chore(memory): model-router wave 0 — plan, LRN-203/204, BLK-029, journal

2026-10-08 15:25:11 +02:00
parent f24682b3f3
commit b73d1b127e
5 changed files with 188 additions and 0 deletions
@@ -48,6 +48,7 @@ rules:
 | BLK-026 | 2026-10-06 | `make test` red on macOS: 13 suites, GNU-only idioms in suite + 7 libs (SIGPIPE under pipefail, `sed -i`, `wc` padding, `stat -c`, `realpath -m`, bare `timeout`, `grep -oP`) | resolved |
 | BLK-027 | 2026-10-06 | this machine never ran `make link`/`make plugin`: no global `core.hooksPath` → post-commit push never fired, branches landed ahead of upstream; 11 vendored skills + `~/.claude/.env` missing | resolved (link) / open (plugin) |
 | BLK-028 | 2026-10-06 | notify-attention on a VS Code client: bell + toast silent-degradation faults (merge of BLK-019 + BLK-020): terminalBell sound default off, ext hooks only terminals born after activation, Code muted in Windows mixer | resolved |
+| BLK-029 | 2026-10-08 | Claude Code mods (2.1.294): alias → id resolution for a model set by a hook lags the Agent tool's (`sonnet` → `claude-sonnet-5`, 404); Agent tool schema refuses full ids | upstream |

 ---

@@ -305,3 +306,11 @@ rules:
 - **Probe order (do FIRST, before server archaeology)**: fresh VS Code terminal, `printf '\a\a\033]777;notify;Test;hello\033\\'` → splits terminal path from client renderer; palette `Help: List Signal Sounds` → Terminal Bell preview bypasses terminal/BEL/hook/dtach/ext, isolates renderer audio in one step.
 - **Status**: resolved (BLK-019 2026-09-01, BLK-020 A+B 2026-09-02/03). Sources superseded by this entry; bodies kept for history.
 - **Reference**: `~/.claude/hooks/notify-attention.sh` header documents the setting; [[LRN-145]] terminalSequence-not-/dev/tty; silent-degradation class [[LRN-047]]; sources [[BLK-019]], [[BLK-020]].
+
+## BLK-029 — Mods: model alias set by a hook resolves to a stale id (`sonnet` → `claude-sonnet-5`, 404) — 2026-10-08
+- **Friction**: model-router spike. Sub-agent routed by `agent.spawn` or `tool.call Agent` rewrite with alias `sonnet`/`haiku` → "model_not_found HTTP 404, model sent to the API: claude-sonnet-5". Same alias typed by the model in the Agent tool param → `claude-sonnet-5-5`, OK.
+- **Real cause**: two alias tables in the CLI (2.1.294): the Agent tool's is current, the function-hooks path's is stale. Not an access issue (`/model` lists all four tiers; explicit haiku/sonnet dispatches answered).
+- **Solution**: hooks write full ids (`claude-sonnet-5-5`, `claude-haiku-4-5-20251001`, `claude-opus-5-5`, `claude-fable-5-1`) from the mod's own table; Agent tool param rewrite stays alias-only (schema enum) → the mod sets the model at `agent.spawn`, not at the param.
+- **Status**: upstream (report to anthropics/claude-code with the request id `req_011Cfppp7VFt8z2Pd3zJrpUi`); workaround in model-router.
+- **Reference**: [[LRN-203]], plan `.claude/tasks/plans/2026-10-08-model-router-mod.md`.
+
@@ -577,3 +577,5 @@ rules:
 - Merge (user go "tu peux merge dans develop"): final full suite green (46 suites minus the declared env red) + Health Stack shellcheck clean on the branch tip → `gitflow finish feature manual-push-mode` → develop 669db06, pushed, branch removed local + origin. 19 commits (runs A, B, C1/C2, D1/D2/D3 + docs + memory). User answered: only pushes change; commits/branches/local merges untouched; invalid value now fail-closed everywhere. User plan: dotfiles installer prompts for `gitflow.autopush` (default false) — told them the gitconfig template also needs `core.hooksPath` (the install wiped it). Open: user probe `! git push --dry-run` under autopush=false; AC6 env red (design-tool-gate); post-run-D residuals in TODO.
 - User tested manual-push mode on their machine: works (no auto push under `false`, bang-prefixed dry-run passes). Prompt handed over for the dotfiles repo: gitconfig template gets `core.hooksPath = ~/.claude/githooks` + `[gitflow] autopush = @AUTOPUSH@` rendered from an install question (default false, true/false only, unrendered placeholder = render failure). BDR-112 amended.

+## 2026-10-08
+- model-router mod, wave 0 spike (user ask: one mod routes model + effort per request, replaces effort-* shifters + pins). Plan `.claude/tasks/plans/2026-10-08-model-router-mod.md`, 4 decisions by AskUserQuestion (spike-first main-loop switch, `CLAUDE_CODE_PLUGIN_DIRS` load, migration wave 2, names model-router / route / /route), rule "pin = entry default, sub-tasks route finer". Spike in dev-mods, hot reload on: `turn.step` effort rewrite proven (transcript `effort` field is the oracle, not `CLAUDE_EFFORT`); sub-agent model at `agent.spawn` + effort per step by agentId proven; main-loop fable → sonnet-5-5/low for 3 steps then back: works, one cold-cache step per switch INTO a model, return free. Found: hook-side alias resolver stale (`sonnet` → `claude-sonnet-5`, 404; Agent tool enum resolves the same alias to 5-5) → mod writes full ids only. feature/model-router-mod open, nothing committed yet (plan + TODO + journal pending).
@@ -1748,3 +1748,12 @@ Rule: when editing a doctrine file under structure locks, grep the test's lock s
 ## LRN-202 — Reading a stderr-then-stdout verb from a hook: `out=$(cmd 2>&1)`, last line = word, prefix line = reason; no temp file; lib path absolute before any cd
 - **Context**: unpushed-guard plan used `2>"${TMPDIR:-/tmp}/x.$$"` + cat + rm: fail-OPEN when TMPDIR is full (`|| mode=auto`), predictable path, symlink-followable on shared /tmp, leaked on kill; `mode=$(cmd 2>&1 >/dev/null)` captures ONLY stderr. push-guard's `mktemp` variant added `set -u` trap hazards. The verb writes its stderr line BEFORE its stdout word in one process, so `${out##*$'\n'}` is the word and the `gitflow.sh push-mode:` line is the reason (select by prefix, not `head -1`: a bash startup warning could precede it). Resolve the lib path to an absolute one BEFORE the hook's `cd "$cwd"` (a relative invocation otherwise resolves into the target repo).
 - **Apply**: hooks never touch temp files for a one-line capture; anything but the expected word is treated as the fail-closed state, never as the default. Links [[BDR-114]], [[LRN-196]], [[LRN-199]].
+
+## LRN-203 — Mod hooks: model ALIAS set by a hook resolves through stale table (`sonnet` → `claude-sonnet-5`, 404); write full ids; oracle = transcript fields, not `CLAUDE_EFFORT`
+- **Context**: model-router spike 2026-10-08, CLI 2.1.294. `agent.spawn` or `tool.call Agent` param rewrite with alias `sonnet` → API got `claude-sonnet-5`, HTTP 404 model_not_found. Same alias passed by the model in the Agent tool param → `claude-sonnet-5-5`, fine. Full id `claude-sonnet-5-5` from hook → fine, every step answered by 5-5. Agent tool schema enum refuses full ids, so full ids reach API only via hooks. Effort rewrite at `turn.step` proven by transcript record field `effort` (high → medium); `CLAUDE_EFFORT` env + `perTurnEffort` stay at turn setting, blind to per-request rewrite.
+- **Apply**: any mod that sets a model carries its own alias → full-id table (one place to bump per tier release). Verify routing with transcript `message.model` + record `effort`, never env vars. Links [[BDR-108]] (aliases as pins: still right at the Agent-tool call site, wrong inside hooks), [[BLK-029]].
+
+## LRN-204 — Main-loop model switch mid-turn works, costs one cold-cache step on the full context per switch INTO a model; return free (per-model cache, 1 h TTL)
+- **Context**: spike 2026-10-08, fable → `claude-sonnet-5-5`/low for 3 steps on ~260k context, then back. Conversation intact (tools, results, thinking blocks from another model in history: no error). First sonnet step cache_read 0 (full 260k billed), next steps 237k cached; return to fable step read 263k cached.
+- **Apply**: switch the main loop only for spans long enough to amortize one uncached read of the whole context (many mechanical steps), never per tool call; short mechanical work → small-context haiku sub-agent. Haiku 4.5 window 200k: a long main loop cannot go to haiku at all. Flag off by default in model-router. Links [[LRN-203]], [[BDR-107]].
+