返回 CodeWhale
2026-08-04-fleet-rebuild.md
根目录 / docs / decisions / 2026-08-04-fleet-rebuild.md
1 # Fleet rebuild decisions (2026-08-04)
2
3 Owner-steered requirements captured this session. Evidence-first: every claim
4 below is anchored to code or a user statement; nothing is aspiration.
5
6 ## 1. Fleets are saved configurations (not shadowed role files)
7
8 Today two parallel stores exist:
9 - per-role profile files (`~/.codewhale/agents/*.toml`, `.codewhale/agents/*.toml`,
10 `[fleet.profiles]`, built-ins) merged by id with `ShadowedProfile` receipts
11 (crates/tui/src/fleet/roster.rs) — the "⚠ shadow" pile the UI must not show;
12 - named fleet files (`fleets/<name>.toml` at `$CODEWHALE_HOME` or workspace)
13 consumed by the workflow crate (`crates/workflow/src/named_fleet.rs`,
14 `fleet_exact.rs`) with `QualifiedFleetId` ambiguity rejection.
15
16 **Decision:** one concept — a named Fleet file. A Fleet contains:
17 - `name` (editable display name), `description`;
18 - `operator`: provider + exact model + reasoning (absent = inherit session route);
19 - `members`: role → model pin or inherit, provider (pins only, never inferred),
20 reasoning (only when the resolved route supports it), instructions,
21 capability requirements (e.g. `vision`);
22 - scope + source: personal (`$CODEWHALE_HOME/fleets/`) or workspace
23 (`.codewhale/fleets/`), the exact file path shown in the UI.
24
25 New schema kind `schema = "fleet"` v2 owned by the TUI crate. The workflow
26 crate's `exact`/legacy parsers stay untouched; binding a v2 file to a Workflow
27 fails closed with "unknown fleet schema" until workflows are decoupled.
28
29 ## 2. Selection: personal default + explicit workspace selection
30
31 - `$CODEWHALE_HOME/fleets/selected` — user-global default Fleet (plain text name).
32 - `.codewhale/fleets/selected` — workspace selection, intentional and labeled.
33 - Resolution: workspace selection wins, then personal, then none (legacy roster).
34 - A new session starts on the selected Fleet's operator route unless CLI/env
35 override exists.
36 - Workspace config never silently shadows a personal Fleet: both remain listed,
37 scopes labeled, ambiguity surfaced (qualified `origin/name`).
38
39 ## 3. Session route changes are temporary by default
40
41 `/model` and `/provider` change only the live session route. Persistence is an
42 explicit choice: Update this Fleet / Save as a new Fleet / Save as my default
43 (no fleet selected) / Keep for session only. Never silently rewrite a saved
44 Fleet or the config file after an in-session route change. Every save receipt
45 names the exact file and scope it changed.
46
47 ## 4. Scout replaces "faster"
48
49 One visible fast exploratory role: Scout. `faster`/`model_strength`/`Fast`
50 loadout disappear from pickers, tool schemas, and Fleet config. Legacy parsing
51 survives for compatibility and migrates to the Scout policy with a receipt.
52 Unpinned Scout may receive a suggested fast companion from explicit catalog
53 metadata (provider/family → verified offering), never name-guessing; the exact
54 resolved Scout provider/model is shown before a run; explicit Scout pins win
55 and survive operator changes; no verified companion → show that clearly or
56 inherit deliberately.
57
58 ## 5. Provider → model family → exact model picker
59
60 The flat model list (model_picker.rs) becomes provider → family → exact model.
61 Each row truthfully shows provider + exact route id, readiness, explicit Fleet
62 role recommendation, input modalities (vision), tool + structured-output
63 support, reasoning levels, context/output limits. Compact by default; endpoint/
64 cost/advanced behind disclosure. No `max`/`ultra` unless genuinely supported.
65 Capability-aware routing: vision work never silently goes to a non-vision
66 member.
67
68 ## 6. Config/credential scope: user-global credentials, layered selection
69
70 Verified: credentials/keychain/OAuth are already user-global
71 (crates/secrets, ~/.codex/auth.json, ~/.grok/auth.json,
72 $CODEWHALE_HOME/credentials/*). The "authorized here, locked there" failure is
73 a config-file-selection problem: an explicit --config/CODEWHALE_CONFIG_PATH
74 pointing at a workspace file makes THAT file the loaded config, and the
75 readiness gate (`provider_is_configured_for_active`) can then fail on a
76 credential that exists user-globally. Fix: readiness consults the user-global
77 credential sources regardless of which config file is loaded; every active
78 route shows its source (provider/model, credential source, config layer,
79 temporary vs saved, precise unavailable reason).
80
81 ## 7. Members/roles pin to folders or users
82
83 Fleet save scope is either user (personal) or folder (workspace); both are
84 always listed with scope + source labels; same-name fleets in two scopes are
85 distinct and never silently shadow — the UI surfaces the ambiguity and the
86 user picks a qualified origin/name.
87
88 ## 8. Workflows will be simplified toward kimicode-swarm/grokbuild-style
89 ## orchestration (no fleet-file dependency)
90
91 Research lane open (scout dispatched). Workflow runs should need only roles +
92 session route, with a default roster from built-ins; fleet files become an
93 optional pin layer, not a requirement. Do not redesign workflows in this
94 session beyond the research; the fleet rebuild must not deepen the coupling.
95
96 ## 9. Copy is truthful
97
98 Every label, receipt, status, and confirmation states only what the code can
99 prove: save receipts name the exact file changed; readiness states carry a
100 precise reason; "saved"/"selected"/"authorized" are never claimed without the
101 underlying write/credential check having succeeded.
102
102 lines MARKDOWN