPersistentAction(input): |
| 1982 | return deliveryIntentPersistentAction |
| 1983 | case deliveryTaskIsConversationOnly(input): |
| 1984 | return deliveryIntentConversation |
| 1985 | case !heuristicInputIsTask(input): |
| 1986 | return deliveryIntentConversation |
| 1987 | case deliveryTaskIsAdvisory(input): |
| 1988 | return deliveryIntentAdvisory |
| 1989 | default: |
| 1990 | return deliveryIntentObservableRead |
| 1991 | } |
| 1992 | } |
| 1993 | |
| 1994 | func deliveryTaskNeedsEvidence(input string) bool { |
| 1995 | intent := classifyDeliveryTaskIntent(input) |
| 1996 | return intent == deliveryIntentObservableRead || intent == deliveryIntentMutation || intent == deliveryIntentPersistentAction |
| 1997 | } |
| 1998 | |
| 1999 | var deliveryMutationNeedles = []string{ |
| 2000 | "fix", "repair", "resolve", "create", "add", "write", "edit", "update", "change", "delete", "remove", "rename", |
| 2001 | "implement", "refactor", "apply", "install", "publish", "commit", "push", "continue work", |
| 2002 | "modify", "patch", "replace", "move", "configure", "upgrade", "downgrade", "bump", "enable", "disable", "merge", |
| 2003 | "make changes", "make a change", "make the changes", "make the requested changes", "make the necessary changes", "make these changes", "make those changes", "make code changes", |
| 2004 | "修复", "解决", "创建", "新建", "添加", "编写", "编辑", "修改", "更新", "删除", "移除", "重命名", "实现", "重构", |
| 2005 | "实施", "落地", "安装", "发布", "提交", "继续处理", "调整", "替换", "移动", "升级", "降级", "启用", "禁用", "合并", "改动", "打补丁", |
| 2006 | } |
| 2007 | |
| 2008 | var deliveryAdvisoryPhrases = []string{ |
| 2009 | "what's wrong", "what is wrong", "why", "what should i do", "what can i do", "how should i", "how do i", "how can i", |
| 2010 | "can you explain", "could you explain", "give me advice", "any advice", "help me understand", |
| 2011 | "为什么", "怎么回事", "怎么办", "怎么", "怎样", "如何", "是什么问题", "什么原因", "的原因", "给我建议", "有什么建议", |
| 2012 | } |
| 2013 | |
| 2014 | func deliveryTaskNeedsMutation(input string) bool { |
| 2015 | intent := classifyDeliveryTaskIntent(input) |
| 2016 | return intent == deliveryIntentMutation || intent == deliveryIntentPersistentAction |
| 2017 | } |
| 2018 | |
| 2019 | func deliveryTaskHasMutationIntent(input string) bool { |
| 2020 | affirmative, _ := deliveryTaskMutationIntent(input) |
| 2021 | return affirmative |
| 2022 | } |
| 2023 | |
| 2024 | func deliveryTaskNeedsPersistentAction(input string) bool { |
| 2025 | normalized := strings.ToLower(strings.TrimSpace(input)) |
| 2026 | if normalized == "" { |
| 2027 | return false |
| 2028 | } |
| 2029 | actionNeedles := []string{ |
| 2030 | "remember", "save", "store", "keep this", "keep that", |
| 2031 | "记住", "记下来", "保存", "存下来", "记录下来", |
| 2032 | } |
| 2033 | durableNeedles := []string{ |
| 2034 | "permanently", "durable", "long-term", "long term", "across sessions", "future sessions", "every session", "after restart", "after restarting", |
| 2035 | "永久", "长期", "持久", "跨会话", "以后每次", "未来会话", "重启后", "下次启动", |
| 2036 | } |
| 2037 | for _, clause := range deliveryTaskClauses(normalized) { |
| 2038 | action := false |
| 2039 | for _, needle := range actionNeedles { |
| 2040 | affirmative, _ := deliveryTaskNeedleIntent(clause, needle) |
| 2041 | action = action || affirmative |
| 2042 | } |
| 2043 | durable := false |
| 2044 | for _, needle := range durableNeedles { |
| 2045 | affirmative, _ := deliveryTaskNeedleIntent(clause, needle) |
| 2046 | durable = durable || affirmative |
| 2047 | } |
| 2048 | if action && durable && !deliveryTaskClauseIsAdvisory(clause) { |
| 2049 | return true |
| 2050 | } |
| 2051 | } |
| 2052 | return false |
| 2053 | } |
| 2054 | |
| 2055 | func deliveryTaskIsConversationOnly(input string) bool { |
| 2056 | normalized := strings.ToLower(strings.TrimSpace(input)) |
| 2057 | if normalized == "" || deliveryTaskHasHostAnchor(normalized) || deliveryTaskHasCommand(normalized) { |
| 2058 | return false |
| 2059 | } |
| 2060 | localCue := containsAnySubstring(normalized, []string{ |
| 2061 | "next turn", "next message", "later in this chat", "this conversation", "when i ask again", "when i ask next", |
| 2062 | "下一轮", "下轮", "下一条消息", "稍后再问", "待会再问", "这个对话", "本次对话", "本轮会话", |
| 2063 | }) |
| 2064 | conversationAction := containsAnySubstring(normalized, []string{ |
| 2065 | "remember", "keep in mind", "keep this", "keep that", "answer", "respond", "reply", |
| 2066 | "记住", "记一下", "回答", "回复", "再告诉我", |
| 2067 | }) |
| 2068 | return localCue && conversationAction |
| 2069 | } |
| 2070 | |
| 2071 | // TaskNeedsMutation reports whether a task text looks like a mutation request |
| 2072 | // under the existing task-intent classification. The host uses it to pick a |
| 2073 | // Goal budget class; it never gates permissions or whether writes are allowed. |
| 2074 | func TaskNeedsMutation(input string) bool { |
| 2075 | return deliveryTaskNeedsMutation(input) |
| 2076 | } |
| 2077 | |
| 2078 | func deliveryTaskMutationIntent(input string) (affirmative, negated bool) { |
| 2079 | normalized := strings.ToLower(strings.TrimSpace(input)) |
| 2080 | for _, clause := range deliveryTaskClauses(normalized) { |
| 2081 | clauseAffirmative := false |
| 2082 | clauseNegated := false |
| 2083 | if deliveryMutationClauseNegated(clause) { |
| 2084 | clauseNegated = true |
| 2085 | } |
| 2086 | for _, needle := range deliveryMutationNeedles { |
| 2087 | hasAffirmative, hasNegated := deliveryTaskNeedleIntent(clause, needle) |
| 2088 | clauseAffirmative = clauseAffirmative || hasAffirmative |
| 2089 | clauseNegated = clauseNegated || hasNegated |
| 2090 | } |
| 2091 | if clauseAffirmative && deliveryTaskClauseIsAdvisory(clause) && !deliveryTaskAdvisoryClauseRequestsMutation(clause) { |
| 2092 | clauseAffirmative = false |
| 2093 | clauseNegated = true |
| 2094 | } |
| 2095 | affirmative = affirmative || clauseAffirmative |
| 2096 | negated = negated || clauseNegated |
| 2097 | } |
| 2098 | return affirmative, negated |
| 2099 | } |
| 2100 | |
| 2101 | func deliveryTaskIsAdvisory(input string) bool { |
| 2102 | normalized := strings.ToLower(strings.TrimSpace(input)) |
| 2103 | |
| 2104 | // Concrete targets and commands always remain host-observable, including |
| 2105 | // when the request is phrased as a "why" question. |
| 2106 | if deliveryTaskHasHostAnchor(normalized) || deliveryTaskHasCommand(normalized) { |
| 2107 | return false |
| 2108 | } |
| 2109 | |
| 2110 | // Question wording is scoped per clause. This keeps remote troubleshooting |
| 2111 | // such as "analyze why WPS won't open" advisory, while a separate imperative |
| 2112 | // clause such as "reproduce the crash" still requires observable work. |
| 2113 | sawAdvisory := false |
| 2114 | for _, clause := range deliveryTaskClauses(normalized) { |
| 2115 | if deliveryTaskClauseIsAdvisory(clause) { |
| 2116 | sawAdvisory = true |
| 2117 | continue |
| 2118 | } |
| 2119 | if deliveryTaskClauseHasObservableWork(clause) { |
| 2120 | return false |
| 2121 | } |
| 2122 | } |
| 2123 | if sawAdvisory { |
| 2124 | return true |
| 2125 | } |
| 2126 | |
| 2127 | // A standalone refusal, inability, or constraint around a mutation verb is |
| 2128 | // advisory rather than work Reasonix can perform. Affirmative mixed intent is |
| 2129 | // handled by deliveryTaskNeedsMutation before this function is consulted. |
| 2130 | _, negatedMutation := deliveryTaskMutationIntent(normalized) |
| 2131 | return negatedMutation |
| 2132 | } |
| 2133 | |
| 2134 | func deliveryTaskHasHostAnchor(input string) bool { |
| 2135 | for _, anchor := range []string{ |
| 2136 | "this repo", "this repository", "current repository", "codebase", "workspace", "pull request", "this pr", "ci job", |
| 2137 | "/pull/", "actions/runs/", |
| 2138 | "当前仓库", "这个仓库", "当前项目", "这个项目", "代码库", "工作区", "这个 pr", "这个pr", "此 pr", "此pr", |
| 2139 | } { |
| 2140 | if strings.Contains(input, anchor) { |
| 2141 | return true |
| 2142 | } |
| 2143 | } |
| 2144 | return deliveryTaskHasFileReference(input) |
| 2145 | } |
| 2146 | |
| 2147 | func deliveryTaskHasFileReference(input string) bool { |
| 2148 | previous := rune(0) |
| 2149 | for index, current := range input { |
| 2150 | if current == '@' && index+1 < len(input) && |
| 2151 | (index == 0 || strings.ContainsRune(" \t\r\n([{<,:;（【《，。；：", previous)) { |
| 2152 | next, _ := utf8.DecodeRuneInString(input[index+1:]) |
| 2153 | if !strings.ContainsRune(" \t\r\n", next) { |
| 2154 | return true |
| 2155 | } |
| 2156 | } |
| 2157 | previous = current |
| 2158 | } |
| 2159 | |
| 2160 | for _, raw := range strings.FieldsFunc(input, func(r rune) bool { |
| 2161 | switch r { |
| 2162 | case ' ', '\t', '\r', '\n', '`', '\'', '"', '(', ')', '[', ']', '{', '}', '<', '>', ',', '，', ';', '；', '!', '！', '?', '？': |
| 2163 | return true |
| 2164 | default: |
| 2165 | return false |
| 2166 | } |
| 2167 | }) { |
| 2168 | token := strings.ToLower(strings.TrimSpace(raw)) |
| 2169 | if token == "" || strings.Contains(token, "://") { |
| 2170 | continue |
| 2171 | } |
| 2172 | if strings.HasPrefix(token, "./") || strings.HasPrefix(token, "../") || |
| 2173 | strings.HasPrefix(token, "/") || strings.Contains(token, `\`) { |
| 2174 | return true |
| 2175 | } |
| 2176 | base := token |
| 2177 | if slash := strings.LastIndexByte(base, '/'); slash >= 0 { |
| 2178 | base = base[slash+1:] |
| 2179 | } |
| 2180 | switch base { |
| 2181 | case "dockerfile", "makefile", "cmakelists.txt", "justfile", "license", "readme", "changelog": |
| 2182 | return true |
| 2183 | } |
| 2184 | dot := strings.LastIndexByte(base, '.') |
| 2185 | if dot < 0 { |
| 2186 | continue |
| 2187 | } |
| 2188 | switch base[dot:] { |
| 2189 | case ".go", ".mod", ".sum", ".js", ".jsx", ".ts", ".tsx", ".py", ".rs", ".java", ".kt", ".swift", |
| 2190 | ".c", ".cc", ".cpp", ".h", ".hpp", ".cs", ".rb", ".php", ".sh", ".zsh", ".fish", ".ps1", |
| 2191 | ".md", ".json", ".yaml", ".yml", ".toml", ".xml", ".sql", ".proto", ".html", ".css", ".scss", |
| 2192 | ".vue", ".svelte", ".txt", ".log", ".csv", ".pdf", ".env", ".ini", ".conf", ".lock": |
| 2193 | return true |
| 2194 | } |
| 2195 | } |
| 2196 | return false |
| 2197 | } |
| 2198 | |
| 2199 | func deliveryTaskHasCommand(input string) bool { |
| 2200 | tokens := strings.FieldsFunc(strings.ToLower(input), func(r rune) bool { |
| 2201 | asciiWord := r >= 'a' && r <= 'z' || r >= '0' && r <= '9' |
| 2202 | return !asciiWord && r != '_' && r != '-' && r != '.' && r != '/' && r != '\\' && r != ':' |
| 2203 | }) |
| 2204 | for i := range tokens { |
| 2205 | if deliveryCommandStartsAt(tokens, i) { |
| 2206 | return true |
| 2207 | } |
| 2208 | } |
| 2209 | return false |
| 2210 | } |
| 2211 | |
| 2212 | func deliveryCommandStartsAt(tokens []string, index int) bool { |
| 2213 | command := strings.TrimSpace(tokens[index]) |
| 2214 | if command == "" { |
| 2215 | return false |
| 2216 | } |
| 2217 | if strings.HasPrefix(command, "./") || strings.HasPrefix(command, "../") || |
| 2218 | strings.HasPrefix(command, "/") || strings.Contains(command, `\`) { |
| 2219 | return true |
| 2220 | } |
| 2221 | next := "" |
| 2222 | if index+1 < len(tokens) { |
| 2223 | next = tokens[index+1] |
| 2224 | } |
| 2225 | if next != "--" && len(next) > 1 && strings.HasPrefix(next, "-") { |
| 2226 | return true |
| 2227 | } |
| 2228 | previous := "" |
| 2229 | if index > 0 { |
| 2230 | previous = tokens[index-1] |
| 2231 | } |
| 2232 | switch command { |
| 2233 | case "go": |
| 2234 | switch next { |
| 2235 | case "build", "clean", "doc", "env", "fmt", "generate", "get", "install", "list", "mod", "run", "test", "tool", "version", "vet", "work": |
| 2236 | return true |
| 2237 | } |
| 2238 | case "git", "npm", "npx", "pnpm", "yarn", "bun", "deno", "cargo", "rustc", "python", "python3", |
| 2239 | "bash", "sh", "zsh", "fish", "powershell", "pwsh", "docker", "docker-compose", "kubectl", "helm", "terraform", |
| 2240 | "gradle", "gradlew", "mvn", "dotnet", "xcodebuild", "gcc", "g++", "clang", "clang++": |
| 2241 | return deliveryCommandHasExplicitCue(previous) || deliveryCommandHasSubcommand(next) |
| 2242 | case "node": |
| 2243 | return deliveryCommandHasExplicitCue(previous) || next == "inspect" || next == "test" |
| 2244 | case "swift": |
| 2245 | return next == "build" || next == "package" || next == "run" || next == "test" |
| 2246 | case "make", "just": |
| 2247 | switch next { |
| 2248 | case "all", "build", "check", "clean", "fail", "failed", "failing", "install", "lint", "test": |
| 2249 | return true |
| 2250 | } |
| 2251 | case "pytest", "cmake", "ninja", "eslint", "tsc", "vitest", "jest": |
| 2252 | return deliveryCommandHasExplicitCue(previous) || next == "fail" || next == "failed" || next == "failing" |
| 2253 | } |
| 2254 | return false |
| 2255 | } |
| 2256 | |
| 2257 | func deliveryCommandHasExplicitCue(previous string) bool { |
| 2258 | switch previous { |
| 2259 | case "command", "execute", "executing", "run", "running", "using", "with": |
| 2260 | return true |
| 2261 | default: |
| 2262 | return false |
| 2263 | } |
| 2264 | } |
| 2265 | |
| 2266 | func deliveryCommandHasSubcommand(next string) bool { |
| 2267 | switch next { |
| 2268 | case "add", "apply", "branch", "build", "check", "checkout", "clean", "clone", "commit", "config", "container", |
| 2269 | "deploy", "describe", "destroy", "dev", "diff", "down", "env", "exec", "fetch", "fmt", "generate", "get", "image", |
| 2270 | "init", "install", "lint", "list", "log", "logs", "login", "logout", "merge", "mod", "package", "plan", "ps", "publish", |
| 2271 | "pull", "push", "rebase", "remote", "remove", "reset", "restore", "run", "serve", "show", "start", "stash", "status", |
| 2272 | "switch", "tag", "test", "tool", "uninstall", "up", "update", "upgrade", "version", "vet", "work", "worktree": |
| 2273 | return true |
| 2274 | default: |
| 2275 | return false |
| 2276 | } |
| 2277 | } |
| 2278 | |
| 2279 | func deliveryTaskClauseHasObservableWork(clause string) bool { |
| 2280 | for _, needle := range []string{ |
| 2281 | "review", "inspect", "analyze", "check", "reproduce", "audit", "verify", |
| 2282 | "评审", "审查", "检查", "分析", "复现", "审计", "验证", |
| 2283 | } { |
| 2284 | affirmative, _ := deliveryTaskNeedleIntent(clause, needle) |
| 2285 | if affirmative { |
| 2286 | return true |
| 2287 | } |
| 2288 | } |
| 2289 | return false |
| 2290 | } |
| 2291 | |
| 2292 | func deliveryTaskClauseIsAdvisory(clause string) bool { |
| 2293 | for _, phrase := range deliveryAdvisoryPhrases { |
| 2294 | if strings.Contains(clause, phrase) { |
| 2295 | return true |
| 2296 | } |
| 2297 | } |
| 2298 | return false |
| 2299 | } |
| 2300 | |
| 2301 | func deliveryTaskAdvisoryClauseRequestsMutation(clause string) bool { |
| 2302 | advisoryIndex := len(clause) |
| 2303 | for _, phrase := range deliveryAdvisoryPhrases { |
| 2304 | if index := strings.Index(clause, phrase); index >= 0 && index < advisoryIndex { |
| 2305 | advisoryIndex = index |
| 2306 | } |
| 2307 | } |
| 2308 | if advisoryIndex == len(clause) { |
| 2309 | return false |
| 2310 | } |
| 2311 | if deliveryTaskStartsWithMutation(clause[:advisoryIndex]) { |
| 2312 | return true |
| 2313 | } |
| 2314 | |
| 2315 | for _, cue := range []string{" please ", " then ", " so ", " therefore ", "然后", "所以", "而是", "转而"} { |
| 2316 | for rest := clause[advisoryIndex:]; ; { |
| 2317 | index := strings.Index(rest, cue) |
| 2318 | if index < 0 { |
| 2319 | break |
| 2320 | } |
| 2321 | rest = rest[index+len(cue):] |
| 2322 | if deliveryTaskStartsWithMutation(rest) { |
| 2323 | return true |
| 2324 | } |
| 2325 | } |
| 2326 | } |
| 2327 | for rest, offset := clause[advisoryIndex:], advisoryIndex; ; { |
| 2328 | index := strings.Index(rest, "请") |
| 2329 | if index < 0 { |
| 2330 | break |
| 2331 | } |
| 2332 | absolute := offset + index |
| 2333 | after := clause[absolute+len("请"):] |
| 2334 | requestWord := strings.HasSuffix(clause[:absolute], "申") || strings.HasPrefix(after, "求") |
| 2335 | if !requestWord && deliveryTaskStartsWithMutation(after) { |
| 2336 | return true |
| 2337 | } |
| 2338 | offset = absolute + len("请") |
| 2339 | rest = clause[offset:] |
| 2340 | } |
| 2341 | |
| 2342 | for _, cue := range []string{" and ", "并且", "并"} { |
| 2343 | if index := strings.LastIndex(clause[advisoryIndex:], cue); index >= 0 { |
| 2344 | cueStart := advisoryIndex + index |
| 2345 | tail := clause[cueStart+len(cue):] |
| 2346 | if !deliveryTaskClauseHasNegation(clause[:cueStart]) && deliveryTaskStartsWithMutation(tail) { |
| 2347 | return true |
| 2348 | } |
| 2349 | } |
| 2350 | } |
| 2351 | return false |
| 2352 | } |
| 2353 | |
| 2354 | func deliveryTaskStartsWithMutation(input string) bool { |
| 2355 | input = strings.TrimSpace(input) |
| 2356 | for { |
| 2357 | stripped := false |
| 2358 | for _, prefix := range []string{"please ", "can you ", "could you ", "would you ", "you should ", "帮我", "请你", "直接", "继续", "再"} { |
| 2359 | if strings.HasPrefix(input, prefix) { |
| 2360 | input = strings.TrimSpace(strings.TrimPrefix(input, prefix)) |
| 2361 | stripped = true |
| 2362 | break |
| 2363 | } |
| 2364 | } |
| 2365 | if !stripped { |
| 2366 | break |
| 2367 | } |
| 2368 | } |
| 2369 | for _, needle := range deliveryMutationNeedles { |
| 2370 | if containsTaskNeedle(input, needle) { |
| 2371 | if containsNonASCII(needle) { |
| 2372 | return strings.HasPrefix(input, needle) |
| 2373 | } |
| 2374 | tokens := strings.FieldsFunc(input, func(r rune) bool { |
| 2375 | return !(r >= 'a' && r <= 'z') && !(r >= '0' && r <= '9') && r != '_' && r != '\'' |
| 2376 | }) |
| 2377 | needleTokens := strings.Fields(needle) |
| 2378 | if len(tokens) >= len(needleTokens) { |
| 2379 | matches := true |
| 2380 | for i := range needleTokens { |
| 2381 | matches = matches && tokens[i] == needleTokens[i] |
| 2382 | } |
| 2383 | if matches { |
| 2384 | return true |
| 2385 | } |
| 2386 | } |
| 2387 | } |
| 2388 | } |
| 2389 | return false |
| 2390 | } |
| 2391 | |
| 2392 | func deliveryTaskClauseHasNegation(clause string) bool { |
| 2393 | clause = strings.ReplaceAll(clause, "’", "'") |
| 2394 | for _, phrase := range []string{ |
| 2395 | " not ", " never ", " without ", "cannot", "can't", " cant ", "don't", " dont ", "won't", " wont ", "unable", |
| 2396 | "不要", "别", "勿", "不能", "无法", "不想", "不敢", "无需", "不需要", "不可", "没法", "没有", "禁止", "拒绝", |
| 2397 | } { |
| 2398 | if strings.Contains(" "+clause+" ", phrase) { |
| 2399 | return true |
| 2400 | } |
| 2401 | } |
| 2402 | return false |
| 2403 | } |
| 2404 | |
| 2405 | func deliveryTaskClauses(input string) []string { |
| 2406 | input = strings.NewReplacer( |
| 2407 | " but ", "\n", |
| 2408 | " however ", "\n", |
| 2409 | " nevertheless ", "\n", |
| 2410 | "但请", "\n请", |
| 2411 | "但是", "\n", |
| 2412 | "不过", "\n", |
| 2413 | ).Replace(input) |
| 2414 | return strings.FieldsFunc(input, func(r rune) bool { |
| 2415 | switch r { |
| 2416 | case '\n', '\r', '.', '。', ',', '，', ';', '；', '!', '！', '?', '？': |
| 2417 | return true |
| 2418 | default: |
| 2419 | return false |
| 2420 | } |
| 2421 | }) |
| 2422 | } |
| 2423 | |
| 2424 | func deliveryMutationClauseNegated(clause string) bool { |
| 2425 | for _, phrase := range []string{ |
| 2426 | "without changing", "without modifying", "analysis only", "review only", |
| 2427 | "不要改动", "只分析", "仅分析", "只检查", "仅检查", "只评审", "仅评审", |
| 2428 | } { |
| 2429 | if strings.Contains(clause, phrase) { |
| 2430 | return true |
| 2431 | } |
| 2432 | } |
| 2433 | return false |
| 2434 | } |
| 2435 | |
| 2436 | func deliveryTaskNeedleIntent(clause, needle string) (affirmative, negated bool) { |
| 2437 | if containsNonASCII(needle) { |
| 2438 | for offset := 0; offset < len(clause); { |
| 2439 | relative := strings.Index(clause[offset:], needle) |
| 2440 | if relative < 0 { |
| 2441 | break |
| 2442 | } |
| 2443 | index := offset + relative |
| 2444 | prefix := []rune(clause[:index]) |
| 2445 | if deliveryMutationRunesNegated(prefix) { |
| 2446 | negated = true |
| 2447 | } else { |
| 2448 | affirmative = true |
| 2449 | } |
| 2450 | offset = index + len(needle) |
| 2451 | } |
| 2452 | return affirmative, negated |
| 2453 | } |
| 2454 | |
| 2455 | clause = strings.ReplaceAll(clause, "’", "'") |
| 2456 | tokens := strings.FieldsFunc(clause, func(r rune) bool { |
| 2457 | return !(r >= 'a' && r <= 'z') && !(r >= '0' && r <= '9') && r != '_' && r != '\'' |
| 2458 | }) |
| 2459 | needleTokens := strings.Fields(needle) |
| 2460 | for i := 0; i+len(needleTokens) <= len(tokens); i++ { |
| 2461 | matches := true |
| 2462 | for j, token := range needleTokens { |
| 2463 | if tokens[i+j] != token { |
| 2464 | matches = false |
| 2465 | break |
| 2466 | } |
| 2467 | } |
| 2468 | if !matches { |
| 2469 | continue |
| 2470 | } |
| 2471 | if deliveryMutationTokensNegated(tokens[:i]) { |
| 2472 | negated = true |
| 2473 | } else { |
| 2474 | affirmative = true |
| 2475 | } |
| 2476 | } |
| 2477 | return affirmative, negated |
| 2478 | } |
| 2479 | |
| 2480 | func deliveryMutationTokensNegated(prefix []string) bool { |
| 2481 | if len(prefix) > 6 { |
| 2482 | prefix = prefix[len(prefix)-6:] |
| 2483 | } |
| 2484 | boundary := -1 |
| 2485 | for i, token := range prefix { |
| 2486 | switch token { |
| 2487 | case "but", "however", "nevertheless", "instead", "so", "then", "therefore", "please": |
| 2488 | boundary = i |
| 2489 | } |
| 2490 | } |
| 2491 | if boundary >= 0 { |
| 2492 | prefix = prefix[boundary+1:] |
| 2493 | } |
| 2494 | for i, token := range prefix { |
| 2495 | if token == "not" && i+1 < len(prefix) && prefix[i+1] == "only" { |
| 2496 | continue |
| 2497 | } |
| 2498 | switch token { |
| 2499 | case "not", "never", "without", "cannot", "can't", "cant", "don't", "dont", "won't", "wont", "unable", "avoid", "avoiding", "afraid", "refuse", "refusing", "needn't": |
| 2500 | return true |
| 2501 | case "no": |
| 2502 | if i+1 < len(prefix) && prefix[i+1] == "need" { |
| 2503 | return true |
| 2504 | } |
| 2505 | } |
| 2506 | } |
| 2507 | return false |
| 2508 | } |
| 2509 | |
| 2510 | func deliveryMutationRunesNegated(prefix []rune) bool { |
| 2511 | if len(prefix) > 12 { |
| 2512 | prefix = prefix[len(prefix)-12:] |
| 2513 | } |
| 2514 | window := string(prefix) |
| 2515 | scopeStart := 0 |
| 2516 | for _, boundary := range []string{"所以", "然后", "而是", "转而", "改为"} { |
| 2517 | if index := strings.LastIndex(window, boundary); index >= 0 { |
| 2518 | end := index + len(boundary) |
| 2519 | if end > scopeStart { |
| 2520 | scopeStart = end |
| 2521 | } |
| 2522 | } |
| 2523 | } |
| 2524 | if index := strings.LastIndex(window, "请"); index >= 0 { |
| 2525 | before, after := window[:index], window[index+len("请"):] |
| 2526 | requestWord := strings.HasSuffix(before, "申") || strings.HasPrefix(after, "求") |
| 2527 | negatedRequest := false |
| 2528 | for _, marker := range []string{"不要", "不能", "无法", "不想", "不敢", "无需", "不需要", "不可", "没法", "禁止", "拒绝"} { |
| 2529 | if strings.HasSuffix(before, marker) || strings.Contains(after, marker) { |
| 2530 | negatedRequest = true |
| 2531 | break |
| 2532 | } |
| 2533 | } |
| 2534 | if !requestWord && !negatedRequest && index+len("请") > scopeStart { |
| 2535 | scopeStart = index + len("请") |
| 2536 | } |
| 2537 | } |
| 2538 | window = window[scopeStart:] |
| 2539 | for _, marker := range []string{"不要", "别", "勿", "不能", "无法", "不想", "不敢", "无需", "不需要", "不可", "没法", "没有", "禁止", "拒绝"} { |
| 2540 | if strings.Contains(window, marker) { |
| 2541 | return true |
| 2542 | } |
| 2543 | } |
| 2544 | return false |
| 2545 | } |
| 2546 | |
| 2547 | // advanceCanonicalTodo flips the canonical todo matching a signed-off step to |
| 2548 | // completed (promoting the next pending item to in_progress) and emits a |
| 2549 | // synthetic todo_write so the task panel reflects it without the model |
| 2550 | // re-sending the whole list. No-op when nothing matches or it is already done. |
| 2551 | func (a *Agent) advanceCanonicalTodo(step string) { |
| 2552 | a.todoMu.Lock() |
| 2553 | if len(a.todoState) == 0 { |
| 2554 | a.todoMu.Unlock() |
| 2555 | return |
| 2556 | } |
| 2557 | m, ok := evidence.MatchStep(step, a.todoState) |
| 2558 | if !ok || !evidence.AdvanceSerialTodo(a.todoState, m.Index-1) { |
| 2559 | a.todoMu.Unlock() |
| 2560 | return |
| 2561 | } |
| 2562 | snapshot := append([]evidence.TodoItem(nil), a.todoState...) |
| 2563 | a.todoMu.Unlock() |
| 2564 | a.recordTodoState(snapshot) |
| 2565 | a.emitTodoState(snapshot, m.Index) |
| 2566 | } |
| 2567 | |
| 2568 | // recordTodoState logs the host-advanced list as a synthetic todo_write receipt |
| 2569 | // so the per-turn final gate (which reads the ledger's latest todo_write) sees |
| 2570 | // the advance — the model no longer has to re-send a todo_write to mark the |
| 2571 | // completion. It bypasses the todo_write tool, so the completion-transition |
| 2572 | // guard never runs on it. |
| 2573 | func (a *Agent) recordTodoState(todos []evidence.TodoItem) { |
| 2574 | if a.evidence == nil { |
| 2575 | return |
| 2576 | } |
| 2577 | args, err := json.Marshal(map[string]any{"todos": todos}) |
| 2578 | if err != nil { |
| 2579 | return |
| 2580 | } |
| 2581 | a.evidence.Record(evidence.ReceiptFromToolCall("todo_write", json.RawMessage(args), true, true)) |
| 2582 | } |
| 2583 | |
| 2584 | func canonicalTodoStatus(s string) string { |
| 2585 | s = strings.TrimSpace(s) |
| 2586 | if s == "" { |
| 2587 | return "pending" |
| 2588 | } |
| 2589 | return s |
| 2590 | } |
| 2591 | |
| 2592 | // emitTodoState emits a synthetic todo_write event so the frontend task panel |
| 2593 | // reflects a host-advanced completion without the model re-sending the list. |
| 2594 | // itemIndex is the 1-based position of the completed todo in the panel. |
| 2595 | func (a *Agent) emitTodoState(todos []evidence.TodoItem, itemIndex int) { |
| 2596 | args, err := json.Marshal(map[string]any{"todos": todos}) |
| 2597 | if err != nil { |
| 2598 | return |
| 2599 | } |
| 2600 | id := fmt.Sprintf("host-advance-%d-%d", a.hostAdvanceSeq.Add(1), itemIndex) |
| 2601 | t := event.Tool{ID: id, Name: "todo_write", Args: string(args), ReadOnly: true} |
| 2602 | a.sink.Emit(event.Event{Kind: event.ToolDispatch, Tool: t}) |
| 2603 | t.Output = "task list advanced by complete_step" |
| 2604 | a.sink.Emit(event.Event{Kind: event.ToolResult, Tool: t}) |
| 2605 | } |
| 2606 | |
| 2607 | // RebuildTodoState re-derives canonical task state from the current session |
| 2608 | // transcript. Call after externally truncating the session (e.g. after a |
| 2609 | // user-cancel strip) so Agent.todoState stays consistent with the messages. |
| 2610 | func (a *Agent) RebuildTodoState() { |
| 2611 | a.rebuildTodoState(a.Session().Snapshot()) |
| 2612 | } |
| 2613 | |
| 2614 | // rebuildTodoState reconstructs the canonical task list from a transcript: the |
| 2615 | // latest successful todo_write is the base, then every complete_step after it |
| 2616 | // advances an item. Deterministic from persisted messages, so it survives a |
| 2617 | // fresh load or a rewind (the truncated history yields the historical state). |
| 2618 | // Empty after compaction drops the todo_write — no worse than no canonical list. |
| 2619 | func (a *Agent) rebuildTodoState(msgs []provider.Message) { |
| 2620 | successful := successfulToolCallIDs(msgs) |
| 2621 | var todos []evidence.TodoItem |
| 2622 | baseIdx := -1 |
| 2623 | for i, msg := range msgs { |
| 2624 | for _, tc := range msg.ToolCalls { |
| 2625 | if tc.Name != "todo_write" || !successful[tc.ID] { |
| 2626 | continue |
| 2627 | } |
| 2628 | rec := evidence.ReceiptFromToolCall(tc.Name, json.RawMessage(tc.Arguments), true, true) |
| 2629 | // A successful empty todo_write is an explicit clear. Preserve it as the |
| 2630 | // latest base so history reloads do not resurrect an older non-empty list. |
| 2631 | todos = evidence.NormalizeSerialTodos(rec.Todos) |
| 2632 | baseIdx = i |
| 2633 | } |
| 2634 | } |
| 2635 | if baseIdx < 0 { |
| 2636 | a.setTodoState(nil) |
| 2637 | return |
| 2638 | } |
| 2639 | for i := baseIdx; i < len(msgs); i++ { |
| 2640 | for _, tc := range msgs[i].ToolCalls { |
| 2641 | if tc.Name != "complete_step" || !successful[tc.ID] { |
| 2642 | continue |
| 2643 | } |
| 2644 | rec := evidence.ReceiptFromToolCall(tc.Name, json.RawMessage(tc.Arguments), true, true) |
| 2645 | if m, ok := evidence.MatchStep(rec.Step, todos); ok { |
| 2646 | evidence.AdvanceSerialTodo(todos, m.Index-1) |
| 2647 | } |
| 2648 | } |
| 2649 | } |
| 2650 | a.setTodoState(todos) |
| 2651 | } |
| 2652 | |
| 2653 | func successfulToolCallIDs(msgs []provider.Message) map[string]bool { |
| 2654 | successful := map[string]bool{} |
| 2655 | for _, msg := range msgs { |
| 2656 | if msg.Role != provider.RoleTool || msg.ToolCallID == "" { |
| 2657 | continue |
| 2658 | } |
| 2659 | if !toolResultFailed(msg.Content) { |
| 2660 | successful[msg.ToolCallID] = true |
| 2661 | } |
| 2662 | } |
| 2663 | return successful |
| 2664 | } |
| 2665 | |
| 2666 | func toolResultFailed(content string) bool { |
| 2667 | content = strings.TrimSpace(content) |
| 2668 | return strings.HasPrefix(content, "error:") || |
| 2669 | strings.HasPrefix(content, "blocked:") || |
| 2670 | strings.HasPrefix(content, "Error:") || |
| 2671 | strings.HasPrefix(content, "[error") |
| 2672 | } |
| 2673 | |
| 2674 | func finalReadinessCheckSource(check instruction.VerifyCheck) string { |
| 2675 | source := strings.TrimSpace(check.SourcePath) |
| 2676 | if source == "" { |
| 2677 | source = "project memory" |
| 2678 | } |
| 2679 | if check.Line > 0 { |
| 2680 | return fmt.Sprintf("%s:%d", source, check.Line) |
| 2681 | } |
| 2682 | return source |
| 2683 | } |
| 2684 | |
| 2685 | func shouldNudgeExecutorHandoff(input, answer string) bool { |
| 2686 | return !executorHandoffAllowsTextOnly(input, answer) |
| 2687 | } |
| 2688 | |
| 2689 | func executorHandoffAllowsTextOnly(input, answer string) bool { |
| 2690 | if looksLikeExecutorHandoffDeferral(answer) { |
| 2691 | return false |
| 2692 | } |
| 2693 | task, plan, ok := parseExecutorHandoff(input) |
| 2694 | if !ok { |
| 2695 | return false |
| 2696 | } |
| 2697 | if handoffTaskLooksTextOnly(task) { |
| 2698 | return true |
| 2699 | } |
| 2700 | return handoffPlanLooksTextOnly(plan) |
| 2701 | } |
| 2702 | |
| 2703 | func parseExecutorHandoff(input string) (task, plan string, ok bool) { |
| 2704 | input = StripTransientUserBlocks(input) |
| 2705 | marker := "# " + executorHandoffMarker |
| 2706 | i := strings.Index(input, marker) |
| 2707 | if i < 0 { |
| 2708 | return "", "", false |
| 2709 | } |
| 2710 | input = input[i+len(marker):] |
| 2711 | _, input, ok = strings.Cut(input, "\n\nOriginal task:\n") |
| 2712 | if !ok { |
| 2713 | return "", "", false |
| 2714 | } |
| 2715 | task, input, ok = strings.Cut(input, "\n\nPlanner output:\n") |
| 2716 | if !ok { |
| 2717 | return "", "", false |
| 2718 | } |
| 2719 | plan, _, ok = strings.Cut(input, "\n\nExecutor instructions:") |
| 2720 | if !ok { |
| 2721 | return "", "", false |
| 2722 | } |
| 2723 | if beforeToolContext, _, found := strings.Cut(plan, "\n\nExecutor tool context:"); found { |
| 2724 | plan = beforeToolContext |
| 2725 | } |
| 2726 | return strings.TrimSpace(task), strings.TrimSpace(plan), true |
| 2727 | } |
| 2728 | |
| 2729 | func looksLikeExecutorHandoffDeferral(answer string) bool { |
| 2730 | lower := strings.ToLower(strings.TrimSpace(answer)) |
| 2731 | if lower == "" { |
| 2732 | return true |
| 2733 | } |
| 2734 | if containsAnySubstring(lower, executorHandoffDeferralPhrases) { |
| 2735 | return true |
| 2736 | } |
| 2737 | switch strings.Trim(lower, " \t\r\n.!?。！？") { |
| 2738 | case "ok", "okay", "sounds good", "done", "好的", "可以", "没问题", "收到": |
| 2739 | return true |
| 2740 | default: |
| 2741 | return false |
| 2742 | } |
| 2743 | } |
| 2744 | |
| 2745 | func handoffTaskLooksTextOnly(task string) bool { |
| 2746 | lower := strings.ToLower(strings.TrimSpace(task)) |
| 2747 | if lower == "" { |
| 2748 | return false |
| 2749 | } |
| 2750 | if containsAnySubstring(lower, executorHandoffWorkRequestTerms) { |
| 2751 | return false |
| 2752 | } |
| 2753 | return containsAnySubstring(lower, executorHandoffTextOnlyTaskTerms) |
| 2754 | } |
| 2755 | |
| 2756 | func handoffPlanLooksTextOnly(plan string) bool { |
| 2757 | lower := strings.ToLower(strings.TrimSpace(plan)) |
| 2758 | if lower == "" { |
| 2759 | return false |
| 2760 | } |
| 2761 | if containsAnySubstring(lower, executorHandoffLocalActionTerms) { |
| 2762 | return false |
| 2763 | } |
| 2764 | if containsAnySubstring(lower, executorHandoffTextOnlyPlanTerms) { |
| 2765 | return true |
| 2766 | } |
| 2767 | return strings.Contains(lower, "?") |
| 2768 | } |
| 2769 | |
| 2770 | func containsAnySubstring(s string, terms []string) bool { |
| 2771 | for _, term := range terms { |
| 2772 | if strings.Contains(s, term) { |
| 2773 | return true |
| 2774 | } |
| 2775 | } |
| 2776 | return false |
| 2777 | } |
| 2778 | |
| 2779 | var executorHandoffDeferralPhrases = []string{ |
| 2780 | "plan looks", "looks good", "should be easy", "should be straightforward", |
| 2781 | "i can implement", "i'll implement", "i will implement", "i'll get started", |
| 2782 | "let me ", "i will now", "i'll now", "i can do that", |
| 2783 | "计划看起来", "可以实现", "我会", "我将", "接下来我", "马上开始", |
| 2784 | } |
| 2785 | |
| 2786 | var executorHandoffWorkRequestTerms = []string{ |
| 2787 | "implement", "fix", "refactor", "migrate", "edit", "write", "create", "delete", |
| 2788 | "update", "remove", "add ", "test", "build", "repair", "patch", |
| 2789 | "修改", "修复", "实现", "新增", "重构", "迁移", "补齐", "更新", "删除", "移除", |
| 2790 | } |
| 2791 | |
| 2792 | var executorHandoffTextOnlyTaskTerms = []string{ |
| 2793 | "now what", "what next", "tl;dr", "tldr", "summarize", "summary", "explain", |
| 2794 | "i installed", "i just installed", "i turned on", "i enabled", "it's on", "it is on", |
| 2795 | "怎么办", "下一步", "然后呢", "总结", "解释", "说明", "装了", "装好了", "安装了", "开了", "开启了", "打开了", |
| 2796 | } |
| 2797 | |
| 2798 | var executorHandoffLocalActionTerms = []string{ |
| 2799 | "write_file", "read_file", "apply_patch", "bash", |
| 2800 | "workspace", "repo", "repository", "codebase", "file", "path", |
| 2801 | "write ", "edit ", "modify ", "create ", "delete ", "remove ", "update ", "add ", "patch ", "refactor ", "implement ", |
| 2802 | "run ", "command", "test", "build", |
| 2803 | "文件", "路径", "仓库", "代码", "写入", "编辑", "修改", "创建", "删除", "移除", "更新", "新增", "运行", "命令", "测试", "构建", |
| 2804 | } |
| 2805 | |
| 2806 | var executorHandoffTextOnlyPlanTerms = []string{ |
| 2807 | "tell the user", "ask the user", "guide the user", "explain to the user", |
| 2808 | "summarize", "summary", "tl;dr", "tldr", "answer the user", "respond to the user", |
| 2809 | "provide guidance", "walk the user", "instruct the user", "have the user", |
| 2810 | "user should", "the user should", "user can", "the user can", "manual", "manually", |
| 2811 | "no tools needed", "no tool calls needed", "does not need tools", "needs no tools", |
| 2812 | "listen", "play a song", "compare the difference", "checkbox", |
| 2813 | "告诉用户", "询问用户", "问用户", "让用户", "请用户", "指导用户", "解释", "总结", "回答", |
| 2814 | "手动", "无需工具", "不需要工具", "试听", "听歌", "对比", "勾选", |
| 2815 | } |
| 2816 | |
| 2817 | func executorHandoffRetryMessage() string { |
| 2818 | return `You are already in the executor phase. The planner's read-only limitations do not apply to you. |
| 2819 | |
| 2820 | The tool schema is still attached to this executor request. Do not invent that MCP servers or tools are unavailable; only report an unavailable tool after a real tool call or host error proves it. |
| 2821 | |
| 2822 | Do not answer as the planner and do not ask how to trigger the executor. |
| 2823 | Use your available tools now to carry out the task. If carrying out the planner's instructions requires a user-owned choice or review, call the ask tool with concrete options and wait for its tool result; do not ask in prose, and do not claim the user answered unless an actual ask tool result or a new user message says so. If a write or command is blocked by permissions or workspace boundaries, state that specific blocker and ask for the needed approval/path.` |
| 2824 | } |
| 2825 | |
| 2826 | func hasVisibleFinalAnswer(text string) bool { |
| 2827 | return strings.TrimSpace(text) != "" |
| 2828 | } |
| 2829 | |
| 2830 | // reasoningOnlyFinishHonoured reports whether the model finished with a stop |
| 2831 | // signal but placed its answer in the reasoning stream rather than the content |
| 2832 | // block. DeepSeek thinking mode does this occasionally: it streams a long |
| 2833 | // reasoning_content, then returns finish_reason="stop" with an empty content. |
| 2834 | // The model has signalled completion, so the host accepts the turn instead of |
| 2835 | // retrying and forcing another expensive thinking round. |
| 2836 | // |
| 2837 | // The accept is scoped to DeepSeek thinking mode (ToolCallReasoningPolicy): |
| 2838 | // for other providers a reasoning-only turn keeps the empty-final retry |
| 2839 | // safety net — local <think>-tag models often recover a visible answer on |
| 2840 | // the second attempt, and a gateway that mislabels truncation as "stop" |
| 2841 | // must not have a degenerate turn committed as the final answer. |
| 2842 | func reasoningOnlyFinishHonoured(p provider.Provider, u *provider.Usage, reasoning string) bool { |
| 2843 | if !provider.RequiresToolCallReasoning(p) { |
| 2844 | return false |
| 2845 | } |
| 2846 | if u == nil || u.FinishReason != "stop" { |
| 2847 | return false |
| 2848 | } |
| 2849 | return strings.TrimSpace(reasoning) != "" |
| 2850 | } |
| 2851 | |
| 2852 | func emptyFinalRetryMessage() string { |
| 2853 | return "The previous assistant response finished without any visible answer text. Continue the same task now and provide a concise visible answer to the user. Do not send reasoning only." |
| 2854 | } |
| 2855 | |
| 2856 | func emptyFinalNotice() string { |
| 2857 | return "No visible answer was produced; asking the assistant to respond again." |
| 2858 | } |
| 2859 | |
| 2860 | func emptyFinalNoticeDetail(prov string, u *provider.Usage, reasoningLen int) string { |
| 2861 | finish := "unknown" |
| 2862 | if u != nil && u.FinishReason != "" { |
| 2863 | finish = u.FinishReason |
| 2864 | } |
| 2865 | return fmt.Sprintf("empty final answer blocked: %s returned no visible answer text (finish=%s, reasoning=%d chars); retrying", prov, finish, reasoningLen) |
| 2866 | } |
| 2867 | |
| 2868 | func executorHandoffNoticeText() string { |
| 2869 | return "The assistant answered before taking action; asking it to use the required tools." |
| 2870 | } |
| 2871 | |
| 2872 | func toolBudgetNoticeText() string { |
| 2873 | return "Tool round limit reached; asking the assistant to summarize progress." |
| 2874 | } |
| 2875 | |
| 2876 | // samplingRequest is a once-prepared, frozen provider request for one model |
| 2877 | // round. All stream retries replay this exact payload — no synthetic recovery |
| 2878 | // messages, no schema reorder, no previous_response_id drift from failed attempts. |
| 2879 | type samplingRequest struct { |
| 2880 | req provider.Request |
| 2881 | } |
| 2882 | |
| 2883 | // prepareSamplingRequest runs interceptors and schema fetch once per model |
| 2884 | // round. Callers deep-copy via freezeProviderRequest before each Stream so |
| 2885 | // providers cannot mutate the shared freeze across retries. |
| 2886 | func (a *Agent) prepareSamplingRequest(ctx context.Context) (samplingRequest, error) { |
| 2887 | // CreatedAt is durable UI metadata, not model input. Strip it from the |
| 2888 | // transport copy so wall-clock differences never invalidate the provider's |
| 2889 | // prompt-cache prefix (and custom providers cannot accidentally send it). |
| 2890 | requestMessages := append([]provider.Message(nil), provider.ModelMessages(a.session.Messages)...) |
| 2891 | for i := range requestMessages { |
| 2892 | requestMessages[i].CreatedAt = 0 |
| 2893 | } |
| 2894 | // context.prepare: extensions may rewrite the message copy feeding THIS |
| 2895 | // request. The session log is never touched — the replacement is |
| 2896 | // ephemeral, so the next request starts from the unmodified history and |
| 2897 | // the prompt-cache prefix stays intact across turns. |
| 2898 | requestMessages, err := a.interceptContextPrepare(ctx, requestMessages) |
| 2899 | if err != nil { |
| 2900 | return samplingRequest{}, err |
| 2901 | } |
| 2902 | req := provider.Request{ |
| 2903 | Messages: requestMessages, |
| 2904 | Tools: a.tools.Schemas(), |
| 2905 | MaxTokens: a.maxOutputTokens, |
| 2906 | Temperature: provider.OptionalTemperature(a.temperature), |
| 2907 | ResponseFormat: responseFormatFromRequest(ctx), |
| 2908 | } |
| 2909 | // provider.request: the fully assembled request gets one last ruling |
| 2910 | // (revalidated by the payload registry) before it goes on the wire. |
| 2911 | req, err = a.interceptProviderRequest(ctx, req) |
| 2912 | if err != nil { |
| 2913 | return samplingRequest{}, err |
| 2914 | } |
| 2915 | return samplingRequest{req: freezeProviderRequest(req)}, nil |
| 2916 | } |
| 2917 | |
| 2918 | // freezeProviderRequest deep-copies the provider-visible request surface so |
| 2919 | // retries share identical messages, tools order, temperature, and format. |
| 2920 | func freezeProviderRequest(req provider.Request) provider.Request { |
| 2921 | out := req |
| 2922 | if len(req.Messages) > 0 { |
| 2923 | out.Messages = append([]provider.Message(nil), req.Messages...) |
| 2924 | for i := range out.Messages { |
| 2925 | if len(out.Messages[i].ToolCalls) > 0 { |
| 2926 | out.Messages[i].ToolCalls = append([]provider.ToolCall(nil), out.Messages[i].ToolCalls...) |
| 2927 | } |
| 2928 | if len(out.Messages[i].Images) > 0 { |
| 2929 | out.Messages[i].Images = append([]string(nil), out.Messages[i].Images...) |
| 2930 | } |
| 2931 | if len(out.Messages[i].ResponsesItems) > 0 { |
| 2932 | items := make([]json.RawMessage, len(out.Messages[i].ResponsesItems)) |
| 2933 | for j, item := range out.Messages[i].ResponsesItems { |
| 2934 | items[j] = append(json.RawMessage(nil), item...) |
| 2935 | } |
| 2936 | out.Messages[i].ResponsesItems = items |
| 2937 | } |
| 2938 | } |
| 2939 | } |
| 2940 | if len(req.Tools) > 0 { |
| 2941 | out.Tools = make([]provider.ToolSchema, len(req.Tools)) |
| 2942 | for i, schema := range req.Tools { |
| 2943 | out.Tools[i] = schema |
| 2944 | if len(schema.Parameters) > 0 { |
| 2945 | out.Tools[i].Parameters = append(json.RawMessage(nil), schema.Parameters...) |
| 2946 | } |
| 2947 | } |
| 2948 | } |
| 2949 | if req.Temperature != nil { |
| 2950 | t := *req.Temperature |
| 2951 | out.Temperature = &t |
| 2952 | } |
| 2953 | if req.ResponseFormat != nil { |
| 2954 | rf := *req.ResponseFormat |
| 2955 | out.ResponseFormat = &rf |
| 2956 | } |
| 2957 | return out |
| 2958 | } |
| 2959 | |
| 2960 | // stream runs one completion, emitting reasoning and text deltas as typed |
| 2961 | // events and collecting complete tool calls. A Message event closes the text |
| 2962 | // stream so a sink can re-render the streamed raw text as styled markdown. The |
| 2963 | // accumulated text and reasoning are also returned so the caller can round-trip |
| 2964 | // reasoning on the next turn. |
| 2965 | // |
| 2966 | // When frozen is non-nil, the request is not rebuilt from session — retries |
| 2967 | // must replay the same provider-visible body. |
| 2968 | func (a *Agent) stream(ctx context.Context, turn int, sink event.Sink) (string, string, string, string, string, []provider.ToolCall, []json.RawMessage, *provider.Usage, bool, bool, []provider.ToolCall, int, error) { |
| 2969 | return a.streamWithFrozen(ctx, turn, sink, nil, "") |
| 2970 | } |
| 2971 | |
| 2972 | func (a *Agent) streamWithFrozen(ctx context.Context, turn int, sink event.Sink, frozen *samplingRequest, attemptID string) (string, string, string, string, string, []provider.ToolCall, []json.RawMessage, *provider.Usage, bool, bool, []provider.ToolCall, int, error) { |
| 2973 | ctx = provider.WithRetryNotify(ctx, func(info provider.RetryInfo) { |
| 2974 | sink.Emit(event.Event{Kind: event.Retrying, RetryAttempt: info.Attempt, RetryMax: info.Max, RetryScope: event.RetryScopeHeaders}) |
| 2975 | }) |
| 2976 | // Reuse a parent attempt counter when present so stream retries accumulate |
| 2977 | // into one RequestCount; otherwise install a fresh counter for this call. |
| 2978 | ctx = provider.WithRequestAttemptCounter(ctx) |
| 2979 | // A stream can terminate locally before the provider channel closes (for |
| 2980 | // example when the client-side reasoning guard fires). Own a child context |
| 2981 | // here so every return path aborts the HTTP request and releases the provider |
| 2982 | // reader instead of leaving generation and billing running in the background. |
| 2983 | ctx, cancel := context.WithCancel(ctx) |
| 2984 | defer cancel() |
| 2985 | |
| 2986 | var req provider.Request |
| 2987 | var err error |
| 2988 | if frozen != nil { |
| 2989 | req = freezeProviderRequest(frozen.req) |
| 2990 | } else { |
| 2991 | prepared, perr := a.prepareSamplingRequest(ctx) |
| 2992 | if perr != nil { |
| 2993 | return "", "", "", "", "", nil, nil, nil, false, false, nil, 0, perr |
| 2994 | } |
| 2995 | req = prepared.req |
| 2996 | } |
| 2997 | // After #7725 Goal token request admission was removed, stream goes |
| 2998 | // directly to the provider. Provider-visible cache controls stay stable |
| 2999 | // across retries and request timing because they are derived from req alone. |
| 3000 | ch, err := a.prov.Stream(ctx, req) |
| 3001 | if err != nil { |
| 3002 | return "", "", "", "", "", nil, nil, provider.UsageWithRequestAttemptCount(ctx, nil), false, false, nil, 0, err |
| 3003 | } |
| 3004 | |
| 3005 | // A PostLLMCall hook rewrites the whole reasoning block, so when one is wired |
| 3006 | // up we buffer reasoning silently and emit the transformed text once after the |
| 3007 | // stream. With no such hook the reasoning streams live, chunk by chunk, as |
| 3008 | // before — the common case must not lose its live "thinking…" display. |
| 3009 | transformReasoning := a.hooks != nil && a.hooks.HasPostLLMCall() |
| 3010 | |
| 3011 | var text, reasoning strings.Builder |
| 3012 | var signature string // provider-issued proof for the reasoning (Anthropic thinking) |
| 3013 | var reasoningID, reasoningStatus string // Responses reasoning item id/status (meta chunk) |
| 3014 | var calls []provider.ToolCall |
| 3015 | var responsesItems []json.RawMessage |
| 3016 | var partialCalls []provider.ToolCall |
| 3017 | var usage *provider.Usage |
| 3018 | var partialToolStarted bool |
| 3019 | var maxArgChars int |
| 3020 | var lastArgProgress time.Time |
| 3021 | finishReasoning := func() (stored, display string) { |
| 3022 | original := reasoning.String() |
| 3023 | display = original |
| 3024 | if transformReasoning && original != "" { |
| 3025 | display = a.hooks.PostLLMCall(ctx, original, turn) |
| 3026 | if display != "" { |
| 3027 | sink.Emit(event.Event{Kind: event.Reasoning, Text: display}) |
| 3028 | } |
| 3029 | } |
| 3030 | stored = display |
| 3031 | providerBound := signature != "" || reasoningID != "" || reasoningStatus != "" |
| 3032 | if providerBound || provider.RequiresReasoningRoundTrip(a.prov) || (len(calls) > 0 && provider.RequiresToolCallReasoning(a.prov)) { |
| 3033 | stored = original |
| 3034 | } |
| 3035 | return stored, display |
| 3036 | } |
| 3037 | for { |
| 3038 | var chunk provider.Chunk |
| 3039 | select { |
| 3040 | case <-ctx.Done(): |
| 3041 | stored, _ := finishReasoning() |
| 3042 | usage = bestEffortStreamUsage(usage, text.Len(), reasoning.Len(), "interrupted") |
| 3043 | usage = provider.UsageWithRequestAttemptCount(ctx, usage) |
| 3044 | return text.String(), stored, signature, reasoningID, reasoningStatus, calls, responsesItems, usage, false, partialToolStarted, partialCalls, maxArgChars, ctx.Err() |
| 3045 | case c, ok := <-ch: |
| 3046 | if !ok { |
| 3047 | if err := ctx.Err(); err != nil { |
| 3048 | stored, _ := finishReasoning() |
| 3049 | usage = bestEffortStreamUsage(usage, text.Len(), reasoning.Len(), "interrupted") |
| 3050 | usage = provider.UsageWithRequestAttemptCount(ctx, usage) |
| 3051 | return text.String(), stored, signature, reasoningID, reasoningStatus, calls, responsesItems, usage, false, partialToolStarted, partialCalls, maxArgChars, err |
| 3052 | } |
| 3053 | stored, display := finishReasoning() |
| 3054 | // provider.response: extensions rule on the assembled terminal |
| 3055 | // response before it is persisted. A replacement becomes the |
| 3056 | // visible assistant turn (the user's transcript); a block fails |
| 3057 | // the turn. |
| 3058 | providerSignature := signature |
| 3059 | finalText, finalReasoning, signature, calls, usage, err := a.interceptProviderResponse( |
| 3060 | ctx, text.String(), stored, signature, calls, usage) |
| 3061 | if err != nil { |
| 3062 | return "", "", "", "", "", nil, nil, nil, false, partialToolStarted, partialCalls, maxArgChars, err |
| 3063 | } |
| 3064 | // Responses reasoning IDs/status and Anthropic signatures are |
| 3065 | // provider-bound metadata. Never attach the provider's metadata |
| 3066 | // to reasoning that an extension replaced. |
| 3067 | if finalReasoning != stored || signature != providerSignature { |
| 3068 | reasoningID, reasoningStatus = "", "" |
| 3069 | } |
| 3070 | if finalReasoning != stored { |
| 3071 | // The extension replaced the reasoning: what is persisted |
| 3072 | // and what the closing Message event re-renders must agree. |
| 3073 | display = finalReasoning |
| 3074 | } |
| 3075 | if finalText != "" || display != "" { |
| 3076 | sink.Emit(event.Event{ |
| 3077 | Kind: event.Message, |
| 3078 | Text: DisplayAssistantText(finalText), |
| 3079 | Reasoning: display, |
| 3080 | }) |
| 3081 | } |
| 3082 | usage = provider.UsageWithRequestAttemptCount(ctx, usage) |
| 3083 | return finalText, finalReasoning, signature, reasoningID, reasoningStatus, calls, responsesItems, usage, false, false, partialCalls, maxArgChars, nil |
| 3084 | } |
| 3085 | chunk = c |
| 3086 | } |
| 3087 | switch chunk.Type { |
| 3088 | case provider.ChunkReasoning: |
| 3089 | reasoning.WriteString(chunk.Text) |
| 3090 | if chunk.Signature != "" { |
| 3091 | signature = chunk.Signature |
| 3092 | } |
| 3093 | // 元数据 chunk（空 Text）：reasoning item id/status 贯通 |
| 3094 | // SSE → session → 下一轮回传（评审 #7234 第 1 点）。 |
| 3095 | if chunk.ReasoningID != "" { |
| 3096 | reasoningID = chunk.ReasoningID |
| 3097 | } |
| 3098 | if chunk.ReasoningStatus != "" { |
| 3099 | reasoningStatus = chunk.ReasoningStatus |
| 3100 | } |
| 3101 | if chunk.Text != "" && !transformReasoning { |
| 3102 | sink.Emit(event.Event{Kind: event.Reasoning, Text: chunk.Text}) |
| 3103 | } |
| 3104 | if a.reasoningByteLimit > 0 && reasoning.Len() > a.reasoningByteLimit { |
| 3105 | stored, _ := finishReasoning() |
| 3106 | usage = bestEffortStreamUsage(usage, text.Len(), reasoning.Len(), finishReasonClientReasoningLimit) |
| 3107 | usage = provider.UsageWithRequestAttemptCount(ctx, usage) |
| 3108 | a.lastUsage.Store(usage) |
| 3109 | return text.String(), stored, signature, reasoningID, reasoningStatus, calls, responsesItems, usage, false, partialToolStarted, partialCalls, maxArgChars, errReasoningByteLimitExceeded |
| 3110 | } |
| 3111 | case provider.ChunkText: |
| 3112 | text.WriteString(chunk.Text) |
| 3113 | sink.Emit(event.Event{Kind: event.Text, Text: chunk.Text}) |
| 3114 | case provider.ChunkToolCallStart: |
| 3115 | partialToolStarted = true |
| 3116 | // Surface the tool card as soon as the call begins — before its |
| 3117 | // (possibly large) arguments finish streaming — so the user sees it |
| 3118 | // working instead of a stall. executeBatch emits the full dispatch |
| 3119 | // (with args) once the call completes; the frontend merges by ID. |
| 3120 | if tc := chunk.ToolCall; tc != nil { |
| 3121 | partialCalls = upsertPartialToolCall(partialCalls, *tc) |
| 3122 | sink.Emit(event.Event{Kind: event.ToolDispatch, Tool: event.Tool{ |
| 3123 | ID: tc.ID, Name: tc.Name, ReadOnly: a.toolReadOnly(tc.Name), Partial: true, AttemptID: attemptID, |
| 3124 | }}) |
| 3125 | } |
| 3126 | case provider.ChunkToolCallArgsDelta: |
| 3127 | partialToolStarted = true |
| 3128 | // Liveness ticks while a large argument payload streams: re-emit the |
| 3129 | // partial dispatch with the cumulative size (time-throttled) so the |
| 3130 | // UI can show progress instead of a dead counter for the duration of |
| 3131 | // a 30KB write_file body. |
| 3132 | if chunk.ArgChars > maxArgChars { |
| 3133 | maxArgChars = chunk.ArgChars |
| 3134 | } |
| 3135 | if tc := chunk.ToolCall; tc != nil && time.Since(lastArgProgress) >= 250*time.Millisecond { |
| 3136 | partialCalls = upsertPartialToolCall(partialCalls, *tc) |
| 3137 | lastArgProgress = time.Now() |
| 3138 | sink.Emit(event.Event{Kind: event.ToolDispatch, Tool: event.Tool{ |
| 3139 | ID: tc.ID, Name: tc.Name, ReadOnly: a.toolReadOnly(tc.Name), Partial: true, ArgChars: chunk.ArgChars, AttemptID: attemptID, |
| 3140 | }}) |
| 3141 | } |
| 3142 | case provider.ChunkToolCall: |
| 3143 | partialToolStarted = true |
| 3144 | if chunk.ToolCall != nil { |
| 3145 | calls = append(calls, *chunk.ToolCall) |
| 3146 | partialCalls = upsertPartialToolCall(partialCalls, *chunk.ToolCall) |
| 3147 | if n := len(chunk.ToolCall.Arguments); n > maxArgChars { |
| 3148 | maxArgChars = n |
| 3149 | } |
| 3150 | } |
| 3151 | case provider.ChunkResponsesItem: |
| 3152 | if len(chunk.ResponsesItem) > 0 { |
| 3153 | responsesItems = append(responsesItems, append(json.RawMessage(nil), chunk.ResponsesItem...)) |
| 3154 | } |
| 3155 | case provider.ChunkUsage: |
| 3156 | usage = chunk.Usage |
| 3157 | a.lastUsage.Store(chunk.Usage) |
| 3158 | a.sessCacheHit.Add(int64(chunk.Usage.CacheHitTokens)) |
| 3159 | a.sessCacheMiss.Add(int64(chunk.Usage.CacheMissTokens)) |
| 3160 | case provider.ChunkError: |
| 3161 | if provider.IsStreamInterrupted(chunk.Err) { |
| 3162 | stored, _ := finishReasoning() |
| 3163 | usage = bestEffortStreamUsage(usage, text.Len(), reasoning.Len(), "interrupted") |
| 3164 | usage = provider.UsageWithRequestAttemptCount(ctx, usage) |
| 3165 | return text.String(), stored, signature, reasoningID, reasoningStatus, calls, responsesItems, usage, true, partialToolStarted, partialCalls, maxArgChars, chunk.Err |
| 3166 | } |
| 3167 | stored, _ := finishReasoning() |
| 3168 | if errors.Is(chunk.Err, context.Canceled) || errors.Is(chunk.Err, context.DeadlineExceeded) { |
| 3169 | usage = bestEffortStreamUsage(usage, text.Len(), reasoning.Len(), "interrupted") |
| 3170 | } |
| 3171 | usage = provider.UsageWithRequestAttemptCount(ctx, usage) |
| 3172 | return text.String(), stored, signature, reasoningID, reasoningStatus, calls, responsesItems, usage, false, partialToolStarted, partialCalls, maxArgChars, chunk.Err |
| 3173 | } |
| 3174 | } |
| 3175 | } |
| 3176 | |
| 3177 | func bestEffortStreamUsage(current *provider.Usage, textBytes, reasoningBytes int, finishReason string) *provider.Usage { |
| 3178 | if current == nil && textBytes == 0 && reasoningBytes == 0 { |
| 3179 | return nil |
| 3180 | } |
| 3181 | var usage provider.Usage |
| 3182 | if current != nil { |
| 3183 | usage = *current |
| 3184 | } |
| 3185 | if finishReason != "" { |
| 3186 | usage.FinishReason = finishReason |
| 3187 | } |
| 3188 | reasoningTokens := estimateTokensFromBytes(reasoningBytes) |
| 3189 | textTokens := estimateTokensFromBytes(textBytes) |
| 3190 | completionTokens := reasoningTokens + textTokens |
| 3191 | if usage.ReasoningTokens < reasoningTokens { |
| 3192 | usage.ReasoningTokens = reasoningTokens |
| 3193 | usage.Estimated = true |
| 3194 | } |
| 3195 | if usage.CompletionTokens < completionTokens { |
| 3196 | usage.CompletionTokens = completionTokens |
| 3197 | usage.Estimated = true |
| 3198 | } |
| 3199 | if minTotal := usage.PromptTokens + usage.CompletionTokens; usage.TotalTokens < minTotal { |
| 3200 | usage.TotalTokens = minTotal |
| 3201 | usage.Estimated = true |
| 3202 | } |
| 3203 | return &usage |
| 3204 | } |
| 3205 | |
| 3206 | func estimateTokensFromBytes(n int) int { |
| 3207 | if n <= 0 { |
| 3208 | return 0 |
| 3209 | } |
| 3210 | tokens := n / 4 |
| 3211 | if n%4 != 0 { |
| 3212 | tokens++ |
| 3213 | } |
| 3214 | if tokens <= 0 { |
| 3215 | return 1 |
| 3216 | } |
| 3217 | return tokens |
| 3218 | } |
| 3219 | |
| 3220 | func upsertPartialToolCall(calls []provider.ToolCall, call provider.ToolCall) []provider.ToolCall { |
| 3221 | for i := range calls { |
| 3222 | if call.ID != "" && calls[i].ID == call.ID { |
| 3223 | calls[i] = call |
| 3224 | return calls |
| 3225 | } |
| 3226 | } |
| 3227 | return append(calls, call) |
| 3228 | } |
| 3229 | |
| 3230 | func (a *Agent) recordInterruptedDisplay(text, reasoning string, calls []provider.ToolCall, pending bool, workDurationMs int64) { |
| 3231 | displayCalls := make([]provider.ToolCall, 0, len(calls)) |
| 3232 | interrupted := make([]string, 0, len(calls)) |
| 3233 | seen := make(map[string]struct{}, len(calls)) |
| 3234 | for _, call := range calls { |
| 3235 | name := strings.TrimSpace(call.Name) |
| 3236 | key := call.ID + "\x00" + name |
| 3237 | if _, ok := seen[key]; ok { |
| 3238 | continue |
| 3239 | } |
| 3240 | seen[key] = struct{}{} |
| 3241 | displayCalls = append(displayCalls, provider.ToolCall{ID: call.ID, Name: name}) |
| 3242 | if name != "" { |
| 3243 | interrupted = append(interrupted, name) |
| 3244 | } |
| 3245 | } |
| 3246 | a.session.Add(provider.Message{ |
| 3247 | Role: provider.RoleTool, |
| 3248 | Content: text, |
| 3249 | ReasoningContent: reasoning, |
| 3250 | ToolCalls: displayCalls, |
| 3251 | ToolCallID: provider.LocalOnlyToolID, |
| 3252 | Name: provider.LocalOnlyToolName, |
| 3253 | WorkDurationMs: workDurationMs, |
| 3254 | LocalOnly: true, |
| 3255 | InterruptedTurn: &provider.InterruptedTurnRecovery{ |
| 3256 | Pending: pending, |
| 3257 | InterruptedTools: interrupted, |
| 3258 | DroppedPartialText: strings.TrimSpace(text) != "", |
| 3259 | DroppedPartialReasoning: strings.TrimSpace(reasoning) != "", |
| 3260 | }, |
| 3261 | }) |
| 3262 | } |
| 3263 | |
| 3264 | func (a *Agent) capturePrefixShape(schemas []provider.ToolSchema) PrefixShape { |
| 3265 | return CaptureShape(a.systemPrompt(), schemas, a.session.RewriteVersion()) |
| 3266 | } |
| 3267 | |
| 3268 | func (a *Agent) systemPrompt() string { |
| 3269 | var b strings.Builder |
| 3270 | for _, m := range a.session.Messages { |
| 3271 | if m.Role != provider.RoleSystem { |
| 3272 | continue |
| 3273 | } |
| 3274 | if b.Len() > 0 { |
| 3275 | b.WriteByte('\n') |
| 3276 | } |
| 3277 | b.WriteString(m.Content) |
| 3278 | } |
| 3279 | return b.String() |
| 3280 | } |
| 3281 | |
| 3282 | // batchExecution is the result of one provider tool-call batch. |
| 3283 | type batchExecution struct { |
| 3284 | results []string |
| 3285 | images [][]string |
| 3286 | executions []*tool.ShellExecution |
| 3287 | recoveryStopTurn bool |
| 3288 | recoveryStopReason string |
| 3289 | } |
| 3290 | |
| 3291 | // executeBatch dispatches one model turn's tool calls. A ToolDispatch event is |
| 3292 | // emitted for every call up front, in call order, so a frontend can show the |
| 3293 | // timeline chronologically. Contiguous known ReadOnly calls fan out across |
| 3294 | // goroutines; unknown and writer calls run as single-call serial segments so |
| 3295 | // write/read ordering stays provider-ordered. ToolResult events are emitted |
| 3296 | // after the batch in call order, so emission stays serial even when execution |
| 3297 | // parallelised. Images are aligned by index with results. |
| 3298 | func (a *Agent) executeBatch(ctx context.Context, calls []provider.ToolCall) batchExecution { |
| 3299 | // The assistant message already stored this slice in Session. Keep execution |
| 3300 | // state separate so refreshing a dependent preview never mutates shared |
| 3301 | // session memory outside Session's lock. |
| 3302 | calls = append([]provider.ToolCall(nil), calls...) |
| 3303 | for _, c := range calls { |
| 3304 | a.emitFullToolDispatch(c, false) |
| 3305 | } |
| 3306 | |
| 3307 | results := make([]string, len(calls)) |
| 3308 | outcomes := make([]toolOutcome, len(calls)) |
| 3309 | durations := make([]int64, len(calls)) |
| 3310 | completedStepInBatch := false |
| 3311 | // Snapshot the receipt count before the batch runs: if a loop guard fires |
| 3312 | // for this batch, successes recorded during it (a mixed batch where only one |
| 3313 | // call was guard-blocked) must already count as progress against the pass. |
| 3314 | receiptMark := 0 |
| 3315 | if a.evidence != nil { |
| 3316 | receiptMark = a.evidence.Len() |
| 3317 | } |
| 3318 | // Full dispatches are prepared against the batch's initial file state. After |
| 3319 | // one writer runs, a dependent later writer may only become previewable (or |
| 3320 | // its original preview may become stale). Refresh even after a failed writer: |
| 3321 | // commands and filesystem calls can mutate disk before reporting an error. |
| 3322 | // The first writer stays on the single-preview fast path. |
| 3323 | earlierWriterRan := false |
| 3324 | surfaceWriters := make([]bool, len(calls)) |
| 3325 | run := func(i int) { |
| 3326 | t, _, ambiguous := a.tools.ResolveCall(calls[i].Name) |
| 3327 | known := t != nil && len(ambiguous) == 0 |
| 3328 | writer := known && !t.ReadOnly() |
| 3329 | surfaceWriters[i] = writer |
| 3330 | if earlierWriterRan && writer { |
| 3331 | if refreshed, changed := refreshCurrentFileDiff(t, calls[i]); changed { |
| 3332 | calls[i] = refreshed |
| 3333 | a.session.UpdateToolCallPreview(refreshed) |
| 3334 | a.emitFullToolDispatch(refreshed, true) |
| 3335 | } |
| 3336 | } |
| 3337 | start := time.Now() |
| 3338 | if calls[i].Name == "complete_step" && completedStepInBatch { |
| 3339 | output := "blocked: only one successful complete_step is allowed per tool-call round. Continue from the newly promoted in_progress todo in the next round instead of batching sign-offs." |
| 3340 | outcomes[i] = toolOutcome{output: output, blocked: true, errMsg: "blocked: complete_step sign-offs must be serial"} |
| 3341 | if a.evidence != nil { |
| 3342 | a.evidence.Record(evidence.ReceiptFromToolCall(calls[i].Name, json.RawMessage(calls[i].Arguments), false, true)) |
| 3343 | } |
| 3344 | durations[i] = time.Since(start).Milliseconds() |
| 3345 | results[i] = output |
| 3346 | return |
| 3347 | } |
| 3348 | outcomes[i] = a.executeOne(ctx, calls[i]) |
| 3349 | if outcomes[i].resolved { |
| 3350 | readOnly := outcomes[i].resolvedReadOnly |
| 3351 | calls[i].ResolvedName = outcomes[i].resolvedName |
| 3352 | calls[i].CapabilityID = outcomes[i].capabilityID |
| 3353 | calls[i].ResolvedReadOnly = &readOnly |
| 3354 | surfaceWriters[i] = !readOnly |
| 3355 | } |
| 3356 | if calls[i].Name == "complete_step" && outcomes[i].errMsg == "" { |
| 3357 | completedStepInBatch = true |
| 3358 | } |
| 3359 | durations[i] = time.Since(start).Milliseconds() |
| 3360 | results[i] = outcomes[i].output |
| 3361 | } |
| 3362 | finalize := func(i int) { |
| 3363 | if calls[i].ResolvedReadOnly != nil { |
| 3364 | a.session.UpdateToolCallResolution(calls[i]) |
| 3365 | a.emitResolvedToolDispatch(calls[i]) |
| 3366 | } |
| 3367 | if surfaceWriters[i] || (outcomes[i].resolved && !outcomes[i].resolvedReadOnly) { |
| 3368 | earlierWriterRan = true |
| 3369 | } |
| 3370 | } |
| 3371 | cancelled := false |
| 3372 | markCancelled := func(start int) { |
| 3373 | errMsg := context.Canceled.Error() |
| 3374 | if err := ctx.Err(); err != nil { |
| 3375 | errMsg = err.Error() |
| 3376 | } |
| 3377 | output := "cancelled: context cancelled before execution" |
| 3378 | for j := start; j < len(calls); j++ { |
| 3379 | results[j] = output |
| 3380 | outcomes[j] = toolOutcome{output: output, errMsg: errMsg} |
| 3381 | } |
| 3382 | cancelled = true |
| 3383 | } |
| 3384 | |
| 3385 | // recoveryBatchStop blocks remaining tools after Episode budgets are |
| 3386 | // exhausted so tool-call / result pairs stay complete for the provider. |
| 3387 | recoveryBatchStop := false |
| 3388 | recoveryStopReason := "" |
| 3389 | markRecoveryStopped := func(start int, reason string) { |
| 3390 | msg := "blocked: Auto recovery paused this turn; do not call more tools. Summarize completed work for the user." |
| 3391 | for j := start; j < len(calls); j++ { |
| 3392 | if results[j] != "" { |
| 3393 | continue |
| 3394 | } |
| 3395 | results[j] = msg |
| 3396 | outcomes[j] = toolOutcome{ |
| 3397 | output: msg, |
| 3398 | blocked: true, |
| 3399 | errMsg: firstLine(msg), |
| 3400 | recoveryStopTurn: true, |
| 3401 | recoveryStopReason: reason, |
| 3402 | } |
| 3403 | } |
| 3404 | recoveryBatchStop = true |
| 3405 | if reason != "" { |
| 3406 | recoveryStopReason = reason |
| 3407 | } |
| 3408 | } |
| 3409 | |
| 3410 | // mutationBatchStop is the deterministic dependency barrier: after any |
| 3411 | // mutating call fails or is blocked, later mutating and verification calls |
| 3412 | // in the same provider batch are skipped (not_run/dependency). Host-proven |
| 3413 | // read-only diagnosis may still run. executeOne also re-checks after proxy |
| 3414 | // resolution so use_capability cannot bypass this pass. |
| 3415 | mutationBatchStop := false |
| 3416 | a.mutationDependencyBarrier.Store(false) |
| 3417 | markDependencySkipped := func(start int) { |
| 3418 | a.mutationDependencyBarrier.Store(true) |
| 3419 | for j := start; j < len(calls); j++ { |
| 3420 | if results[j] != "" { |
| 3421 | continue |
| 3422 | } |
| 3423 | // Pre-classify when statically certain. Proxies and ambiguous |
| 3424 | // targets fall through to run() so executeOne can resolve the real |
| 3425 | // target and re-apply the barrier before Commit/Execute. |
| 3426 | if !batchCallStaticallySkippable(a, calls[j]) { |
| 3427 | continue |
| 3428 | } |
| 3429 | isVerification := calls[j].Name == "bash" && evidence.IsDeliveryVerificationCommand(bashCommandFromArgs(json.RawMessage(calls[j].Arguments))) |
| 3430 | msg := "blocked: skipped because an earlier modification in this tool batch failed or was blocked. " + |
| 3431 | "Fix or re-run the failed change first; verification was not executed." |
| 3432 | var ex *tool.ShellExecution |
| 3433 | if calls[j].Name == "bash" { |
| 3434 | ex = &tool.ShellExecution{ |
| 3435 | Kind: "shell", |
| 3436 | State: tool.ShellStateNotRun, |
| 3437 | FailurePhase: tool.ShellPhaseDependency, |
| 3438 | MutationRisk: tool.ShellMutationNotStarted, |
| 3439 | Verification: tool.ShellVerificationNotVerification, |
| 3440 | } |
| 3441 | if isVerification { |
| 3442 | ex.Verification = tool.ShellVerificationNotRun |
| 3443 | } |
| 3444 | if t, _, amb := a.tools.ResolveCall(calls[j].Name); t != nil && len(amb) == 0 { |
| 3445 | if bt, ok := t.(tool.DetailedExecutor); ok { |
| 3446 | if desc := bt.ExecutionDescriptor(json.RawMessage(calls[j].Arguments)); desc != nil { |
| 3447 | ex.Shell = desc.Shell |
| 3448 | ex.ShellVersion = desc.ShellVersion |
| 3449 | ex.Platform = desc.Platform |
| 3450 | ex.SupportsAndAnd = desc.SupportsAndAnd |
| 3451 | } |
| 3452 | } |
| 3453 | } |
| 3454 | } |
| 3455 | results[j] = msg |
| 3456 | outcomes[j] = toolOutcome{ |
| 3457 | output: msg, |
| 3458 | blocked: true, |
| 3459 | errMsg: firstLine(msg), |
| 3460 | execution: ex, |
| 3461 | } |
| 3462 | durations[j] = 0 |
| 3463 | } |
| 3464 | mutationBatchStop = true |
| 3465 | } |
| 3466 | |
| 3467 | for _, batch := range partitionToolCalls(a.tools, calls) { |
| 3468 | if ctx.Err() != nil { |
| 3469 | markCancelled(batch.start) |
| 3470 | break |
| 3471 | } |
| 3472 | if recoveryBatchStop { |
| 3473 | markRecoveryStopped(batch.start, recoveryStopReason) |
| 3474 | break |
| 3475 | } |
| 3476 | if batch.parallel && batch.end-batch.start > 1 { |
| 3477 | // Parallel segments are read-only by construction; no mutation barrier. |
| 3478 | ranUntil := runParallel(ctx, batch.start, batch.end, run) |
| 3479 | for i := batch.start; i < ranUntil; i++ { |
| 3480 | finalize(i) |
| 3481 | } |
| 3482 | // After parallel execution completes, check if context was cancelled. |
| 3483 | // The individual tool executions should have detected ctx.Done(), but |
| 3484 | // we verify here to ensure we don't continue to subsequent batches. |
| 3485 | if ctx.Err() != nil { |
| 3486 | markCancelled(ranUntil) |
| 3487 | break |
| 3488 | } |
| 3489 | for i := batch.start; i < batch.end; i++ { |
| 3490 | if outcomes[i].recoveryStopTurn { |
| 3491 | recoveryBatchStop = true |
| 3492 | recoveryStopReason = outcomes[i].recoveryStopReason |
| 3493 | markRecoveryStopped(batch.end, recoveryStopReason) |
| 3494 | break |
| 3495 | } |
| 3496 | } |
| 3497 | if recoveryBatchStop { |
| 3498 | break |
| 3499 | } |
| 3500 | continue |
| 3501 | } |
| 3502 | for i := batch.start; i < batch.end; i++ { |
| 3503 | // Before executing the next tool, check if context was cancelled. |
| 3504 | // This prevents starting new tools when a previous tool's execution |
| 3505 | // triggered cancellation. |
| 3506 | if ctx.Err() != nil { |
| 3507 | markCancelled(i) |
| 3508 | break |
| 3509 | } |
| 3510 | if recoveryBatchStop { |
| 3511 | markRecoveryStopped(i, recoveryStopReason) |
| 3512 | break |
| 3513 | } |
| 3514 | if mutationBatchStop { |
| 3515 | // Fill dependency skips for remaining mutating/verify calls, then |
| 3516 | // allow any residual read-only diagnosis to run individually. |
| 3517 | if results[i] != "" { |
| 3518 | continue |
| 3519 | } |
| 3520 | t, _, ambiguous := a.tools.ResolveCall(calls[i].Name) |
| 3521 | known := t != nil && len(ambiguous) == 0 |
| 3522 | readOnly := known && t.ReadOnly() |
| 3523 | if calls[i].Name == "bash" && permission.BashCommandIsReadOnly(json.RawMessage(calls[i].Arguments)) { |
| 3524 | readOnly = true |
| 3525 | } |
| 3526 | isVerification := calls[i].Name == "bash" && evidence.IsDeliveryVerificationCommand(bashCommandFromArgs(json.RawMessage(calls[i].Arguments))) |
| 3527 | mutates := evidence.ToolCallMutates(calls[i].Name, json.RawMessage(calls[i].Arguments), readOnly) |
| 3528 | if mutates || isVerification { |
| 3529 | markDependencySkipped(i) |
| 3530 | // markDependencySkipped fills this index; move on. |
| 3531 | if results[i] != "" { |
| 3532 | continue |
| 3533 | } |
| 3534 | } |
| 3535 | } |
| 3536 | if results[i] != "" { |
| 3537 | // Pre-filled dependency skip. |
| 3538 | finalize(i) |
| 3539 | continue |
| 3540 | } |
| 3541 | run(i) |
| 3542 | finalize(i) |
| 3543 | if outcomes[i].recoveryStopTurn { |
| 3544 | recoveryBatchStop = true |
| 3545 | recoveryStopReason = outcomes[i].recoveryStopReason |
| 3546 | markRecoveryStopped(i+1, recoveryStopReason) |
| 3547 | break |
| 3548 | } |
| 3549 | // Mutation/verification failure barrier for the rest of this batch. |
| 3550 | if batchCallIsMutatingFailure(a, calls[i], outcomes[i]) { |
| 3551 | mutationBatchStop = true |
| 3552 | markDependencySkipped(i + 1) |
| 3553 | } |
| 3554 | // After each tool execution, also check if the context was cancelled. |
| 3555 | // If so, stop executing remaining tools and return immediately so |
| 3556 | // the agent loop can detect the cancellation and exit. |
| 3557 | if ctx.Err() != nil { |
| 3558 | markCancelled(i + 1) |
| 3559 | break |
| 3560 | } |
| 3561 | } |
| 3562 | if cancelled || recoveryBatchStop { |
| 3563 | break |
| 3564 | } |
| 3565 | } |
| 3566 | |
| 3567 | for i, c := range calls { |
| 3568 | o := outcomes[i] |
| 3569 | t, _, ambiguous := a.tools.ResolveCall(c.Name) |
| 3570 | ok := t != nil && len(ambiguous) == 0 |
| 3571 | readOnly := ok && t.ReadOnly() |
| 3572 | if c.ResolvedReadOnly != nil { |
| 3573 | readOnly = *c.ResolvedReadOnly |
| 3574 | } |
| 3575 | tr := event.Tool{ |
| 3576 | ID: c.ID, |
| 3577 | Name: c.Name, |
| 3578 | Args: c.Arguments, |
| 3579 | ResolvedName: c.ResolvedName, |
| 3580 | CapabilityID: c.CapabilityID, |
| 3581 | Output: o.output, |
| 3582 | Err: o.errMsg, |
| 3583 | ReadOnly: readOnly, |
| 3584 | Truncated: o.truncated, |
| 3585 | DurationMs: durations[i], |
| 3586 | Execution: toEventShellExecution(o.execution, durations[i]), |
| 3587 | } |
| 3588 | a.sink.Emit(event.Event{Kind: event.ToolResult, Tool: tr}) |
| 3589 | if o.truncated && o.truncMsg != "" { |
| 3590 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, Text: o.truncMsg}) |
| 3591 | } |
| 3592 | } |
| 3593 | if !cancelled { |
| 3594 | a.applyStormBreaker(calls, outcomes, results, receiptMark) |
| 3595 | } |
| 3596 | images := make([][]string, len(calls)) |
| 3597 | executions := make([]*tool.ShellExecution, len(calls)) |
| 3598 | for i := range outcomes { |
| 3599 | images[i] = outcomes[i].images |
| 3600 | executions[i] = outcomes[i].execution |
| 3601 | if outcomes[i].recoveryStopTurn { |
| 3602 | recoveryBatchStop = true |
| 3603 | if outcomes[i].recoveryStopReason != "" { |
| 3604 | recoveryStopReason = outcomes[i].recoveryStopReason |
| 3605 | } |
| 3606 | } |
| 3607 | } |
| 3608 | return batchExecution{ |
| 3609 | results: results, |
| 3610 | images: images, |
| 3611 | executions: executions, |
| 3612 | recoveryStopTurn: recoveryBatchStop, |
| 3613 | recoveryStopReason: recoveryStopReason, |
| 3614 | } |
| 3615 | } |
| 3616 | |
| 3617 | func toEventShellExecution(in *tool.ShellExecution, durationMs int64) *event.ShellExecution { |
| 3618 | if in == nil { |
| 3619 | return nil |
| 3620 | } |
| 3621 | out := &event.ShellExecution{ |
| 3622 | Kind: in.Kind, |
| 3623 | Shell: in.Shell, |
| 3624 | ShellVersion: in.ShellVersion, |
| 3625 | Platform: in.Platform, |
| 3626 | SupportsAndAnd: in.SupportsAndAnd, |
| 3627 | State: in.State, |
| 3628 | FailurePhase: in.FailurePhase, |
| 3629 | OutputTail: in.OutputTail, |
| 3630 | MutationRisk: in.MutationRisk, |
| 3631 | Verification: in.Verification, |
| 3632 | DurationMs: in.DurationMs, |
| 3633 | } |
| 3634 | if out.DurationMs == 0 && durationMs > 0 { |
| 3635 | out.DurationMs = durationMs |
| 3636 | } |
| 3637 | if in.ExitCode != nil { |
| 3638 | code := *in.ExitCode |
| 3639 | out.ExitCode = &code |
| 3640 | } |
| 3641 | return out |
| 3642 | } |
| 3643 | |
| 3644 | func toProviderToolExecution(in *tool.ShellExecution) *provider.ToolExecution { |
| 3645 | if in == nil { |
| 3646 | return nil |
| 3647 | } |
| 3648 | out := &provider.ToolExecution{ |
| 3649 | Kind: in.Kind, |
| 3650 | Shell: in.Shell, |
| 3651 | ShellVersion: in.ShellVersion, |
| 3652 | Platform: in.Platform, |
| 3653 | SupportsAndAnd: in.SupportsAndAnd, |
| 3654 | State: in.State, |
| 3655 | FailurePhase: in.FailurePhase, |
| 3656 | OutputTail: in.OutputTail, |
| 3657 | MutationRisk: in.MutationRisk, |
| 3658 | Verification: in.Verification, |
| 3659 | DurationMs: in.DurationMs, |
| 3660 | } |
| 3661 | if in.ExitCode != nil { |
| 3662 | code := *in.ExitCode |
| 3663 | out.ExitCode = &code |
| 3664 | } |
| 3665 | return out |
| 3666 | } |
| 3667 | |
| 3668 | // batchCallIsMutatingFailure reports whether a finished call was a mutation |
| 3669 | // (file write / non-readonly bash mutation) that failed or was blocked, so later |
| 3670 | // mutations and verifications in the same batch must not run. |
| 3671 | func batchCallIsMutatingFailure(a *Agent, call provider.ToolCall, o toolOutcome) bool { |
| 3672 | if o.errMsg == "" && !o.blocked { |
| 3673 | return false |
| 3674 | } |
| 3675 | readOnly := false |
| 3676 | t, _, ambiguous := a.tools.ResolveCall(call.Name) |
| 3677 | known := t != nil && len(ambiguous) == 0 |
| 3678 | if known { |
| 3679 | readOnly = t.ReadOnly() |
| 3680 | } |
| 3681 | if call.ResolvedReadOnly != nil { |
| 3682 | readOnly = *call.ResolvedReadOnly |
| 3683 | } |
| 3684 | if o.resolved { |
| 3685 | readOnly = o.resolvedReadOnly |
| 3686 | } |
| 3687 | if call.Name == "bash" && permission.BashCommandIsReadOnly(json.RawMessage(call.Arguments)) { |
| 3688 | readOnly = true |
| 3689 | } |
| 3690 | // Verification failures do not open the dependency barrier by themselves — |
| 3691 | // only a failed modification does. |
| 3692 | if call.Name == "bash" && evidence.IsDeliveryVerificationCommand(bashCommandFromArgs(json.RawMessage(call.Arguments))) { |
| 3693 | return false |
| 3694 | } |
| 3695 | // Resolved writers (including MCP targets behind use_capability) count even |
| 3696 | // when the provider-visible proxy advertised ReadOnly. |
| 3697 | if o.resolved && !o.resolvedReadOnly { |
| 3698 | return true |
| 3699 | } |
| 3700 | if evidence.ToolCallMutates(call.Name, json.RawMessage(call.Arguments), readOnly) { |
| 3701 | return true |
| 3702 | } |
| 3703 | // Fail closed only for a target the host could not classify at all. A blanket |
| 3704 | // !readOnly fallback here would re-admit exactly the writers ToolCallMutates |
| 3705 | // deliberately exempts (todo_write, complete_step, ask, bash_output, wait and |
| 3706 | // the other non-mutation meta tools): a failed todo update would then block |
| 3707 | // every real edit left in the batch. Resolved writer proxies already returned |
| 3708 | // true above, so narrowing this does not reopen the use_capability path. |
| 3709 | return !known |
| 3710 | } |
| 3711 | |
| 3712 | // batchCallStaticallySkippable reports whether a remaining call can be marked |
| 3713 | // not_run/dependency without resolving a proxy. Proxies and unknown tools |
| 3714 | // return false so executeOne can resolve the real target first. |
| 3715 | func batchCallStaticallySkippable(a *Agent, call provider.ToolCall) bool { |
| 3716 | t, _, ambiguous := a.tools.ResolveCall(call.Name) |
| 3717 | if t == nil || len(ambiguous) > 0 { |
| 3718 | // Unknown / ambiguous: fail closed via executeOne path. |
| 3719 | return false |
| 3720 | } |
| 3721 | // Proxy resolution may consult a live connected capability and its result |
| 3722 | // can change between calls. Do not resolve here merely to pre-fill a skip: |
| 3723 | // executeOne resolves exactly once, then applyMutationDependencyBarrier |
| 3724 | // classifies the real target before Commit or Execute. |
| 3725 | if _, ok := t.(tool.CallResolver); ok { |
| 3726 | return false |
| 3727 | } |
| 3728 | readOnly := t.ReadOnly() |
| 3729 | if call.Name == "bash" && permission.BashCommandIsReadOnly(json.RawMessage(call.Arguments)) { |
| 3730 | readOnly = true |
| 3731 | } |
| 3732 | isVerification := call.Name == "bash" && evidence.IsDeliveryVerificationCommand(bashCommandFromArgs(json.RawMessage(call.Arguments))) |
| 3733 | if isVerification { |
| 3734 | return true |
| 3735 | } |
| 3736 | return !readOnly || evidence.ToolCallMutates(call.Name, json.RawMessage(call.Arguments), readOnly) |
| 3737 | } |
| 3738 | |
| 3739 | func (a *Agent) emitFullToolDispatch(c provider.ToolCall, refreshed bool) { |
| 3740 | t, _, ambiguous := a.tools.ResolveCall(c.Name) |
| 3741 | ok := t != nil && len(ambiguous) == 0 |
| 3742 | ev := event.Tool{ID: c.ID, Name: c.Name, Args: c.Arguments, ReadOnly: ok && t.ReadOnly(), Refreshed: refreshed} |
| 3743 | ev.FileDiff = event.FileDiff{Diff: c.Diff, Added: c.Added, Removed: c.Removed} |
| 3744 | if ok && ev.Diff == "" && ev.Added == 0 && ev.Removed == 0 { |
| 3745 | if ch, ok := tool.PreviewChange(t, json.RawMessage(c.Arguments)); ok { |
| 3746 | ev.FileDiff = event.FileDiff{Diff: ch.Diff, Added: ch.Added, Removed: ch.Removed} |
| 3747 | } |
| 3748 | } |
| 3749 | if ok { |
| 3750 | if pr, ok := t.(interface { |
| 3751 | ResolveProfile(json.RawMessage) *event.Profile |
| 3752 | }); ok { |
| 3753 | ev.Profile = pr.ResolveProfile(json.RawMessage(c.Arguments)) |
| 3754 | } |
| 3755 | } |
| 3756 | a.sink.Emit(event.Event{Kind: event.ToolDispatch, Tool: ev}) |
| 3757 | } |
| 3758 | |
| 3759 | // emitResolvedToolDispatch upserts the real target classification of a stable |
| 3760 | // proxy call without changing the provider-visible Name/Args. Append-only sinks |
| 3761 | // ignore Refreshed events; stateful frontends replace the existing card by ID. |
| 3762 | func (a *Agent) emitResolvedToolDispatch(c provider.ToolCall) { |
| 3763 | if c.ResolvedReadOnly == nil { |
| 3764 | return |
| 3765 | } |
| 3766 | if c.ResolvedName != "" && c.ResolvedName != c.Name { |
| 3767 | EmitProxyAudit(a.sink, tool.ResolvedCall{ |
| 3768 | DisplayName: c.Name, |
| 3769 | TargetName: c.ResolvedName, |
| 3770 | CapabilityID: c.CapabilityID, |
| 3771 | }) |
| 3772 | } |
| 3773 | a.sink.Emit(event.Event{Kind: event.ToolDispatch, Tool: event.Tool{ |
| 3774 | ID: c.ID, |
| 3775 | Name: c.Name, |
| 3776 | Args: c.Arguments, |
| 3777 | ResolvedName: c.ResolvedName, |
| 3778 | CapabilityID: c.CapabilityID, |
| 3779 | ReadOnly: *c.ResolvedReadOnly, |
| 3780 | Refreshed: true, |
| 3781 | FileDiff: event.FileDiff{ |
| 3782 | Diff: c.Diff, Added: c.Added, Removed: c.Removed, |
| 3783 | }, |
| 3784 | }}) |
| 3785 | } |
| 3786 | |
| 3787 | // refreshCurrentFileDiff recomputes a writer preview against the state left by |
| 3788 | // earlier successful writers in the same provider batch. Preview failures clear |
| 3789 | // any stale initial diff; a later Execute will then fail or ask for recovery |
| 3790 | // without presenting the user with a preview that no longer describes disk. |
| 3791 | func refreshCurrentFileDiff(t tool.Tool, call provider.ToolCall) (provider.ToolCall, bool) { |
| 3792 | pv, ok := t.(tool.Previewer) |
| 3793 | if !ok { |
| 3794 | return call, false |
| 3795 | } |
| 3796 | refreshed := call |
| 3797 | refreshed.Diff = "" |
| 3798 | refreshed.Added = 0 |
| 3799 | refreshed.Removed = 0 |
| 3800 | if change, err := pv.Preview(json.RawMessage(call.Arguments)); err == nil { |
| 3801 | refreshed.Diff = change.Diff |
| 3802 | refreshed.Added = change.Added |
| 3803 | refreshed.Removed = change.Removed |
| 3804 | } |
| 3805 | changed := refreshed.Diff != call.Diff || refreshed.Added != call.Added || refreshed.Removed != call.Removed |
| 3806 | return refreshed, changed |
| 3807 | } |
| 3808 | |
| 3809 | func (a *Agent) withPreviewFileDiffs(calls []provider.ToolCall) []provider.ToolCall { |
| 3810 | if len(calls) == 0 { |
| 3811 | return calls |
| 3812 | } |
| 3813 | out := make([]provider.ToolCall, len(calls)) |
| 3814 | copy(out, calls) |
| 3815 | for i := range out { |
| 3816 | if out[i].Diff != "" || out[i].Added != 0 || out[i].Removed != 0 { |
| 3817 | continue |
| 3818 | } |
| 3819 | t, _, ambiguous := a.tools.ResolveCall(out[i].Name) |
| 3820 | ok := t != nil && len(ambiguous) == 0 |
| 3821 | if !ok { |
| 3822 | continue |
| 3823 | } |
| 3824 | if ch, ok := tool.PreviewChange(t, json.RawMessage(out[i].Arguments)); ok { |
| 3825 | out[i].Diff = ch.Diff |
| 3826 | out[i].Added = ch.Added |
| 3827 | out[i].Removed = ch.Removed |
| 3828 | } |
| 3829 | } |
| 3830 | return out |
| 3831 | } |
| 3832 | |
| 3833 | type toolCallBatch struct { |
| 3834 | start int |
| 3835 | end int |
| 3836 | parallel bool |
| 3837 | } |
| 3838 | |
| 3839 | // partitionToolCalls keeps provider order while letting contiguous known |
| 3840 | // read-only tools run together. Unknown and writer tools are single-call serial |
| 3841 | // batches so they cannot reorder around reads or produce surprising errors. |
| 3842 | // complete_step and todo_write read the turn's evidence ledger. wait and |
| 3843 | // bash_output can merge a background task's receipts into that ledger. These |
| 3844 | // evidence-sensitive tools never join a parallel run, so provider order stays |
| 3845 | // receipt order. use_capability is always serial because its provider-visible |
| 3846 | // read-only surface can resolve to a real MCP writer only inside executeOne; |
| 3847 | // batching it as a reader would let multiple database/API mutations race. |
| 3848 | func partitionToolCalls(r *tool.Registry, calls []provider.ToolCall) []toolCallBatch { |
| 3849 | var batches []toolCallBatch |
| 3850 | for i := 0; i < len(calls); { |
| 3851 | if parallelisable(r, calls[i].Name) { |
| 3852 | start := i |
| 3853 | i++ |
| 3854 | for i < len(calls) && parallelisable(r, calls[i].Name) { |
| 3855 | i++ |
| 3856 | } |
| 3857 | batches = append(batches, toolCallBatch{start: start, end: i, parallel: true}) |
| 3858 | continue |
| 3859 | } |
| 3860 | batches = append(batches, toolCallBatch{start: i, end: i + 1}) |
| 3861 | i++ |
| 3862 | } |
| 3863 | return batches |
| 3864 | } |
| 3865 | |
| 3866 | func parallelisable(r *tool.Registry, name string) bool { |
| 3867 | switch name { |
| 3868 | case "complete_step", "todo_write", "wait", "bash_output", "use_capability": |
| 3869 | return false |
| 3870 | } |
| 3871 | t, _, ambiguous := r.ResolveCall(name) |
| 3872 | return t != nil && len(ambiguous) == 0 && t.ReadOnly() |
| 3873 | } |
| 3874 | |
| 3875 | func runParallel(ctx context.Context, start, end int, run func(int)) int { |
| 3876 | const maxParallel = 8 |
| 3877 | sem := make(chan struct{}, maxParallel) |
| 3878 | var wg sync.WaitGroup |
| 3879 | ranUntil := start |
| 3880 | launch: |
| 3881 | for i := start; i < end; i++ { |
| 3882 | if ctx.Err() != nil { |
| 3883 | break |
| 3884 | } |
| 3885 | select { |
| 3886 | case sem <- struct{}{}: |
| 3887 | case <-ctx.Done(): |
| 3888 | break launch |
| 3889 | } |
| 3890 | if ctx.Err() != nil { |
| 3891 | <-sem |
| 3892 | break |
| 3893 | } |
| 3894 | i := i |
| 3895 | wg.Add(1) |
| 3896 | ranUntil = i + 1 |
| 3897 | go func() { |
| 3898 | defer wg.Done() |
| 3899 | defer func() { <-sem }() |
| 3900 | run(i) |
| 3901 | }() |
| 3902 | } |
| 3903 | wg.Wait() |
| 3904 | return ranUntil |
| 3905 | } |
| 3906 | |
| 3907 | // stormBreakThreshold is how many times in a row the same tool may fail the same |
| 3908 | // way before the loop stops echoing the raw error back and instead returns a |
| 3909 | // directive to change approach. Two natural self-corrections are healthy; the |
| 3910 | // third identical failure is a death-spiral — the dominant case being a tool call |
| 3911 | // whose arguments are truncated at the output-token ceiling, which the model then |
| 3912 | // re-emits (re-worded but still over-long), truncating the same way again. |
| 3913 | const stormBreakThreshold = 3 |
| 3914 | |
| 3915 | // repeatSuccessBreakThreshold is how many identical write-like successes the |
| 3916 | // agent allows before refusing another copy in the same user turn. Two gives the |
| 3917 | // model room for a natural self-correction; the third repeat is usually a |
| 3918 | // no-op/write loop and should be redirected to a different tool or final answer. |
| 3919 | const repeatSuccessBreakThreshold = 2 |
| 3920 | |
| 3921 | const ( |
| 3922 | // todoProgressNudgeRounds is the first adaptive checkpoint. The host asks |
| 3923 | // the model to reassess, but keeps the turn alive so it can recover. |
| 3924 | todoProgressNudgeRounds = 8 |
| 3925 | // maxTodoStallRounds pauses only after the reassessment also failed to |
| 3926 | // produce a new completion or unique host-observed work receipt. |
| 3927 | maxTodoStallRounds = 16 |
| 3928 | ) |
| 3929 | |
| 3930 | func todoProgressNudgeMessage(rounds int) string { |
| 3931 | return fmt.Sprintf("Host progress check: the current todo has produced no new completion, unique read, command, or mutation for %d tool-call rounds. Reassess before using more tools: sign off the current item if it is done, narrow the remaining work without replacing the active item, or explain/ask about a real blocker. Do not repeat reads, commands, or writes just to reset this guard.", rounds) |
| 3932 | } |
| 3933 | |
| 3934 | // loopGuardBlockErrMsg is the errMsg carried by a repeat-success loop-guard |
| 3935 | // block. applyStormBreaker matches it to arm the final-readiness loop-guard |
| 3936 | // pass, since that guard also invites the model to report the blocker. |
| 3937 | const loopGuardBlockErrMsg = "blocked by loop guard" |
| 3938 | |
| 3939 | // applyStormBreaker detects a run of zero-progress turns and, past the |
| 3940 | // threshold, rewrites the model-facing result (results[0]) into a directive to |
| 3941 | // change approach. Two detectors, because a stuck model varies its retries two |
| 3942 | // ways. The signature detector keys on each call's (tool, error/blocker) — not |
| 3943 | // its args — since a stuck model reworks the arguments cosmetically while |
| 3944 | // hitting the same host refusal or failure (see the stormSig field doc). The |
| 3945 | // streak detector counts consecutive turns in which every call was blocked, |
| 3946 | // regardless of shape: rotating tools, reordering a batch, or a blocker whose |
| 3947 | // text varies per attempt escapes the signature but is still zero progress — |
| 3948 | // only a host refusal (not a plain error) proves that, so the streak requires |
| 3949 | // blocked outcomes. Any success resets both. When a guard fires — or when a |
| 3950 | // call in the batch was already blocked by the per-call repeat-success guard — |
| 3951 | // the final-readiness loop-guard pass is armed so the model may report the |
| 3952 | // blocker (see loopGuardAllowsFinal). The hard maxSteps guard remains the |
| 3953 | // ultimate backstop; this just keeps the loop from burning that whole budget |
| 3954 | // bouncing off the same host refusals. |
| 3955 | func (a *Agent) applyStormBreaker(calls []provider.ToolCall, outcomes []toolOutcome, results []string, receiptMark int) { |
| 3956 | allBlocked := len(outcomes) > 0 |
| 3957 | for _, outcome := range outcomes { |
| 3958 | if !outcome.blocked { |
| 3959 | allBlocked = false |
| 3960 | break |
| 3961 | } |
| 3962 | } |
| 3963 | if allBlocked { |
| 3964 | a.blockedTurnStreak++ |
| 3965 | } else { |
| 3966 | a.blockedTurnStreak = 0 |
| 3967 | } |
| 3968 | for _, outcome := range outcomes { |
| 3969 | if outcome.blocked && outcome.errMsg == loopGuardBlockErrMsg { |
| 3970 | a.armLoopGuardPass(receiptMark) |
| 3971 | break |
| 3972 | } |
| 3973 | } |
| 3974 | |
| 3975 | sig, ok := batchStormSignature(calls, outcomes) |
| 3976 | switch { |
| 3977 | case !ok: |
| 3978 | a.stormSig, a.stormCount = "", 0 |
| 3979 | case sig != a.stormSig: |
| 3980 | a.stormSig, a.stormCount = sig, 1 |
| 3981 | default: |
| 3982 | a.stormCount++ |
| 3983 | } |
| 3984 | stormHit := ok && a.stormCount >= stormBreakThreshold |
| 3985 | streakHit := allBlocked && a.blockedTurnStreak >= stormBreakThreshold |
| 3986 | if !stormHit && !streakHit { |
| 3987 | return |
| 3988 | } |
| 3989 | |
| 3990 | const blockedAdvice = "Change approach: do not keep retrying a blocked tool by changing the tool, command, or arguments. Respect the permission, plan-mode, hook, or loop-guard blocker; use an already-allowed tool, ask the user for the specific approval or choice if appropriate, or explain the blocker in your final answer." |
| 3991 | var guard, detail string |
| 3992 | if stormHit { |
| 3993 | subject := fmt.Sprintf("%q", calls[0].Name) |
| 3994 | short := calls[0].Name |
| 3995 | if len(calls) > 1 { |
| 3996 | subject = fmt.Sprintf("this batch of %d tool calls", len(calls)) |
| 3997 | short = fmt.Sprintf("a batch of %d calls", len(calls)) |
| 3998 | } |
| 3999 | anyBlocked := false |
| 4000 | for _, outcome := range outcomes { |
| 4001 | if outcome.blocked { |
| 4002 | anyBlocked = true |
| 4003 | break |
| 4004 | } |
| 4005 | } |
| 4006 | action := "failed" |
| 4007 | advice := "Change approach: if an argument is being truncated, write less in one call and split the work into several smaller calls; otherwise fix the arguments, use a different tool, or explain the blocker in your final answer." |
| 4008 | if anyBlocked { |
| 4009 | action = "been blocked or failed" |
| 4010 | advice = blockedAdvice |
| 4011 | } |
| 4012 | guard = fmt.Sprintf( |
| 4013 | "[loop guard] %s has now %s %d times in a row with the same host response. Re-sending it — even with the wording changed — will not help: the calls keep hitting the same outcome. %s", |
| 4014 | subject, action, a.stormCount, advice) |
| 4015 | detail = fmt.Sprintf( |
| 4016 | "loop guard: %s hit the same host response %d× — nudging the model to change approach", |
| 4017 | short, a.stormCount) |
| 4018 | } else { |
| 4019 | guard = fmt.Sprintf( |
| 4020 | "[loop guard] every tool call in the last %d turns has been blocked by the host (permission, plan mode, hook, or loop guard). Switching tools, reordering calls, or rewording arguments will not help while the blockers stand. %s", |
| 4021 | a.blockedTurnStreak, blockedAdvice) |
| 4022 | detail = fmt.Sprintf( |
| 4023 | "loop guard: every tool call blocked %d turns in a row — nudging the model to change approach", |
| 4024 | a.blockedTurnStreak) |
| 4025 | } |
| 4026 | results[0] = outcomes[0].output + "\n\n" + guard |
| 4027 | a.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, Code: event.NoticeCodeLoopGuard, Text: loopGuardNoticeText(), Detail: detail}) |
| 4028 | a.armLoopGuardPass(receiptMark) |
| 4029 | } |
| 4030 | |
| 4031 | func loopGuardNoticeText() string { |
| 4032 | return "The assistant is not making progress; asking it to change approach." |
| 4033 | } |
| 4034 | |
| 4035 | // batchStormSignature returns a per-turn fixation signature — each call's |
| 4036 | // (name, error/blocker) in order — and ok=true only when every call errored or |
| 4037 | // was blocked. ok=false (any success) means the turn made progress, so the |
| 4038 | // caller resets the counter. Keying on the host response rather than the args is |
| 4039 | // deliberate: a stuck model reworks the arguments while hitting the same |
| 4040 | // response, so identical-args matching would miss the loop. |
| 4041 | func batchStormSignature(calls []provider.ToolCall, outcomes []toolOutcome) (string, bool) { |
| 4042 | if len(calls) == 0 { |
| 4043 | return "", false |
| 4044 | } |
| 4045 | var sb strings.Builder |
| 4046 | for i := range calls { |
| 4047 | if outcomes[i].errMsg == "" { |
| 4048 | return "", false |
| 4049 | } |
| 4050 | sb.WriteString(calls[i].Name) |
| 4051 | sb.WriteByte(0) |
| 4052 | sb.WriteString(outcomes[i].errMsg) |
| 4053 | sb.WriteByte(0) |
| 4054 | } |
| 4055 | return sb.String(), true |
| 4056 | } |
| 4057 | |
| 4058 | // toolOutcome is one tool call's result, split into the model-facing output and |
| 4059 | // the display-facing notice bits. errMsg is the short failure reason (empty on |
| 4060 | // success) — a refused call, an unknown tool, or an execution error — so a sink |
| 4061 | // renders the result as failed ("⊘ name <errMsg>" / a red card) instead of OK; |
| 4062 | // blocked narrows that to a refusal (plan mode / permission). truncMsg is set |
| 4063 | // (without the "· " prefix) when the output was head+tailed. images carries |
| 4064 | // data URLs from a tool.ImageTool result; they ride outside output so text |
| 4065 | // truncation can never corrupt an image payload. |
| 4066 | type toolOutcome struct { |
| 4067 | output string |
| 4068 | images []string |
| 4069 | blocked bool |
| 4070 | errMsg string |
| 4071 | truncated bool |
| 4072 | truncMsg string |
| 4073 | resolved bool |
| 4074 | resolvedName string |
| 4075 | capabilityID string |
| 4076 | resolvedReadOnly bool |
| 4077 | // execution is local shell metadata (optional). Provider messages strip it |
| 4078 | // via ModelMessages; UI/event sinks surface it on ToolResult cards. |
| 4079 | execution *tool.ShellExecution |
| 4080 | // recoveryGeneration is the gate generation captured before execution so |
| 4081 | // ObserveResult can ignore stale results after a mode switch. |
| 4082 | recoveryGeneration uint64 |
| 4083 | // recoveryStopTurn is set when Auto Episode budgets are exhausted. |
| 4084 | recoveryStopTurn bool |
| 4085 | recoveryStopReason string |
| 4086 | } |
| 4087 | |
| 4088 | // completedMCPConnect recognizes a synthetic cache-miss connect call whose |
| 4089 | // background discovery finished after the provider request was serialized. The |
| 4090 | // connect placeholder is intentionally absent once real tools replace it, but |
| 4091 | // the already-advertised call still completed its only job and must not surface |
| 4092 | // as an unknown tool. |
| 4093 | func completedMCPConnect(reg *tool.Registry, name string) (string, bool) { |
| 4094 | server, rawName, ok := tool.SplitMCPName(name) |
| 4095 | if !ok || rawName != "connect" { |
| 4096 | return "", false |
| 4097 | } |
| 4098 | prefix := tool.MCPNamePrefix + server + "__" |
| 4099 | for _, current := range reg.Names() { |
| 4100 | if current != name && strings.HasPrefix(current, prefix) { |
| 4101 | return server, true |
| 4102 | } |
| 4103 | } |
| 4104 | return "", false |
| 4105 | } |
| 4106 | |
| 4107 | // recoveryPlanTransition detects structural rewrites of an active canonical |
| 4108 | // task list. Initial plans and progress-only status updates stay on the fast |
| 4109 | // path; changing step identity, order, or hierarchy while work remains is a |
| 4110 | // semantic transition for the independent Auto reviewer. |
| 4111 | func (a *Agent) recoveryPlanTransition(toolName string, args json.RawMessage) (bool, string, string) { |
| 4112 | if a == nil || toolName != "todo_write" || a.planMode.Load() { |
| 4113 | return false, "", "" |
| 4114 | } |
| 4115 | before := a.CanonicalTodoState() |
| 4116 | if len(before) == 0 || len(evidence.IncompleteTodos(before)) == 0 { |
| 4117 | return false, "", "" |
| 4118 | } |
| 4119 | after := evidence.ReceiptFromToolCall("todo_write", args, true, true).Todos |
| 4120 | if len(after) == 0 || evidence.ValidateSerialTodos(after) != nil || !evidence.PreservesCompletedTodoPositions(before, after) { |
| 4121 | // Let todo_write report malformed or invalid state directly; an invalid |
| 4122 | // task list is not a meaningful plan proposal for the reviewer. |
| 4123 | return false, "", "" |
| 4124 | } |
| 4125 | if samePlanStructure(before, after) { |
| 4126 | return false, "", "" |
| 4127 | } |
| 4128 | return true, planReviewText(before), planReviewText(after) |
| 4129 | } |
| 4130 | |
| 4131 | func samePlanStructure(a, b []evidence.TodoItem) bool { |
| 4132 | if len(a) != len(b) { |
| 4133 | return false |
| 4134 | } |
| 4135 | for i := range a { |
| 4136 | if a[i].Level != b[i].Level || normalizePlanStep(a[i].Content) != normalizePlanStep(b[i].Content) { |
| 4137 | return false |
| 4138 | } |
| 4139 | } |
| 4140 | return true |
| 4141 | } |
| 4142 | |
| 4143 | func normalizePlanStep(s string) string { |
| 4144 | return strings.Join(strings.Fields(strings.TrimSpace(s)), " ") |
| 4145 | } |
| 4146 | |
| 4147 | func planReviewText(todos []evidence.TodoItem) string { |
| 4148 | var b strings.Builder |
| 4149 | for i, todo := range todos { |
| 4150 | indent := "" |
| 4151 | if todo.Level == 1 { |
| 4152 | indent = " " |
| 4153 | } |
| 4154 | fmt.Fprintf(&b, "%s%d. %s [%s]", indent, i+1, normalizePlanStep(todo.Content), canonicalTodoStatus(todo.Status)) |
| 4155 | if i+1 < len(todos) { |
| 4156 | b.WriteByte('\n') |
| 4157 | } |
| 4158 | } |
| 4159 | return b.String() |
| 4160 | } |
| 4161 | |
| 4162 | func recoveryTaskScopeID(deliveryScopeID string, runSeq uint64) string { |
| 4163 | if scope := strings.TrimSpace(deliveryScopeID); scope != "" { |
| 4164 | return "goal:" + scope |
| 4165 | } |
| 4166 | return fmt.Sprintf("turn:%d", runSeq) |
| 4167 | } |
| 4168 | |
| 4169 | func (a *Agent) readOnlyExecutionBlock(visible tool.Tool, resolved *tool.ResolvedCall) (toolOutcome, bool) { |
| 4170 | if a == nil || !a.readOnlyExecution { |
| 4171 | return toolOutcome{}, false |
| 4172 | } |
| 4173 | block := func(reason string) (toolOutcome, bool) { |
| 4174 | return toolOutcome{ |
| 4175 | output: "blocked: read-only agent cannot " + reason, |
| 4176 | blocked: true, |
| 4177 | errMsg: "blocked by read-only execution boundary", |
| 4178 | }, true |
| 4179 | } |
| 4180 | // Destructive MCP is left for the Executor; Planner must not misread this |
| 4181 | // as missing configuration or an unavailable MCP server. |
| 4182 | blockDestructiveForExecutor := func(name string) (toolOutcome, bool) { |
| 4183 | msg := "blocked: MCP capability " + name + " is destructive and is reserved for the Executor. Write the required operation into the plan/handoff so the Coordinator can hand it to the Executor; do not treat this as missing MCP configuration or an unavailable capability." |
| 4184 | return toolOutcome{ |
| 4185 | output: msg, |
| 4186 | blocked: true, |
| 4187 | errMsg: "blocked: destructive MCP reserved for executor", |
| 4188 | }, true |
| 4189 | } |
| 4190 | if resolved == nil { |
| 4191 | if a.plannerMCPExecution && isMCPExecutionTarget(visible, "") { |
| 4192 | if !mcpServerAuthorized(visible) { |
| 4193 | return block("execute an MCP capability from an unauthorized server") |
| 4194 | } |
| 4195 | if readOnlyExecutionMCPDestructive(visible) { |
| 4196 | return blockDestructiveForExecutor(visible.Name()) |
| 4197 | } |
| 4198 | return toolOutcome{}, false |
| 4199 | } |
| 4200 | if visible == nil || !visible.ReadOnly() { |
| 4201 | if reasoner, ok := visible.(tool.ReadOnlyExecutionBlockReason); ok && strings.TrimSpace(reasoner.ReadOnlyExecutionBlockReason()) != "" { |
| 4202 | return block(reasoner.ReadOnlyExecutionBlockReason()) |
| 4203 | } |
| 4204 | return block("execute a state-changing tool") |
| 4205 | } |
| 4206 | if isInstalledMCPTool(visible) && !mcpServerAuthorized(visible) { |
| 4207 | return block("execute a reader from an unauthorized MCP server") |
| 4208 | } |
| 4209 | if readOnlyExecutionMCPDestructive(visible) { |
| 4210 | return block("execute a destructive MCP capability") |
| 4211 | } |
| 4212 | if h, ok := visible.(tool.ReadOnlyExecutionHostMutation); ok && h.ReadOnlyExecutionHostMutation() && !readOnlyExecutionAllowsMCPStartup(visible) { |
| 4213 | return block("start or mutate a host capability") |
| 4214 | } |
| 4215 | return toolOutcome{}, false |
| 4216 | } |
| 4217 | |
| 4218 | switch resolved.ProxyAction { |
| 4219 | case "list", "inspect": |
| 4220 | if !resolved.SkipExecute || resolved.Target != nil || !resolved.ReadOnly { |
| 4221 | return block("execute a malformed dynamic inspection") |
| 4222 | } |
| 4223 | return toolOutcome{}, false |
| 4224 | case "decline": |
| 4225 | return block("decline a capability decision") |
| 4226 | case "call": |
| 4227 | if resolved.Target == nil { |
| 4228 | if a.plannerMCPExecution && resolved.HostCompleted && resolved.SkipExecute && resolved.ReadOnly && !resolved.Unavailable { |
| 4229 | if _, ok := parseMCPServerCapabilityID(resolved.CapabilityID); ok { |
| 4230 | return toolOutcome{}, false |
| 4231 | } |
| 4232 | } |
| 4233 | return block("execute an unresolved dynamic capability") |
| 4234 | } |
| 4235 | if a.plannerMCPExecution && plannerAllowsMCPTarget(resolved.Target, resolved.TargetName) { |
| 4236 | if isMCPLifecycleConnectTarget(resolved.Target) { |
| 4237 | if !plannerMCPConnectAllowed(resolved.Target) { |
| 4238 | return block("start an unauthorized MCP server") |
| 4239 | } |
| 4240 | } else if !mcpServerAuthorized(resolved.Target) { |
| 4241 | return block("execute an MCP capability from an unauthorized server") |
| 4242 | } |
| 4243 | if readOnlyExecutionMCPDestructive(resolved.Target) { |
| 4244 | name := resolved.TargetName |
| 4245 | if name == "" { |
| 4246 | name = resolved.CapabilityID |
| 4247 | } |
| 4248 | return blockDestructiveForExecutor(name) |
| 4249 | } |
| 4250 | return toolOutcome{}, false |
| 4251 | } |
| 4252 | if !resolved.ReadOnly { |
| 4253 | if reasoner, ok := resolved.Target.(tool.ReadOnlyExecutionBlockReason); ok && strings.TrimSpace(reasoner.ReadOnlyExecutionBlockReason()) != "" { |
| 4254 | return block(reasoner.ReadOnlyExecutionBlockReason()) |
| 4255 | } |
| 4256 | return block("execute a state-changing dynamic capability") |
| 4257 | } |
| 4258 | if isInstalledMCPTool(resolved.Target) && !mcpServerAuthorized(resolved.Target) { |
| 4259 | return block("execute a dynamic reader from an unauthorized MCP server") |
| 4260 | } |
| 4261 | if readOnlyExecutionMCPDestructive(resolved.Target) { |
| 4262 | return block("execute a destructive MCP capability") |
| 4263 | } |
| 4264 | if h, ok := resolved.Target.(tool.ReadOnlyExecutionHostMutation); ok && h.ReadOnlyExecutionHostMutation() && !readOnlyExecutionAllowsMCPStartup(resolved.Target) { |
| 4265 | return block("start or mutate a host capability") |
| 4266 | } |
| 4267 | return toolOutcome{}, false |
| 4268 | default: |
| 4269 | return block("execute an unknown dynamic capability action") |
| 4270 | } |
| 4271 | } |
| 4272 | |
| 4273 | func readOnlyExecutionMCPDestructive(t tool.Tool) bool { |
| 4274 | return mcpDestructiveHint(t) |
| 4275 | } |
| 4276 | |
| 4277 | func readOnlyExecutionAllowsMCPStartup(t tool.Tool) bool { |
| 4278 | if t == nil || !t.ReadOnly() || readOnlyExecutionMCPDestructive(t) { |
| 4279 | return false |
| 4280 | } |
| 4281 | if !mcpServerAuthorized(t) { |
| 4282 | return false |
| 4283 | } |
| 4284 | meta, ok := t.(tool.MCPMetadata) |
| 4285 | if !ok || strings.TrimSpace(meta.MCPServerName()) == "" || strings.TrimSpace(meta.MCPRawToolName()) == "" { |
| 4286 | return false |
| 4287 | } |
| 4288 | return true |
| 4289 | } |
| 4290 | |
| 4291 | // plannerAllowsMCPTarget reports whether a resolved use_capability target is an |
| 4292 | // MCP tool or lifecycle connect that Planner may consider under |
| 4293 | // PlannerMCPExecution (authorization and destructive checks run separately). |
| 4294 | func plannerAllowsMCPTarget(t tool.Tool, targetName string) bool { |
| 4295 | if t == nil { |
| 4296 | return false |
| 4297 | } |
| 4298 | if isInstalledMCPTool(t) || isMCPLifecycleConnectTarget(t) { |
| 4299 | return true |
| 4300 | } |
| 4301 | return isMCPExecutionTarget(t, targetName) |
| 4302 | } |
| 4303 | |
| 4304 | // isMCPLifecycleConnectTarget identifies on-demand MCP connect-and-list targets |
| 4305 | // (mcp_connect__<server>) used by use_capability action=call on mcp-server ids. |
| 4306 | func isMCPLifecycleConnectTarget(t tool.Tool) bool { |
| 4307 | if t == nil { |
| 4308 | return false |
| 4309 | } |
| 4310 | if _, ok := t.(mcpLifecycleConnect); ok { |
| 4311 | return true |
| 4312 | } |
| 4313 | name := strings.TrimSpace(t.Name()) |
| 4314 | return strings.HasPrefix(name, "mcp_connect__") |
| 4315 | } |
| 4316 | |
| 4317 | // mcpLifecycleConnect is implemented by deferred connect targets so Planner |
| 4318 | // can authorize lifecycle actions without relying on name prefixes alone. |
| 4319 | type mcpLifecycleConnect interface { |
| 4320 | MCPLifecycleConnect() bool |
| 4321 | MCPServerAuthorized() bool |
| 4322 | } |
| 4323 | |
| 4324 | func plannerMCPConnectAllowed(t tool.Tool) bool { |
| 4325 | if life, ok := t.(mcpLifecycleConnect); ok { |
| 4326 | return life.MCPServerAuthorized() |
| 4327 | } |
| 4328 | return mcpServerAuthorized(t) |
| 4329 | } |
| 4330 | |
| 4331 | func isInstalledMCPTool(t tool.Tool) bool { |
| 4332 | meta, ok := t.(tool.MCPMetadata) |
| 4333 | return ok && strings.TrimSpace(meta.MCPServerName()) != "" && strings.TrimSpace(meta.MCPRawToolName()) != "" |
| 4334 | } |
| 4335 | |
| 4336 | func isMCPExecutionTarget(t tool.Tool, name string) bool { |
| 4337 | return isInstalledMCPTool(t) || strings.HasPrefix(strings.TrimSpace(name), "mcp__") |
| 4338 | } |
| 4339 | |
| 4340 | func mcpServerAuthorized(t tool.Tool) bool { |
| 4341 | authority, ok := t.(tool.MCPServerAuthorization) |
| 4342 | return ok && authority.MCPServerAuthorized() |
| 4343 | } |
| 4344 | |
| 4345 | func mcpDestructiveHint(t tool.Tool) bool { |
| 4346 | annotations, ok := t.(tool.MCPAnnotations) |
| 4347 | return ok && annotations.MCPDestructiveHint() |
| 4348 | } |
| 4349 | |
| 4350 | func (a *Agent) planModeDecision(toolName string, readOnly bool, safety planmode.PlanSafety, args json.RawMessage) planmode.Decision { |
| 4351 | return (planmode.Policy{}).Decide(planmode.Call{ |
| 4352 | Name: toolName, |
| 4353 | ReadOnly: readOnly, |
| 4354 | Safety: safety, |
| 4355 | Args: args, |
| 4356 | }) |
| 4357 | } |
| 4358 | |
| 4359 | func (a *Agent) repeatedSuccessBlock(call provider.ToolCall, t tool.Tool) (string, bool) { |
| 4360 | sig, ok := repeatSuccessSignature(call, t) |
| 4361 | if !ok || a.repeatSuccessCounts == nil { |
| 4362 | return "", false |
| 4363 | } |
| 4364 | count := a.repeatSuccessCounts[sig] |
| 4365 | if count < repeatSuccessBreakThreshold { |
| 4366 | return "", false |
| 4367 | } |
| 4368 | return fmt.Sprintf( |
| 4369 | "blocked: [loop guard] %q has already succeeded %d times with the same write-like arguments in this user turn. Re-running it is unlikely to help and may burn tokens or repeat file writes. Change approach: use edit_file or multi_edit for file changes, verify with a read/test command, or explain the blocker in your final answer.", |
| 4370 | call.Name, count), true |
| 4371 | } |
| 4372 | |
| 4373 | func (a *Agent) staleAnchorEditBlock(call provider.ToolCall) (string, bool) { |
| 4374 | if a.evidence == nil || !anchorBasedEditTool(call.Name) { |
| 4375 | return "", false |
| 4376 | } |
| 4377 | rec := evidence.ReceiptFromToolCall(call.Name, json.RawMessage(call.Arguments), true, false) |
| 4378 | if len(rec.Paths) == 0 { |
| 4379 | return "", false |
| 4380 | } |
| 4381 | writeIndex, ok := a.evidence.LatestSuccessfulWriteIndex(rec.Paths) |
| 4382 | if !ok || a.evidence.HasSuccessfulAnchorRefreshReadAfter(rec.Paths, writeIndex) { |
| 4383 | return "", false |
| 4384 | } |
| 4385 | return fmt.Sprintf( |
| 4386 | "blocked: [fresh read required] %q targets %s, which was already modified earlier this turn. Re-read the current file with read_file without offset/limit before another range deletion, or use multi_edit with exact replacements when possible. This prevents stale start/end anchors from selecting an unintended destructive span.", |
| 4387 | call.Name, strings.Join(rec.Paths, ", ")), true |
| 4388 | } |
| 4389 | |
| 4390 | func anchorBasedEditTool(name string) bool { |
| 4391 | switch name { |
| 4392 | // edit_file synchronously reads the current file, requires a unique exact |
| 4393 | // or narrowly fuzzy match, and returns the actual applied diff. Let it try |
| 4394 | // optimistically; a stale old_string fails without writing and tells the |
| 4395 | // model to re-read. delete_range remains guarded because two independently |
| 4396 | // resolved anchors can otherwise select an unintended destructive span. |
| 4397 | case "delete_range": |
| 4398 | return true |
| 4399 | default: |
| 4400 | return false |
| 4401 | } |
| 4402 | } |
| 4403 | |
| 4404 | func (a *Agent) recordRepeatSuccess(call provider.ToolCall, t tool.Tool) { |
| 4405 | sig, ok := repeatSuccessSignature(call, t) |
| 4406 | if !ok { |
| 4407 | return |
| 4408 | } |
| 4409 | if a.repeatSuccessCounts == nil { |
| 4410 | a.repeatSuccessCounts = make(map[string]int) |
| 4411 | } |
| 4412 | a.repeatSuccessCounts[sig]++ |
| 4413 | } |
| 4414 | |
| 4415 | func repeatSuccessSignature(call provider.ToolCall, t tool.Tool) (string, bool) { |
| 4416 | if t.ReadOnly() { |
| 4417 | return "", false |
| 4418 | } |
| 4419 | switch call.Name { |
| 4420 | case "write_file", "edit_file", "multi_edit", "move_file", "notebook_edit": |
| 4421 | return call.Name + "\x00" + canonicalToolArgs(call.Arguments), true |
| 4422 | case "bash": |
| 4423 | var p struct { |
| 4424 | Command string `json:"command"` |
| 4425 | RunInBackground bool `json:"run_in_background"` |
| 4426 | } |
| 4427 | if err := json.Unmarshal([]byte(call.Arguments), &p); err != nil { |
| 4428 | return "", false |
| 4429 | } |
| 4430 | if p.RunInBackground || !isShellFileWriteCommand(p.Command) { |
| 4431 | return "", false |
| 4432 | } |
| 4433 | return "bash\x00" + normalizeShellCommand(p.Command), true |
| 4434 | default: |
| 4435 | return "", false |
| 4436 | } |
| 4437 | } |
| 4438 | |
| 4439 | func canonicalToolArgs(raw string) string { |
| 4440 | var v any |
| 4441 | if err := json.Unmarshal([]byte(raw), &v); err != nil { |
| 4442 | return strings.TrimSpace(raw) |
| 4443 | } |
| 4444 | b, err := json.Marshal(v) |
| 4445 | if err != nil { |
| 4446 | return strings.TrimSpace(raw) |
| 4447 | } |
| 4448 | var compact bytes.Buffer |
| 4449 | if err := json.Compact(&compact, b); err != nil { |
| 4450 | return string(b) |
| 4451 | } |
| 4452 | return compact.String() |
| 4453 | } |
| 4454 | |
| 4455 | func normalizeShellCommand(command string) string { |
| 4456 | if fields, malformed := shellparse.StaticFields(command); malformed == "" && len(fields) > 0 { |
| 4457 | return strings.Join(fields, " ") |
| 4458 | } |
| 4459 | return strings.Join(strings.Fields(command), " ") |
| 4460 | } |
| 4461 | |
| 4462 | func isShellFileWriteCommand(command string) bool { |
| 4463 | lower := strings.ToLower(command) |
| 4464 | switch { |
| 4465 | case shellPythonOpenWrites(lower): |
| 4466 | return true |
| 4467 | case strings.Contains(lower, "set-content") || strings.Contains(lower, "add-content") || strings.Contains(lower, "out-file"): |
| 4468 | return true |
| 4469 | case strings.Contains(lower, "sed -i") || strings.Contains(lower, "perl -pi"): |
| 4470 | return true |
| 4471 | case hasShellWriteRedirect(command): |
| 4472 | return true |
| 4473 | default: |
| 4474 | return false |
| 4475 | } |
| 4476 | } |
| 4477 | |
| 4478 | func shellPythonOpenWrites(lower string) bool { |
| 4479 | if !strings.Contains(lower, "open(") { |
| 4480 | return false |
| 4481 | } |
| 4482 | if strings.Contains(lower, ".write(") { |
| 4483 | return true |
| 4484 | } |
| 4485 | for _, marker := range []string{", 'w", `, "w`, ", 'a", `, "a`, ", 'x", `, "x`, "mode='w", `mode="w`, "mode='a", `mode="a`, "mode='x", `mode="x`} { |
| 4486 | if strings.Contains(lower, marker) { |
| 4487 | return true |
| 4488 | } |
| 4489 | } |
| 4490 | return false |
| 4491 | } |
| 4492 | |
| 4493 | func hasShellWriteRedirect(command string) bool { |
| 4494 | file, err := shellparse.ParseBash(command) |
| 4495 | if err == nil { |
| 4496 | hasWrite := false |
| 4497 | syntax.Walk(file, func(node syntax.Node) bool { |
| 4498 | redir, ok := node.(*syntax.Redirect) |
| 4499 | if !ok { |
| 4500 | return true |
| 4501 | } |
| 4502 | if bashRedirectWritesFile(command, redir) { |
| 4503 | hasWrite = true |
| 4504 | return false |
| 4505 | } |
| 4506 | return true |
| 4507 | }) |
| 4508 | return hasWrite |
| 4509 | } |
| 4510 | return hasShellWriteRedirectFallback(command) |
| 4511 | } |
| 4512 | |
| 4513 | func bashRedirectWritesFile(source string, redir *syntax.Redirect) bool { |
| 4514 | if redir == nil { |
| 4515 | return false |
| 4516 | } |
| 4517 | switch redir.Op { |
| 4518 | case syntax.RdrOut, syntax.AppOut, syntax.RdrClob, syntax.AppClob, |
| 4519 | syntax.RdrAll, syntax.RdrAllClob, syntax.AppAll, syntax.AppAllClob, |
| 4520 | syntax.RdrInOut: |
| 4521 | return !redirectWordIsNullSink(source, redir.Word) |
| 4522 | default: |
| 4523 | return false |
| 4524 | } |
| 4525 | } |
| 4526 | |
| 4527 | func redirectWordIsNullSink(source string, word *syntax.Word) bool { |
| 4528 | if word == nil { |
| 4529 | return false |
| 4530 | } |
| 4531 | if value, ok := shellparse.StaticWord(word); ok { |
| 4532 | if isNullSinkWord(strings.TrimSpace(value)) { |
| 4533 | return true |
| 4534 | } |
| 4535 | } |
| 4536 | value := strings.TrimSpace(redirectWordSource(source, word)) |
| 4537 | if isNullSinkWord(value) { |
| 4538 | return true |
| 4539 | } |
| 4540 | if len(value) >= 2 && ((value[0] == '\'' && value[len(value)-1] == '\'') || (value[0] == '"' && value[len(value)-1] == '"')) { |
| 4541 | return isNullSinkWord(value[1 : len(value)-1]) |
| 4542 | } |
| 4543 | return false |
| 4544 | } |
| 4545 | |
| 4546 | func isNullSinkWord(value string) bool { |
| 4547 | if value == "/dev/null" { |
| 4548 | return true |
| 4549 | } |
| 4550 | return strings.EqualFold(value, "$null") || strings.EqualFold(value, "nul") |
| 4551 | } |
| 4552 | |
| 4553 | func redirectWordSource(source string, word *syntax.Word) string { |
| 4554 | if word == nil || !word.Pos().IsValid() || !word.End().IsValid() { |
| 4555 | return "" |
| 4556 | } |
| 4557 | start := int(word.Pos().Offset()) |
| 4558 | end := int(word.End().Offset()) |
| 4559 | if start < 0 || end < start || end > len(source) { |
| 4560 | return "" |
| 4561 | } |
| 4562 | return source[start:end] |
| 4563 | } |
| 4564 | |
| 4565 | func hasShellWriteRedirectFallback(command string) bool { |
| 4566 | var quote rune |
| 4567 | var prev rune |
| 4568 | for _, r := range command { |
| 4569 | if quote != 0 { |
| 4570 | if r == quote { |
| 4571 | quote = 0 |
| 4572 | } |
| 4573 | prev = r |
| 4574 | continue |
| 4575 | } |
| 4576 | if r == '\'' || r == '"' { |
| 4577 | quote = r |
| 4578 | prev = r |
| 4579 | continue |
| 4580 | } |
| 4581 | if r == '>' { |
| 4582 | if prev == '2' { |
| 4583 | prev = r |
| 4584 | continue |
| 4585 | } |
| 4586 | return true |
| 4587 | } |
| 4588 | prev = r |
| 4589 | } |
| 4590 | return false |
| 4591 | } |
| 4592 | |
| 4593 | // isBackgroundTaskCall reports whether a `task` call set run_in_background, so a |
| 4594 | // fire-and-return dispatch isn't mistaken for a sub-agent that has stopped. |
| 4595 | func isBackgroundTaskCall(args string) bool { |
| 4596 | var p struct { |
| 4597 | RunInBackground bool `json:"run_in_background"` |
| 4598 | } |
| 4599 | _ = json.Unmarshal([]byte(args), &p) |
| 4600 | return p.RunInBackground |
| 4601 | } |
| 4602 | |
| 4603 | // toolReadOnly reports a tool's ReadOnly classification by name (false for an |
| 4604 | // unknown tool), for stamping early ToolDispatch events. |
| 4605 | func (a *Agent) toolReadOnly(name string) bool { |
| 4606 | t, _, ambiguous := a.tools.ResolveCall(name) |
| 4607 | return t != nil && len(ambiguous) == 0 && t.ReadOnly() |
| 4608 | } |
| 4609 | |
| 4610 | // firstLine returns s up to its first newline — a one-line failure summary for |
| 4611 | // the display Err, while the full error stays in the model-facing output. |
| 4612 | func firstLine(s string) string { |
| 4613 | if i := strings.IndexByte(s, '\n'); i >= 0 { |
| 4614 | return s[:i] |
| 4615 | } |
| 4616 | return s |
| 4617 | } |
| 4618 | |
| 4619 | // truncateToolOutput head+tails s when it exceeds maxToolOutputBytes, slicing |
| 4620 | // on rune boundaries so we never split a multibyte glyph. Returns the possibly |
| 4621 | // trimmed body plus a one-line user-facing notice when truncation happened |
| 4622 | // (empty when it didn't, without the "· " display prefix). |
| 4623 | func truncateToolOutput(s string) (string, string) { |
| 4624 | if len(s) <= maxToolOutputBytes { |
| 4625 | return s, "" |
| 4626 | } |
| 4627 | keep := maxToolOutputBytes / 2 |
| 4628 | head := snapToRuneBoundary(s, 0, keep) |
| 4629 | tail := snapToRuneBoundary(s, len(s)-keep, len(s)) |
| 4630 | omitted := len(s) - len(head) - len(tail) |
| 4631 | notice := fmt.Sprintf("tool output truncated: %d of %d bytes elided", omitted, len(s)) |
| 4632 | body := head + fmt.Sprintf("\n\n…[truncated %d of %d bytes — rerun with narrower args to see the middle]…\n\n", omitted, len(s)) + tail |
| 4633 | return body, notice |
| 4634 | } |
| 4635 | |
| 4636 | // snapToRuneBoundary returns s[lo:hi] with the bounds nudged outward until |
| 4637 | // both land on rune-start positions. |
| 4638 | func snapToRuneBoundary(s string, lo, hi int) string { |
| 4639 | for lo > 0 && !utf8.RuneStart(s[lo]) { |
| 4640 | lo-- |
| 4641 | } |
| 4642 | for hi < len(s) && !utf8.RuneStart(s[hi]) { |
| 4643 | hi++ |
| 4644 | } |
| 4645 | return s[lo:hi] |
| 4646 | } |
| 4647 | |
| 4648 | // finishReasonMessage maps an abnormal finish_reason to a one-line warning, |
| 4649 | // returning ok=false for the normal terminations ("stop", "tool_calls") and a |
| 4650 | // nil usage. The sink renders the message; the "! " prefix is presentation. |
| 4651 | func finishReasonMessage(u *provider.Usage) (string, bool) { |
| 4652 | if u == nil { |
| 4653 | return "", false |
| 4654 | } |
| 4655 | switch u.FinishReason { |
| 4656 | case "length": |
| 4657 | return "response truncated: hit max output tokens", true |
| 4658 | case finishReasonClientReasoningLimit: |
| 4659 | return "response stopped: hit the client reasoning safety limit", true |
| 4660 | case "content_filter": |
| 4661 | return "response blocked by content filter", true |
| 4662 | case "repetition_truncation": |
| 4663 | return "response truncated: model repetition detected", true |
| 4664 | default: |
| 4665 | return "", false |
| 4666 | } |
| 4667 | } |
| 4668 |