| 1 | package control |
| 2 | |
| 3 | import ( |
| 4 | "crypto/rand" |
| 5 | "encoding/json" |
| 6 | "fmt" |
| 7 | "log/slog" |
| 8 | "os" |
| 9 | "path/filepath" |
| 10 | "strings" |
| 11 | "sync" |
| 12 | "time" |
| 13 | |
| 14 | "reasonix/internal/agent" |
| 15 | "reasonix/internal/evidence" |
| 16 | "reasonix/internal/fileutil" |
| 17 | fileencoding "reasonix/internal/fileutil/encoding" |
| 18 | "reasonix/internal/goaleval" |
| 19 | "reasonix/internal/store" |
| 20 | "reasonix/internal/tool" |
| 21 | ) |
| 22 | |
| 23 | const ( |
| 24 | goalContinueTurn = "Continue pursuing the active goal under its task contract. Do the next useful work, then call update_goal with your disposition: continue (include the next concrete step in next_action), complete (only when fully done and verified), or blocked (when only the user can unblock)." |
| 25 | goalCompleteNotice = "goal complete" |
| 26 | |
| 27 | // Default no-progress stall limit: four consecutive goal turns without any |
| 28 | // host-verifiable progress pause the goal. |
| 29 | defaultNoProgressLimit = 4 |
| 30 | ) |
| 31 | |
| 32 | // Budget classes select turn quotas only. They never gate permissions, writes, |
| 33 | // or provider request admission. Token usage is still accumulated for display. |
| 34 | const ( |
| 35 | budgetClassSimple = "simple" |
| 36 | budgetClassWrite = "write" |
| 37 | budgetClassResearch = "research" |
| 38 | ) |
| 39 | |
| 40 | // Stop causes distinguish a safe pause (blocked + stopCause) from a genuine |
| 41 | // task block (blocked with empty stopCause). Old clients see blocked either |
| 42 | // way and never fail open. |
| 43 | // |
| 44 | // stopCauseBudgetTokens is retained only so old sidecars that paused on the |
| 45 | // removed token hard-limit can be recognized and auto-resumed on load. |
| 46 | const ( |
| 47 | stopCauseBudgetTurns = "budget_turns" |
| 48 | stopCauseBudgetTokens = "budget_tokens" // legacy; never written by current runtime |
| 49 | stopCauseNoProgress = "no_progress" |
| 50 | stopCauseEvaluator = "evaluator_unavailable" |
| 51 | stopCauseManual = "manual" |
| 52 | ) |
| 53 | |
| 54 | // budgetQuota returns the default turn quota for a budget class. Token hard |
| 55 | // limits were removed; callers no longer receive a token ceiling. |
| 56 | func budgetQuota(class string) (turns int) { |
| 57 | switch class { |
| 58 | case budgetClassResearch: |
| 59 | return 40 |
| 60 | case budgetClassWrite: |
| 61 | return 20 |
| 62 | default: |
| 63 | return 10 |
| 64 | } |
| 65 | } |
| 66 | |
| 67 | // budgetClassFor derives a goal's budget class: AutoResearch always means |
| 68 | // research; otherwise Goal-specific write classification decides whether the |
| 69 | // objective is a write turn budget (including bare fault statements) or simple. |
| 70 | // Ordinary Delivery consultation/diagnosis classification is unchanged. |
| 71 | func budgetClassFor(goal string, researchMode GoalResearchMode) string { |
| 72 | if shouldUseAutoResearch(goal, researchMode) { |
| 73 | return budgetClassResearch |
| 74 | } |
| 75 | if agent.GoalTaskNeedsWriteBudget(goal) { |
| 76 | return budgetClassWrite |
| 77 | } |
| 78 | return budgetClassSimple |
| 79 | } |
| 80 | |
| 81 | // goalMachine owns the active goal's finite-state machine and its persistence. |
| 82 | // It is a strict leaf: its methods take only the machine's own locks and never |
| 83 | // call back into the Controller, so the controller may hold c.mu while invoking |
| 84 | // a getter without risking lock inversion. The FSM is pure — advance() takes |
| 85 | // already-gathered inputs (the update_goal report, readiness, evaluator verdict, |
| 86 | // budget/progress state) and returns what to persist plus a notice, so no disk |
| 87 | // or executor work happens under mu. |
| 88 | type goalMachine struct { |
| 89 | // mu guards the FSM fields below; every critical section under it is short |
| 90 | // and non-blocking (no disk I/O, no executor calls). |
| 91 | mu sync.Mutex |
| 92 | goal string |
| 93 | status string |
| 94 | researchMode GoalResearchMode |
| 95 | autoResearchTaskID string |
| 96 | scopeID string |
| 97 | deliveryCheckpoint evidence.DeliveryCheckpoint |
| 98 | block string |
| 99 | strict bool |
| 100 | continuationEpoch uint64 |
| 101 | |
| 102 | // Runtime budget state, persisted across turns and restarts. |
| 103 | // tokensUsed is observational only (no hard limit). tokensLimit is kept at |
| 104 | // 0 for wire/sidecar compatibility and is never enforced. |
| 105 | budgetClass string |
| 106 | turnsUsed int |
| 107 | turnsLimit int |
| 108 | tokensUsed int |
| 109 | tokensLimit int // always 0 at runtime; deprecated hard limit |
| 110 | noProgressTurns int |
| 111 | noProgressLimit int |
| 112 | lastContinuationReason string |
| 113 | lastEvaluatorReason string |
| 114 | stopCause string |
| 115 | budgetExtensions int // turn extensions from resume (compat field name) |
| 116 | |
| 117 | // statePath is the persisted goal-state sidecar; empty disables persistence. |
| 118 | statePath string |
| 119 | // writeMu serializes goal-state disk writes so concurrent saves don't |
| 120 | // interleave or land out of order. Taken OFF mu by writeState. |
| 121 | writeMu sync.Mutex |
| 122 | } |
| 123 | |
| 124 | // goalState is the serializable form of a running goal. New fields are |
| 125 | // safe-to-omit JSON: old readers ignore them, and restoreFromState re-derives |
| 126 | // defaults when they are missing. |
| 127 | type goalState struct { |
| 128 | Goal string `json:"goal,omitempty"` |
| 129 | Status string `json:"status,omitempty"` |
| 130 | ResearchMode GoalResearchMode `json:"researchMode,omitempty"` |
| 131 | AutoResearchTaskID string `json:"autoResearchTaskID,omitempty"` |
| 132 | ScopeID string `json:"scopeID,omitempty"` |
| 133 | DeliveryCheckpoint evidence.DeliveryCheckpoint `json:"deliveryCheckpoint,omitempty"` |
| 134 | Turns int `json:"turns,omitempty"` |
| 135 | Blocks int `json:"blocks,omitempty"` |
| 136 | Block string `json:"block,omitempty"` |
| 137 | Strict bool `json:"strict,omitempty"` |
| 138 | Todos []evidence.TodoItem `json:"todos,omitempty"` |
| 139 | |
| 140 | BudgetClass string `json:"budgetClass,omitempty"` |
| 141 | TurnsUsed int `json:"turnsUsed,omitempty"` |
| 142 | TurnsLimit int `json:"turnsLimit,omitempty"` |
| 143 | TokensUsed int `json:"tokensUsed,omitempty"` |
| 144 | TokensLimit int `json:"tokensLimit,omitempty"` |
| 145 | NoProgressTurns int `json:"noProgressTurns,omitempty"` |
| 146 | NoProgressLimit int `json:"noProgressLimit,omitempty"` |
| 147 | LastContinuationReason string `json:"lastContinuationReason,omitempty"` |
| 148 | LastEvaluatorReason string `json:"lastEvaluatorReason,omitempty"` |
| 149 | StopCause string `json:"stopCause,omitempty"` |
| 150 | BudgetExtensions int `json:"budgetExtensions,omitempty"` |
| 151 | } |
| 152 | |
| 153 | // goalMachineSnapshot is an in-memory rollback point for durable Goal updates. |
| 154 | // Persistence paths and mutexes are deliberately excluded. |
| 155 | type goalMachineSnapshot struct { |
| 156 | goal string |
| 157 | status string |
| 158 | researchMode GoalResearchMode |
| 159 | autoResearchTaskID string |
| 160 | scopeID string |
| 161 | deliveryCheckpoint evidence.DeliveryCheckpoint |
| 162 | block string |
| 163 | strict bool |
| 164 | } |
| 165 | |
| 166 | // goalAdvanceInput carries everything the FSM needs for one continuation step, |
| 167 | // gathered by the caller off the machine's lock. The FSM is the exclusive |
| 168 | // decision point: it applies readiness, budget, and no-progress gates and |
| 169 | // decides complete / continue / blocked / pause. |
| 170 | type goalAdvanceInput struct { |
| 171 | report *goalTurnReport // validated update_goal report; nil when none |
| 172 | readiness agent.ReadinessResult |
| 173 | evaluator *goalEvaluatorVerdict // evaluator verdict; nil when not run |
| 174 | evaluatorFailed string // evaluator error/timeout text; pause fail-closed |
| 175 | todos []evidence.TodoItem |
| 176 | progressBefore string // host progress signature captured before the turn |
| 177 | progressAfter string // host progress signature captured after the turn |
| 178 | expectedEpoch *uint64 |
| 179 | } |
| 180 | |
| 181 | // goalEvaluatorVerdict is the bounded evaluator's structured outcome. |
| 182 | type goalEvaluatorVerdict struct { |
| 183 | outcome goaleval.Outcome |
| 184 | reason string |
| 185 | } |
| 186 | |
| 187 | // goalAdvanceResult reports the FSM step's outcome. data/path/ok describe the |
| 188 | // state to persist (built under mu when something changed); notice is surfaced |
| 189 | // to the user; cont reports whether the goal loop should continue; intercept |
| 190 | // (with interceptNotice) is the next synthetic turn's prompt. |
| 191 | type goalAdvanceResult struct { |
| 192 | notice string |
| 193 | intercept string |
| 194 | interceptNotice string |
| 195 | cont bool |
| 196 | continuationEpoch uint64 |
| 197 | path string |
| 198 | data []byte |
| 199 | ok bool |
| 200 | } |
| 201 | |
| 202 | // goalContinuationSnapshot binds a continuation to the exact Goal lifecycle |
| 203 | // state admitted for its synthetic turn. The orchestrator uses these captured |
| 204 | // fields throughout the turn instead of re-reading a possibly replaced Goal. |
| 205 | type goalContinuationSnapshot struct { |
| 206 | goal string |
| 207 | researchMode GoalResearchMode |
| 208 | autoResearchTaskID string |
| 209 | scopeID string |
| 210 | } |
| 211 | |
| 212 | // goalStatePath derives a session's persisted goal-state sidecar. |
| 213 | func goalStatePath(sessionPath string) string { |
| 214 | return store.SessionGoalState(sessionPath) |
| 215 | } |
| 216 | |
| 217 | func (g *goalMachine) setStatePath(path string) { |
| 218 | g.mu.Lock() |
| 219 | g.statePath = path |
| 220 | g.mu.Unlock() |
| 221 | } |
| 222 | |
| 223 | func (g *goalMachine) capture() goalMachineSnapshot { |
| 224 | g.mu.Lock() |
| 225 | defer g.mu.Unlock() |
| 226 | return goalMachineSnapshot{ |
| 227 | goal: g.goal, status: g.status, researchMode: g.researchMode, |
| 228 | autoResearchTaskID: g.autoResearchTaskID, scopeID: g.scopeID, |
| 229 | deliveryCheckpoint: g.deliveryCheckpoint, block: g.block, |
| 230 | strict: g.strict, |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | func (g *goalMachine) restore(snapshot goalMachineSnapshot) { |
| 235 | g.mu.Lock() |
| 236 | g.goal, g.status, g.researchMode = snapshot.goal, snapshot.status, snapshot.researchMode |
| 237 | g.autoResearchTaskID, g.scopeID = snapshot.autoResearchTaskID, snapshot.scopeID |
| 238 | g.deliveryCheckpoint, g.block = snapshot.deliveryCheckpoint, snapshot.block |
| 239 | g.strict = snapshot.strict |
| 240 | g.continuationEpoch++ |
| 241 | g.mu.Unlock() |
| 242 | } |
| 243 | |
| 244 | // snapshot returns the fields Compose injects into outgoing turns. |
| 245 | func (g *goalMachine) snapshot() (goal, status string, mode GoalResearchMode, autoResearchTaskID string) { |
| 246 | g.mu.Lock() |
| 247 | defer g.mu.Unlock() |
| 248 | return g.goal, g.status, g.researchMode, g.autoResearchTaskID |
| 249 | } |
| 250 | |
| 251 | func (g *goalMachine) goalText() string { |
| 252 | g.mu.Lock() |
| 253 | defer g.mu.Unlock() |
| 254 | return g.goal |
| 255 | } |
| 256 | |
| 257 | func (g *goalMachine) currentAutoResearchTaskID() string { |
| 258 | g.mu.Lock() |
| 259 | defer g.mu.Unlock() |
| 260 | if strings.TrimSpace(g.goal) == "" || g.status != GoalStatusRunning { |
| 261 | return "" |
| 262 | } |
| 263 | return g.autoResearchTaskID |
| 264 | } |
| 265 | |
| 266 | // continuationToken captures the Goal lifecycle that owns an outgoing turn. |
| 267 | // The matching assistant output may advance the FSM only while this epoch is |
| 268 | // still current. |
| 269 | func (g *goalMachine) continuationToken() uint64 { |
| 270 | g.mu.Lock() |
| 271 | defer g.mu.Unlock() |
| 272 | return g.continuationEpoch |
| 273 | } |
| 274 | |
| 275 | func (g *goalMachine) deliveryScope() (id, task string, ok bool) { |
| 276 | g.mu.Lock() |
| 277 | defer g.mu.Unlock() |
| 278 | if strings.TrimSpace(g.goal) == "" || g.status != GoalStatusRunning { |
| 279 | return "", "", false |
| 280 | } |
| 281 | if g.scopeID == "" { |
| 282 | g.scopeID = newGoalScopeID() |
| 283 | } |
| 284 | return g.scopeID, g.goal, true |
| 285 | } |
| 286 | |
| 287 | // goalScopeIDForTurn resolves the active goal scope for an outgoing turn: the |
| 288 | // continuation snapshot's scope, or the running goal's (assigning one when |
| 289 | // needed). ok=false means no active goal. |
| 290 | func (g *goalMachine) goalScopeIDForTurn(continuation *goalContinuationSnapshot) (string, bool) { |
| 291 | if continuation != nil { |
| 292 | return continuation.scopeID, true |
| 293 | } |
| 294 | id, _, ok := g.deliveryScope() |
| 295 | return id, ok |
| 296 | } |
| 297 | |
| 298 | func newGoalScopeID() string { |
| 299 | var raw [16]byte |
| 300 | if _, err := rand.Read(raw[:]); err == nil { |
| 301 | return fmt.Sprintf("goal-%x", raw[:]) |
| 302 | } |
| 303 | return fmt.Sprintf("goal-fallback-%d-%d", os.Getpid(), time.Now().UnixNano()) |
| 304 | } |
| 305 | |
| 306 | // active reports whether a goal is currently running. |
| 307 | func (g *goalMachine) active() bool { |
| 308 | g.mu.Lock() |
| 309 | defer g.mu.Unlock() |
| 310 | return strings.TrimSpace(g.goal) != "" && g.status == GoalStatusRunning |
| 311 | } |
| 312 | |
| 313 | // statusForDisplay maps the empty zero status to "stopped" for frontends. |
| 314 | func (g *goalMachine) statusForDisplay() string { |
| 315 | g.mu.Lock() |
| 316 | defer g.mu.Unlock() |
| 317 | if g.status == "" { |
| 318 | return GoalStatusStopped |
| 319 | } |
| 320 | return g.status |
| 321 | } |
| 322 | |
| 323 | // budgetExhausted reports whether the goal's turn budget is spent. Token usage |
| 324 | // never exhausts the goal by itself. |
| 325 | func (g *goalMachine) budgetExhausted() bool { |
| 326 | g.mu.Lock() |
| 327 | defer g.mu.Unlock() |
| 328 | return g.turnsLimit > 0 && g.turnsUsed >= g.turnsLimit |
| 329 | } |
| 330 | |
| 331 | // set installs a session-scoped goal (or clears it when goal is empty), resets |
| 332 | // the per-goal budget/runtime counters, and returns the state to persist. ok is |
| 333 | // false (no persistence) when the goal is unchanged or no state path is |
| 334 | // configured. |
| 335 | func (g *goalMachine) set(goal string, mode GoalResearchMode, autoResearchTaskID string, todos []evidence.TodoItem) (string, []byte, bool) { |
| 336 | goal = strings.TrimSpace(goal) |
| 337 | g.mu.Lock() |
| 338 | defer g.mu.Unlock() |
| 339 | if goal != "" && g.goal == goal && g.status == GoalStatusRunning && g.researchMode == mode && g.autoResearchTaskID == autoResearchTaskID { |
| 340 | return "", nil, false |
| 341 | } |
| 342 | g.continuationEpoch++ |
| 343 | g.turnsUsed, g.tokensUsed, g.noProgressTurns = 0, 0, 0 |
| 344 | g.block = "" |
| 345 | g.lastContinuationReason, g.lastEvaluatorReason = "", "" |
| 346 | g.stopCause = "" |
| 347 | g.budgetExtensions = 0 |
| 348 | if goal == "" { |
| 349 | g.goal, g.status, g.researchMode, g.autoResearchTaskID = "", GoalStatusStopped, GoalResearchAuto, "" |
| 350 | g.scopeID = "" |
| 351 | g.deliveryCheckpoint = evidence.DeliveryCheckpoint{} |
| 352 | } else { |
| 353 | g.goal, g.status, g.researchMode, g.autoResearchTaskID = goal, GoalStatusRunning, mode, autoResearchTaskID |
| 354 | g.scopeID = newGoalScopeID() |
| 355 | g.deliveryCheckpoint = evidence.DeliveryCheckpoint{ScopeID: g.scopeID} |
| 356 | g.budgetClass = budgetClassFor(goal, mode) |
| 357 | g.turnsLimit = budgetQuota(g.budgetClass) |
| 358 | g.tokensLimit = 0 // no token hard limit |
| 359 | g.noProgressLimit = defaultNoProgressLimit |
| 360 | } |
| 361 | return g.buildStateLocked(todos) |
| 362 | } |
| 363 | |
| 364 | func (g *goalMachine) setStrict(strict bool, todos []evidence.TodoItem) (string, []byte, bool) { |
| 365 | g.mu.Lock() |
| 366 | defer g.mu.Unlock() |
| 367 | g.strict = strict |
| 368 | return g.buildStateLocked(todos) |
| 369 | } |
| 370 | |
| 371 | // stop transitions a running goal to the given terminal status and clears the |
| 372 | // transient runtime bookkeeping. stopCause is cleared: a host stop is not a |
| 373 | // safe pause. |
| 374 | func (g *goalMachine) stop(status string, todos []evidence.TodoItem) (string, []byte, bool) { |
| 375 | g.mu.Lock() |
| 376 | defer g.mu.Unlock() |
| 377 | g.continuationEpoch++ |
| 378 | if strings.TrimSpace(g.goal) != "" && g.status == GoalStatusRunning { |
| 379 | g.status = status |
| 380 | } |
| 381 | g.stopCause = "" |
| 382 | g.noProgressTurns = 0 |
| 383 | return g.buildStateLocked(todos) |
| 384 | } |
| 385 | |
| 386 | // pauseFor transitions a running goal to a safe pause: status blocked plus a |
| 387 | // stop cause, keeping every budget/runtime counter for a later resume. |
| 388 | func (g *goalMachine) pauseFor(stopCause, reason string, todos []evidence.TodoItem) (string, []byte, bool) { |
| 389 | g.mu.Lock() |
| 390 | defer g.mu.Unlock() |
| 391 | g.continuationEpoch++ |
| 392 | if strings.TrimSpace(g.goal) != "" && g.status == GoalStatusRunning { |
| 393 | g.status = GoalStatusBlocked |
| 394 | } |
| 395 | g.stopCause = stopCause |
| 396 | if reason != "" { |
| 397 | g.block = reason |
| 398 | } |
| 399 | return g.buildStateLocked(todos) |
| 400 | } |
| 401 | |
| 402 | // resume re-enters a recoverable blocked/stopped goal without resetting its |
| 403 | // delivery evidence scope, AutoResearch identity, or runtime history. Turn |
| 404 | // budget pauses (and any pause whose original turn quota is already spent) |
| 405 | // append one turn slice of the current budget class; no-progress counting |
| 406 | // resets but accumulated token usage and budget_extensions are preserved. |
| 407 | // Token hard limits no longer exist, so resume never extends a token ceiling. |
| 408 | func (g *goalMachine) resume(todos []evidence.TodoItem) (path string, data []byte, persist, resumed, extended bool) { |
| 409 | g.mu.Lock() |
| 410 | defer g.mu.Unlock() |
| 411 | if strings.TrimSpace(g.goal) == "" || g.status == GoalStatusComplete { |
| 412 | return "", nil, false, false, false |
| 413 | } |
| 414 | // Legacy budget_tokens pauses are treated like turn-budget pauses so users |
| 415 | // can resume without understanding the removed hard limit. |
| 416 | extend := g.stopCause == stopCauseBudgetTurns || |
| 417 | g.stopCause == stopCauseBudgetTokens || |
| 418 | g.stopCause == stopCauseNoProgress || |
| 419 | (g.turnsLimit > 0 && g.turnsUsed >= g.turnsLimit) |
| 420 | g.continuationEpoch++ |
| 421 | g.status = GoalStatusRunning |
| 422 | g.block = "" |
| 423 | g.stopCause = "" |
| 424 | g.noProgressTurns = 0 |
| 425 | g.tokensLimit = 0 |
| 426 | if g.scopeID == "" { |
| 427 | g.scopeID = newGoalScopeID() |
| 428 | } |
| 429 | if extend { |
| 430 | if g.budgetClass == "" { |
| 431 | g.budgetClass = budgetClassFor(g.goal, g.researchMode) |
| 432 | } |
| 433 | g.turnsLimit += budgetQuota(g.budgetClass) |
| 434 | g.budgetExtensions++ |
| 435 | } |
| 436 | path, data, persist = g.buildStateLocked(todos) |
| 437 | return path, data, persist, true, extend |
| 438 | } |
| 439 | |
| 440 | func (g *goalMachine) setDeliveryCheckpoint(checkpoint evidence.DeliveryCheckpoint, todos []evidence.TodoItem) (string, []byte, bool) { |
| 441 | g.mu.Lock() |
| 442 | defer g.mu.Unlock() |
| 443 | if g.scopeID == "" || checkpoint.ScopeID != g.scopeID { |
| 444 | return "", nil, false |
| 445 | } |
| 446 | g.deliveryCheckpoint = checkpoint |
| 447 | return g.buildStateLocked(todos) |
| 448 | } |
| 449 | |
| 450 | func (g *goalMachine) deliveryState() evidence.DeliveryCheckpoint { |
| 451 | g.mu.Lock() |
| 452 | defer g.mu.Unlock() |
| 453 | return g.deliveryCheckpoint |
| 454 | } |
| 455 | |
| 456 | // acceptContinuation checks an advance result before the orchestrator surfaces |
| 457 | // its notice. admitContinuation revalidates after synchronous notice callbacks |
| 458 | // and captures the Goal state at the synthetic-turn admission boundary. |
| 459 | func (g *goalMachine) acceptContinuation(res goalAdvanceResult) (string, bool) { |
| 460 | g.mu.Lock() |
| 461 | defer g.mu.Unlock() |
| 462 | if !res.cont || |
| 463 | res.continuationEpoch != g.continuationEpoch || |
| 464 | strings.TrimSpace(g.goal) == "" || |
| 465 | g.status != GoalStatusRunning { |
| 466 | return "", false |
| 467 | } |
| 468 | return res.intercept, true |
| 469 | } |
| 470 | |
| 471 | // admitContinuation atomically validates an advance result and captures the |
| 472 | // Goal state used to compose and scope its synthetic turn. Keeping validation |
| 473 | // and capture in one critical section prevents a stale intercept from being |
| 474 | // paired with a replacement Goal between those operations. |
| 475 | func (g *goalMachine) admitContinuation(res goalAdvanceResult) (goalContinuationSnapshot, bool) { |
| 476 | g.mu.Lock() |
| 477 | defer g.mu.Unlock() |
| 478 | if !res.cont || |
| 479 | res.continuationEpoch != g.continuationEpoch || |
| 480 | strings.TrimSpace(g.goal) == "" || |
| 481 | g.status != GoalStatusRunning { |
| 482 | return goalContinuationSnapshot{}, false |
| 483 | } |
| 484 | if g.scopeID == "" { |
| 485 | g.scopeID = newGoalScopeID() |
| 486 | } |
| 487 | return goalContinuationSnapshot{ |
| 488 | goal: g.goal, |
| 489 | researchMode: g.researchMode, |
| 490 | autoResearchTaskID: g.autoResearchTaskID, |
| 491 | scopeID: g.scopeID, |
| 492 | }, true |
| 493 | } |
| 494 | |
| 495 | // advance runs one continuation step of the goal FSM from already-gathered |
| 496 | // inputs. It mutates the machine, decides whether to keep looping, and builds |
| 497 | // the state to persist when the goal reached a terminal/notice point. |
| 498 | // |
| 499 | // Decision priority (the FSM is the exclusive decision point): |
| 500 | // 1. complete + readiness ready (report or evaluator) → complete |
| 501 | // 2. blocked (report or evaluator) → blocked immediately (no triple confirm) |
| 502 | // 3. evaluator failed/uncertain → safe pause (fail closed, never default to continue) |
| 503 | // 4. evaluator failed/uncertain → safe pause (fail closed, never default to continue) |
| 504 | // 5. budget exhausted → safe pause (also vetoes complete claims rejected by |
| 505 | // readiness: those would continue, and continuation past the budget is a |
| 506 | // pause) |
| 507 | // 6. no-progress limit reached → safe pause |
| 508 | // 7. otherwise continue, carrying the missing requirements (complete rejected |
| 509 | // by readiness, or no report with an explicit missing list) or the report's |
| 510 | // next_action as the next turn's prompt. |
| 511 | func (g *goalMachine) advance(in goalAdvanceInput) goalAdvanceResult { |
| 512 | g.mu.Lock() |
| 513 | defer g.mu.Unlock() |
| 514 | if in.expectedEpoch != nil && *in.expectedEpoch != g.continuationEpoch { |
| 515 | return goalAdvanceResult{cont: false} |
| 516 | } |
| 517 | if strings.TrimSpace(g.goal) == "" || g.status != GoalStatusRunning { |
| 518 | return goalAdvanceResult{cont: false} |
| 519 | } |
| 520 | g.continuationEpoch++ |
| 521 | // A top-level goal turn (the first turn or a synthetic continuation) counts |
| 522 | // against the turn budget; the in-Run model/tool loop is never re-counted. |
| 523 | g.turnsUsed++ |
| 524 | // No-progress bookkeeping: only host-verifiable changes reset the stall |
| 525 | // counter. A terminal report is itself host-verifiable progress. |
| 526 | progressed := in.progressAfter != in.progressBefore || in.progressBefore == "" |
| 527 | if in.report != nil && in.report.status != GoalStatusRunning { |
| 528 | progressed = true |
| 529 | } |
| 530 | if progressed { |
| 531 | g.noProgressTurns = 0 |
| 532 | } else { |
| 533 | g.noProgressTurns++ |
| 534 | } |
| 535 | var notice string |
| 536 | var intercept string |
| 537 | var interceptNotice string |
| 538 | evaluatorComplete := in.evaluator != nil && in.evaluator.outcome == goaleval.OutcomeComplete |
| 539 | evaluatorBlocked := in.evaluator != nil && in.evaluator.outcome == goaleval.OutcomeBlocked |
| 540 | // Terminal dispositions first (completing or blocking ends the goal, so the |
| 541 | // budget gates never veto them); then evaluator fail-closed, then the |
| 542 | // budget gates, then the no-progress gate; only then continue. |
| 543 | reportBlocked := in.report != nil && in.report.status == GoalStatusBlocked |
| 544 | reportComplete := in.report != nil && in.report.status == GoalStatusComplete |
| 545 | completeOK := (reportComplete || evaluatorComplete) && formatIncompleteTodos(in.todos, in.readiness.Reason) == "" |
| 546 | switch { |
| 547 | case reportBlocked: |
| 548 | // A single blocked report ends the goal immediately; the host no longer |
| 549 | // repeats a three-turn confirmation ritual. |
| 550 | reason := cleanGoalBlockReason(in.report.reason) |
| 551 | if reason == "" { |
| 552 | reason = "blocked" |
| 553 | } |
| 554 | g.status = GoalStatusBlocked |
| 555 | g.block = reason |
| 556 | g.stopCause = "" |
| 557 | g.lastContinuationReason = clipGoalReason(in.report.reason) |
| 558 | notice = "goal blocked: " + reason |
| 559 | case evaluatorBlocked: |
| 560 | reason := cleanGoalBlockReason(in.evaluator.reason) |
| 561 | if reason == "" { |
| 562 | reason = "blocked" |
| 563 | } |
| 564 | g.status = GoalStatusBlocked |
| 565 | g.block = reason |
| 566 | g.stopCause = "" |
| 567 | g.lastEvaluatorReason = clipGoalReason(in.evaluator.reason) |
| 568 | notice = "goal blocked: " + reason |
| 569 | case completeOK: |
| 570 | g.goal = "" |
| 571 | g.status = GoalStatusComplete |
| 572 | g.block = "" |
| 573 | g.stopCause = "" |
| 574 | g.lastContinuationReason, g.lastEvaluatorReason = "", "" |
| 575 | notice = goalCompleteNotice |
| 576 | case in.evaluatorFailed != "" || (in.evaluator != nil && in.evaluator.outcome == goaleval.OutcomeUncertain): |
| 577 | // Fail closed: an unavailable, erroring, or uncertain evaluator pauses |
| 578 | // the goal instead of defaulting to continue. |
| 579 | reason := "the completion evaluator is unavailable or could not judge the turn" |
| 580 | if in.evaluatorFailed != "" { |
| 581 | reason = "the completion evaluator failed: " + in.evaluatorFailed |
| 582 | } |
| 583 | g.status = GoalStatusBlocked |
| 584 | g.stopCause = stopCauseEvaluator |
| 585 | g.block = clipGoalReason(reason) |
| 586 | g.lastEvaluatorReason = clipGoalReason(reason) |
| 587 | notice = "goal paused: " + reason |
| 588 | case g.turnsLimit > 0 && g.turnsUsed >= g.turnsLimit: |
| 589 | reason := fmt.Sprintf("turn budget exhausted (%d/%d turns used)", g.turnsUsed, g.turnsLimit) |
| 590 | g.status = GoalStatusBlocked |
| 591 | g.stopCause = stopCauseBudgetTurns |
| 592 | g.block = clipGoalReason(reason) |
| 593 | notice = "goal paused: " + reason |
| 594 | case g.noProgressLimit > 0 && g.noProgressTurns >= g.noProgressLimit: |
| 595 | reason := fmt.Sprintf("no host-verifiable progress in the last %d turns", g.noProgressTurns) |
| 596 | g.status = GoalStatusBlocked |
| 597 | g.stopCause = stopCauseNoProgress |
| 598 | g.block = clipGoalReason(reason) |
| 599 | notice = "goal paused: " + reason |
| 600 | default: |
| 601 | // Continue. A complete claim rejected by readiness, or a turn with no |
| 602 | // report but an explicit missing list, carries the missing requirements |
| 603 | // into the next turn; a continue report carries its next_action. |
| 604 | switch { |
| 605 | case reportComplete: |
| 606 | intercept = formatIncompleteTodos(in.todos, in.readiness.Reason) |
| 607 | interceptNotice = "Goal is not ready to complete yet; continuing the remaining work." |
| 608 | g.lastContinuationReason = clipGoalReason("readiness missing: " + in.readiness.Reason) |
| 609 | case in.report != nil && in.report.status == GoalStatusRunning: |
| 610 | g.lastContinuationReason = clipGoalReason(in.report.reason) |
| 611 | if in.report.nextAction != "" { |
| 612 | intercept = in.report.nextAction |
| 613 | } |
| 614 | case len(in.readiness.Missing) > 0: |
| 615 | intercept = formatIncompleteTodos(in.todos, in.readiness.Reason) |
| 616 | interceptNotice = "Goal is not ready to complete yet; continuing the remaining work." |
| 617 | g.lastContinuationReason = clipGoalReason("readiness missing: " + in.readiness.Reason) |
| 618 | case evaluatorComplete: |
| 619 | intercept = formatIncompleteTodos(in.todos, in.readiness.Reason) |
| 620 | interceptNotice = "Goal is not ready to complete yet; continuing the remaining work." |
| 621 | g.lastEvaluatorReason = clipGoalReason(in.evaluator.reason) |
| 622 | g.lastContinuationReason = clipGoalReason("readiness missing: " + in.readiness.Reason) |
| 623 | case in.evaluator != nil && in.evaluator.outcome == goaleval.OutcomeContinue: |
| 624 | g.lastEvaluatorReason = clipGoalReason(in.evaluator.reason) |
| 625 | } |
| 626 | } |
| 627 | res := goalAdvanceResult{ |
| 628 | notice: notice, |
| 629 | intercept: intercept, |
| 630 | interceptNotice: interceptNotice, |
| 631 | cont: notice == "", |
| 632 | continuationEpoch: g.continuationEpoch, |
| 633 | } |
| 634 | if notice != "" { |
| 635 | res.path, res.data, res.ok = g.buildStateLocked(in.todos) |
| 636 | } |
| 637 | return res |
| 638 | } |
| 639 | |
| 640 | // foldUsage attributes a turn's billable tokens to the goal, but only while the |
| 641 | // goal lifecycle still matches the recorder's scope+epoch; stale or replaced |
| 642 | // goals reject late usage. |
| 643 | func (g *goalMachine) foldUsage(scopeID string, epoch uint64, tokens int) bool { |
| 644 | g.mu.Lock() |
| 645 | defer g.mu.Unlock() |
| 646 | if tokens <= 0 || g.scopeID != scopeID || g.continuationEpoch != epoch { |
| 647 | return false |
| 648 | } |
| 649 | g.tokensUsed += tokens |
| 650 | return true |
| 651 | } |
| 652 | |
| 653 | // buildStateLocked marshals the current goal state for persistence. The caller |
| 654 | // holds mu; this only reads in-memory state, never touching disk. Returns ok=false |
| 655 | // when persistence is disabled (no state path). The matching writeState does the |
| 656 | // disk write OFF mu so the per-turn save can't stall a status poll. |
| 657 | func (g *goalMachine) buildStateLocked(todos []evidence.TodoItem) (path string, data []byte, ok bool) { |
| 658 | if g.statePath == "" { |
| 659 | return "", nil, false |
| 660 | } |
| 661 | state := goalState{ |
| 662 | Goal: g.goal, |
| 663 | Status: g.status, |
| 664 | ResearchMode: g.researchMode, |
| 665 | AutoResearchTaskID: g.autoResearchTaskID, |
| 666 | ScopeID: g.scopeID, |
| 667 | DeliveryCheckpoint: g.deliveryCheckpoint, |
| 668 | Turns: g.turnsUsed, |
| 669 | Block: g.block, |
| 670 | Strict: g.strict, |
| 671 | Todos: todos, |
| 672 | BudgetClass: g.budgetClass, |
| 673 | TurnsUsed: g.turnsUsed, |
| 674 | TurnsLimit: g.turnsLimit, |
| 675 | TokensUsed: g.tokensUsed, |
| 676 | TokensLimit: g.tokensLimit, |
| 677 | NoProgressTurns: g.noProgressTurns, |
| 678 | NoProgressLimit: g.noProgressLimit, |
| 679 | LastContinuationReason: g.lastContinuationReason, |
| 680 | LastEvaluatorReason: g.lastEvaluatorReason, |
| 681 | StopCause: g.stopCause, |
| 682 | BudgetExtensions: g.budgetExtensions, |
| 683 | } |
| 684 | b, err := json.Marshal(state) |
| 685 | if err != nil { |
| 686 | slog.Warn("controller: marshal goal state", "err", err) |
| 687 | return "", nil, false |
| 688 | } |
| 689 | return g.statePath, b, true |
| 690 | } |
| 691 | |
| 692 | // writeStateErr persists pre-marshaled goal-state bytes to disk, OFF mu and |
| 693 | // serialized by writeMu so concurrent saves don't interleave or land out of |
| 694 | // order. Atomic replacement keeps the prior state intact when a write fails. |
| 695 | func (g *goalMachine) writeStateErr(path string, data []byte) error { |
| 696 | if path == "" || data == nil { |
| 697 | return nil |
| 698 | } |
| 699 | g.writeMu.Lock() |
| 700 | defer g.writeMu.Unlock() |
| 701 | if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil { |
| 702 | return err |
| 703 | } |
| 704 | return fileutil.AtomicWriteFile(path, data, 0o644) |
| 705 | } |
| 706 | |
| 707 | // writeState preserves the existing best-effort behavior for background Goal |
| 708 | // progress. Callers that need transactional persistence use writeStateErr. |
| 709 | func (g *goalMachine) writeState(path string, data []byte) { |
| 710 | if err := g.writeStateErr(path, data); err != nil { |
| 711 | slog.Warn("controller: write goal state", "err", err) |
| 712 | } |
| 713 | } |
| 714 | |
| 715 | // persistWithTodos re-persists goal state with the given todos, without |
| 716 | // changing any in-memory goal fields. Used after force-completing todos on |
| 717 | // goal completion so a session reload does not revert to the old incomplete |
| 718 | // todo state. |
| 719 | func (g *goalMachine) persistWithTodos(todos []evidence.TodoItem) { |
| 720 | g.mu.Lock() |
| 721 | path, data, ok := g.buildStateLocked(todos) |
| 722 | g.mu.Unlock() |
| 723 | if ok { |
| 724 | g.writeState(path, data) |
| 725 | } |
| 726 | } |
| 727 | |
| 728 | // terminalTodosFromState reads the persisted goal-state sidecar and returns its |
| 729 | // todo snapshot only after the goal has reached a terminal state. Running goal |
| 730 | // state is not refreshed on every todo_write, so its todos may be older than the |
| 731 | // transcript rebuilt by Agent.SetSession. |
| 732 | func (g *goalMachine) terminalTodosFromState(sessionPath string) ([]evidence.TodoItem, bool) { |
| 733 | if strings.TrimSpace(sessionPath) == "" { |
| 734 | return nil, false |
| 735 | } |
| 736 | data, err := fileencoding.ReadFileUTF8(goalStatePath(sessionPath)) |
| 737 | if err != nil { |
| 738 | if !os.IsNotExist(err) { |
| 739 | slog.Warn("controller: read goal state", "err", err) |
| 740 | } |
| 741 | return nil, false |
| 742 | } |
| 743 | var state goalState |
| 744 | if err := json.Unmarshal(data, &state); err != nil { |
| 745 | slog.Warn("controller: parse goal state", "err", err) |
| 746 | return nil, false |
| 747 | } |
| 748 | switch state.Status { |
| 749 | case GoalStatusComplete, GoalStatusBlocked, GoalStatusStopped: |
| 750 | default: |
| 751 | return nil, false |
| 752 | } |
| 753 | if len(state.Todos) == 0 { |
| 754 | return nil, false |
| 755 | } |
| 756 | return append([]evidence.TodoItem(nil), state.Todos...), true |
| 757 | } |
| 758 | |
| 759 | // restoreFromState reloads Goal state from the persisted sidecar during resume. |
| 760 | // The sidecar is authoritative when present: a stale tab profile must not turn |
| 761 | // a blocked or stopped Goal back into a running one during a controller rebuild. |
| 762 | // Recoverable terminal states retain their scope for an explicit ResumeGoal. |
| 763 | // Old sidecars missing the budget fields get their defaults re-derived: the |
| 764 | // existing Turns count carries into the new turn budget, tokens start from 0, |
| 765 | // and the budget class is recomputed from the goal text. |
| 766 | // |
| 767 | // When a legacy budget_tokens pause is cleared, migrated is true and path/data |
| 768 | // carry the rewritten state for immediate atomic persistence (no provider call). |
| 769 | func (g *goalMachine) restoreFromState(sessionPath string) (path string, data []byte, migrated bool) { |
| 770 | if strings.TrimSpace(sessionPath) == "" { |
| 771 | return "", nil, false |
| 772 | } |
| 773 | // Ensure write path is bound even when the controller rebuilds. |
| 774 | if g.statePath == "" { |
| 775 | g.setStatePath(goalStatePath(sessionPath)) |
| 776 | } |
| 777 | raw, err := fileencoding.ReadFileUTF8(goalStatePath(sessionPath)) |
| 778 | if err != nil { |
| 779 | if !os.IsNotExist(err) { |
| 780 | slog.Warn("controller: read goal state", "err", err) |
| 781 | } |
| 782 | return "", nil, false |
| 783 | } |
| 784 | var state goalState |
| 785 | if err := json.Unmarshal(raw, &state); err != nil { |
| 786 | slog.Warn("controller: parse goal state", "err", err) |
| 787 | return "", nil, false |
| 788 | } |
| 789 | g.mu.Lock() |
| 790 | defer g.mu.Unlock() |
| 791 | g.goal = strings.TrimSpace(state.Goal) |
| 792 | g.status = state.Status |
| 793 | if g.status == "" { |
| 794 | g.status = GoalStatusStopped |
| 795 | } |
| 796 | g.researchMode = state.ResearchMode |
| 797 | g.autoResearchTaskID = strings.TrimSpace(state.AutoResearchTaskID) |
| 798 | g.scopeID = strings.TrimSpace(state.ScopeID) |
| 799 | if g.goal != "" && g.scopeID == "" { |
| 800 | g.scopeID = newGoalScopeID() |
| 801 | } |
| 802 | g.deliveryCheckpoint = state.DeliveryCheckpoint |
| 803 | if g.scopeID == "" { |
| 804 | g.deliveryCheckpoint = evidence.DeliveryCheckpoint{} |
| 805 | } else if g.deliveryCheckpoint.ScopeID == "" { |
| 806 | g.deliveryCheckpoint.ScopeID = g.scopeID |
| 807 | } else if g.deliveryCheckpoint.ScopeID != g.scopeID { |
| 808 | g.deliveryCheckpoint = evidence.DeliveryCheckpoint{ScopeID: g.scopeID} |
| 809 | } |
| 810 | g.block = state.Block |
| 811 | g.strict = state.Strict |
| 812 | g.stopCause = state.StopCause |
| 813 | g.budgetExtensions = state.BudgetExtensions |
| 814 | g.lastContinuationReason = state.LastContinuationReason |
| 815 | g.lastEvaluatorReason = state.LastEvaluatorReason |
| 816 | // Budget defaults: old sidecars carry Turns (pre-budget counting); treat it |
| 817 | // as the new turn usage and re-derive the class/limits from the goal text. |
| 818 | g.turnsUsed = state.TurnsUsed |
| 819 | if g.turnsUsed == 0 && state.Turns > 0 { |
| 820 | g.turnsUsed = state.Turns |
| 821 | } |
| 822 | g.tokensUsed = state.TokensUsed |
| 823 | // Token hard limits are gone: keep the field at 0. Old non-zero sidecar |
| 824 | // values are read and ignored so downgrade/upgrade never loses other state. |
| 825 | g.tokensLimit = 0 |
| 826 | migrated = false |
| 827 | if g.goal != "" { |
| 828 | g.budgetClass = state.BudgetClass |
| 829 | if g.budgetClass == "" { |
| 830 | g.budgetClass = budgetClassFor(g.goal, g.researchMode) |
| 831 | } |
| 832 | if state.TurnsLimit > 0 { |
| 833 | g.turnsLimit = state.TurnsLimit |
| 834 | } else { |
| 835 | g.turnsLimit = budgetQuota(g.budgetClass) |
| 836 | } |
| 837 | if state.NoProgressLimit > 0 { |
| 838 | g.noProgressLimit = state.NoProgressLimit |
| 839 | } else { |
| 840 | g.noProgressLimit = defaultNoProgressLimit |
| 841 | } |
| 842 | g.noProgressTurns = state.NoProgressTurns |
| 843 | // Auto-clear legacy token-budget pauses so the next user turn can |
| 844 | // continue without a manual resume. Loading itself never calls a |
| 845 | // provider. |
| 846 | if g.status == GoalStatusBlocked && g.stopCause == stopCauseBudgetTokens { |
| 847 | g.status = GoalStatusRunning |
| 848 | g.stopCause = "" |
| 849 | // The stop cause proves the block belongs to the removed token gate; |
| 850 | // do not leave a stale blocked reason attached to a running goal. |
| 851 | g.block = "" |
| 852 | migrated = true |
| 853 | } |
| 854 | // Also rewrite sidecars that still store a non-zero tokensLimit so the |
| 855 | // next load does not re-surface the deprecated hard ceiling in status. |
| 856 | if state.TokensLimit != 0 { |
| 857 | migrated = true |
| 858 | } |
| 859 | } |
| 860 | g.continuationEpoch++ |
| 861 | if migrated { |
| 862 | // Migration rewrites only the removed budget state. Preserve the todo |
| 863 | // snapshot carried by the authoritative sidecar instead of clearing it. |
| 864 | path, data, ok := g.buildStateLocked(state.Todos) |
| 865 | if ok { |
| 866 | return path, data, true |
| 867 | } |
| 868 | } |
| 869 | return "", nil, false |
| 870 | } |
| 871 | |
| 872 | // formatIncompleteTodos renders the reminder shown when a complete claim |
| 873 | // arrives while the executor's canonical todos or project-readiness checks |
| 874 | // aren't done. Returns empty when nothing is blocking. Pure: the caller gathers |
| 875 | // todos and the readiness reason from the executor off the goal lock. |
| 876 | func formatIncompleteTodos(todos []evidence.TodoItem, readiness string) string { |
| 877 | var parts []string |
| 878 | if len(todos) > 0 { |
| 879 | if incomplete := evidence.IncompleteTodos(todos); len(incomplete) > 0 { |
| 880 | var b strings.Builder |
| 881 | b.WriteString("the following tasks are still incomplete:") |
| 882 | for _, t := range incomplete { |
| 883 | fmt.Fprintf(&b, "\n - %s (%s)", t.Content, t.Status) |
| 884 | } |
| 885 | parts = append(parts, b.String()) |
| 886 | } |
| 887 | } |
| 888 | if readiness != "" { |
| 889 | parts = append(parts, readiness) |
| 890 | } |
| 891 | if len(parts) == 0 { |
| 892 | return "" |
| 893 | } |
| 894 | var b strings.Builder |
| 895 | b.WriteString("Goal signaled complete but issues remain:\n") |
| 896 | for _, p := range parts { |
| 897 | b.WriteString("- ") |
| 898 | b.WriteString(p) |
| 899 | b.WriteString("\n") |
| 900 | } |
| 901 | b.WriteString("Fix or use todo_write/complete_step to mark done, then report complete again via update_goal.") |
| 902 | return b.String() |
| 903 | } |
| 904 | |
| 905 | // clipGoalReason bounds a recorded reason for storage and display. |
| 906 | func clipGoalReason(reason string) string { |
| 907 | reason = strings.TrimSpace(reason) |
| 908 | const max = 400 |
| 909 | if r := []rune(reason); len(r) > max { |
| 910 | return string(r[:max]) + "..." |
| 911 | } |
| 912 | return reason |
| 913 | } |
| 914 | |
| 915 | func cleanGoalBlockReason(reason string) string { |
| 916 | return strings.Trim(strings.TrimSpace(reason), " \t\r\n:：,，.。;；!！?？-—_[]()（）") |
| 917 | } |
| 918 | |
| 919 | // ShortGoalForNotice collapses whitespace and truncates a goal for one-line UI. |
| 920 | func ShortGoalForNotice(goal string) string { |
| 921 | goal = strings.Join(strings.Fields(goal), " ") |
| 922 | runes := []rune(goal) |
| 923 | const max = 160 |
| 924 | if len(runes) <= max { |
| 925 | return goal |
| 926 | } |
| 927 | return string(runes[:max]) + "..." |
| 928 | } |
| 929 | |
| 930 | // goalTodos snapshots the executor's canonical todos for goal-state persistence. |
| 931 | func (c *Controller) goalTodos() []evidence.TodoItem { |
| 932 | if c.executor == nil { |
| 933 | return nil |
| 934 | } |
| 935 | return c.executor.CanonicalTodoState() |
| 936 | } |
| 937 | |
| 938 | // persistGoalState writes a freshly built goal state to disk, off c.mu. The |
| 939 | // executor guard preserves the original behavior of skipping persistence when |
| 940 | // no executor is attached. |
| 941 | func (c *Controller) persistGoalState(path string, data []byte, ok bool) { |
| 942 | if !ok || c.executor == nil { |
| 943 | return |
| 944 | } |
| 945 | c.goals.writeState(path, data) |
| 946 | } |
| 947 | |
| 948 | func (c *Controller) restoreTerminalGoalTodos(sessionPath string) { |
| 949 | if c.executor == nil { |
| 950 | return |
| 951 | } |
| 952 | todos, ok := c.goals.terminalTodosFromState(sessionPath) |
| 953 | if !ok { |
| 954 | return |
| 955 | } |
| 956 | c.executor.ReplaceTodoState(todos) |
| 957 | } |
| 958 | |
| 959 | // GoalRuntimeView is the host-side runtime summary exposed to frontends. |
| 960 | type GoalRuntimeView struct { |
| 961 | TurnsUsed int |
| 962 | TurnsLimit int |
| 963 | TokensUsed int |
| 964 | TokensLimit int |
| 965 | NoProgressTurns int |
| 966 | NoProgressLimit int |
| 967 | LastReason string |
| 968 | StopCause string |
| 969 | BudgetExtensions int |
| 970 | } |
| 971 | |
| 972 | // runtimeView returns the goal's budget/runtime summary. |
| 973 | func (g *goalMachine) runtimeView() GoalRuntimeView { |
| 974 | g.mu.Lock() |
| 975 | defer g.mu.Unlock() |
| 976 | last := g.lastEvaluatorReason |
| 977 | if last == "" { |
| 978 | last = g.lastContinuationReason |
| 979 | } |
| 980 | return GoalRuntimeView{ |
| 981 | TurnsUsed: g.turnsUsed, |
| 982 | TurnsLimit: g.turnsLimit, |
| 983 | TokensUsed: g.tokensUsed, |
| 984 | TokensLimit: 0, // deprecated: no hard token limit |
| 985 | NoProgressTurns: g.noProgressTurns, |
| 986 | NoProgressLimit: g.noProgressLimit, |
| 987 | LastReason: last, |
| 988 | StopCause: g.stopCause, |
| 989 | BudgetExtensions: g.budgetExtensions, |
| 990 | } |
| 991 | } |
| 992 | |
| 993 | func (g *goalMachine) lastContinuationReasonText() string { |
| 994 | g.mu.Lock() |
| 995 | defer g.mu.Unlock() |
| 996 | if g.lastEvaluatorReason != "" { |
| 997 | return g.lastEvaluatorReason |
| 998 | } |
| 999 | return g.lastContinuationReason |
| 1000 | } |
| 1001 | |
| 1002 | func (g *goalMachine) budgetStatusText() string { |
| 1003 | g.mu.Lock() |
| 1004 | defer g.mu.Unlock() |
| 1005 | return fmt.Sprintf("turns: %d/%d used, tokens: %d, no-progress turns: %d/%d", |
| 1006 | g.turnsUsed, g.turnsLimit, g.tokensUsed, g.noProgressTurns, g.noProgressLimit) |
| 1007 | } |
| 1008 | |
| 1009 | // goalTurnRecorder is the per-turn recorder bound to one goal turn's scope and |
| 1010 | // epoch. update_goal calls land here as candidate state; the FSM commits them |
| 1011 | // only when the goal lifecycle still matches (scope + epoch), so late calls |
| 1012 | // from a replaced or cleared goal are rejected. Usage events emitted during the |
| 1013 | // turn are folded through the recorder into the goal's observational token total. |
| 1014 | type goalTurnRecorder struct { |
| 1015 | mu sync.Mutex |
| 1016 | machine *goalMachine |
| 1017 | scopeID string |
| 1018 | epoch uint64 |
| 1019 | recorded bool |
| 1020 | terminal bool |
| 1021 | status string |
| 1022 | reason string |
| 1023 | nextAction string |
| 1024 | tokensUsed int |
| 1025 | progressBefore string |
| 1026 | } |
| 1027 | |
| 1028 | func (g *goalMachine) newTurnRecorder(scopeID string, epoch uint64) *goalTurnRecorder { |
| 1029 | return &goalTurnRecorder{machine: g, scopeID: scopeID, epoch: epoch} |
| 1030 | } |
| 1031 | |
| 1032 | // RecordGoalReport validates the report against the turn's goal lifecycle and |
| 1033 | // records it as the turn's candidate disposition. Same-value repeats are |
| 1034 | // idempotent; continue may upgrade to complete/blocked; complete and blocked |
| 1035 | // are terminal and reject conflicting later calls. |
| 1036 | func (r *goalTurnRecorder) RecordGoalReport(report tool.GoalReport) (string, error) { |
| 1037 | r.mu.Lock() |
| 1038 | defer r.mu.Unlock() |
| 1039 | if !r.machine.turnActive(r.scopeID, r.epoch) { |
| 1040 | return "", fmt.Errorf("update_goal: the active goal changed during this turn — report ignored; no goal state was changed") |
| 1041 | } |
| 1042 | switch { |
| 1043 | case r.terminal: |
| 1044 | return "", fmt.Errorf("update_goal: this turn's disposition is already final (%s); conflicting %q report ignored", r.status, report.Status) |
| 1045 | case !r.recorded: |
| 1046 | // first record |
| 1047 | case r.status == report.Status && r.reason == report.Reason && r.nextAction == report.NextAction: |
| 1048 | return fmt.Sprintf("update_goal: %s already recorded for this turn (identical report).", report.Status), nil |
| 1049 | case r.status == GoalStatusRunning && (report.Status == GoalStatusComplete || report.Status == GoalStatusBlocked): |
| 1050 | // continue → terminal upgrade allowed. |
| 1051 | default: |
| 1052 | return "", fmt.Errorf("update_goal: conflicting reports this turn (%s then %s) — the later report was ignored", r.status, report.Status) |
| 1053 | } |
| 1054 | r.recorded = true |
| 1055 | r.status = report.Status |
| 1056 | r.reason = report.Reason |
| 1057 | r.nextAction = report.NextAction |
| 1058 | if report.Status != GoalStatusRunning { |
| 1059 | r.terminal = true |
| 1060 | } |
| 1061 | return fmt.Sprintf("update_goal: %s recorded for this turn.", report.Status), nil |
| 1062 | } |
| 1063 | |
| 1064 | func (r *goalTurnRecorder) setProgressBefore(sig string) { |
| 1065 | r.mu.Lock() |
| 1066 | r.progressBefore = sig |
| 1067 | r.mu.Unlock() |
| 1068 | } |
| 1069 | |
| 1070 | func (r *goalTurnRecorder) progressBeforeText() string { |
| 1071 | r.mu.Lock() |
| 1072 | defer r.mu.Unlock() |
| 1073 | return r.progressBefore |
| 1074 | } |
| 1075 | |
| 1076 | func (r *goalTurnRecorder) addUsage(tokens int) { |
| 1077 | if tokens <= 0 { |
| 1078 | return |
| 1079 | } |
| 1080 | r.mu.Lock() |
| 1081 | if r.machine.foldUsage(r.scopeID, r.epoch, tokens) { |
| 1082 | r.tokensUsed += tokens |
| 1083 | } |
| 1084 | r.mu.Unlock() |
| 1085 | } |
| 1086 | |
| 1087 | func (r *goalTurnRecorder) usageTokens() int { |
| 1088 | r.mu.Lock() |
| 1089 | defer r.mu.Unlock() |
| 1090 | return r.tokensUsed |
| 1091 | } |
| 1092 | |
| 1093 | // validReport returns the recorded report only when the goal lifecycle still |
| 1094 | // matches the recorder's binding; stale (replaced/cleared) turns report nothing. |
| 1095 | func (r *goalTurnRecorder) validReport(expectedEpoch uint64) *goalTurnReport { |
| 1096 | r.mu.Lock() |
| 1097 | defer r.mu.Unlock() |
| 1098 | if !r.recorded || r.epoch != expectedEpoch || !r.machine.turnActive(r.scopeID, r.epoch) { |
| 1099 | return nil |
| 1100 | } |
| 1101 | return &goalTurnReport{status: r.status, reason: r.reason, nextAction: r.nextAction} |
| 1102 | } |
| 1103 | |
| 1104 | // goalTurnReport is the validated update_goal report for one goal turn. |
| 1105 | type goalTurnReport struct { |
| 1106 | status string |
| 1107 | reason string |
| 1108 | nextAction string |
| 1109 | } |
| 1110 | |
| 1111 | // turnActive reports whether the machine's goal lifecycle matches the binding. |
| 1112 | func (g *goalMachine) turnActive(scopeID string, epoch uint64) bool { |
| 1113 | g.mu.Lock() |
| 1114 | defer g.mu.Unlock() |
| 1115 | return strings.TrimSpace(g.goal) != "" && g.status == GoalStatusRunning && |
| 1116 | g.scopeID == scopeID && g.continuationEpoch == epoch |
| 1117 | } |
| 1118 |