claude/.claude/tasks/plans/2026-10-09-model-router-w2-1546.md

# PLAN r4 — model-router wave 2 (migration), 2026-10-09

r1 → r2 after 3 lenses (simplicity CONCERNS(5), correctness CONCERNS(10),
robustness FATAL(8)); r2 → r3 after the confirmation pass (robustness
FATAL(8): 1 BLOCKER); r3 → r4 after a second confirmation (correctness
CONCERNS(3), wording-level). Every BLOCKER/MAJOR is closed by a NAMED change
(§ Challenge ledger). Mod = the router while on; the tracked frontmatter
(`model:` + `effort:` on agents, `effort:` on skills) STAYS as the
off-state floor, census-locked equal to the rows (r3).
Two sub-runs on `feature/model-router-w2`: W2-A (mod) then, after a user
`/reload-plugins` + live probe, W2-B (repo migration). Docs + registries =
W2-B's own STEP 6/7 (feat pipeline tail); W2-A runs no doc-sync (deviation,
stated: a mod-only diff has no public doc of its own until B lands).

## Decisions (pass B, user 2026-10-09) — amended by the challenge
- D1 the five `effort-*` skills are DELETED in W2-B. The mod's
  `Skill(effort-*)` bridge is deleted in W2-B too (same commit as the citers,
  robustness 7: edits are live on the symlinked tree). The typed `/effort-`
  floor code goes in W2-A (A1). The `ultrathink` floor stays.
- D2 pins reworked, not deleted (r3): rows are PHASES by ROLE, no bare
  levels; the mod routes agents on both axes while on (model written at
  spawn from the tier, only UPWARD in rank; effort per step; explicit
  Agent params win). The tracked frontmatter stays as the OFF-STATE FLOOR
  (mod off / `enabled:false` / unloaded / a hook failing open → the engine
  applies the frontmatter as today: never the parent model, never session
  effort on an xhigh gate agent; r2 robustness 1 + confirmation 8). The
  census locks every row equal to its frontmatter (tier head == `model:`
  alias, phase effort == `effort:`), so there is one declared value, two
  carriers. Deleted: the five shifters, `lib/effort-pins.*` (vendored
  levels become rows; off-state = session level for them, as before
  BDR-108), the witness script.
- D3 the orchestrators declare PHASES: dispatch span → `route(phase=
  "orchestrate")`; own level high → `reflect`, xhigh → `plan`; bookkeeping
  tail → `route(phase="apply")` (work/low); escalation → `escalate`.
  Judgment dispatches of BUILT-INS (`general-purpose` with `model="opus"`
  or `"fable"`) carry an explicit `effort=` param (a main route never
  reaches a child; correctness 8). A best-tier skill row SURVIVES the end
  of the turn in its own slot (A6 `runMain`): a run spans prose gates;
  turn-scoped routes (`route` calls, prompt rules, the bridge) never
  touch it.
- D4 `lib/model-gate.md` = the mod rule; the witness is the `route` tool's
  own answer (correctness 7): self-check big → silent; small → call
  `route(phase=<entry phase>)` and STOP unless the answer names a fable or
  opus id; tool absent or "is off" → STOP with the remedy. `model-check.sh`
  + its test deleted. The route answer ALWAYS names the id the next main
  step runs on (A3d). Relaunch levers in STOP texts: `ultrathink` in the
  relaunch prompt (turn floor) or `/route effort=max` (sticky, `/route
  clear` after). Builtin `/effort` is NOT a lever inside a run (rows and
  routes rank above the engine effort; r2's engine-effort detector dropped
  as unsafe, confirmation 4) — documented, named to the user.

## Phase table (A2a) — two rows added to `phases`
| phase | tier | effort | role |
| plan | best | xhigh | brainstorm, plan, architecture, audit verdict |
| reflect | best | high | diagnosis, review, contract, day-to-day orchestrator entry |
| orchestrate | best | medium | between dispatches |
| escalate | best | max | stuck, cap reached |
| judge | big | xhigh | dispatched challengers, analyzers, audits |
| implement | work | medium | code from a closed plan |
| write | work | **high** (was medium) | docs, commits, refactors, handover prose on sonnet at high (BDR-107 "high judgment on sonnet") |
| verify | work | xhigh | verifier, security-auditor |
| explore | work | medium | Explore |
| **apply** (new) | work | low | low appliers (BDR-107): small fixes, release mechanics, probes, validators; bookkeeping tail of the main loop |
| mechanical | cheap | low | listing, status, profile toggles |

## Row tables (A2b)
skills (56):
- plan: ship-feature init-project onboard tour audit-delta analyze
  code-clean client-handover brainstorming writing-plans
  requesting-code-review 21st-ui-review
- reflect: feat hotfix bugfix refactor web-validate harden seo geo
  site-motion frontend-design emil-design-eng design-motion-principles
  21st-ui-build scroll-world-storytelling build-threejs-scroll-worlds
  scroll-scrubbed-visual-sequence scroll-scrubbed-word-reveal
  scroll-progress-timeline subagent-driven-development writing-skills
  deprecation-and-migration 21st-ai 21st-ui-explore
- implement: gitflow prune-memory pdf-translate ci-cd-and-automation
  observability-and-instrumentation test-driven-development
- apply: commit-change release-candidate doc capitalize close reconcile
  deploy (work tier: never haiku on main even with the switch on,
  robustness 15)
- mechanical: status profile plugin-check skills-perso using-git-worktrees
  21st-cli-use 21st-registry 21st-design-sync
agents (21 + 2 built-ins), model = frontmatter alias, unchanged everywhere;
`effort:` frontmatter updated where the row differs (analyzer → xhigh):
- implement (sonnet/medium): feater bugfixer code-cleaner scaffolder onboarder
- write (sonnet/high): commit-changer doc-syncer handover-doc-writer refactorer
- apply (sonnet/low): hotfixer release-executor plugin-probe validator-analyzer
- verify (sonnet/xhigh): verifier security-auditor
- judge (opus/xhigh): plan-challenger plugin-advisor seo-analyzer
  geo-analyzer analyzer (the ONE level delta: high → xhigh, BDR-108 rung
  "audit before validation"; named at the gate)
- mechanical (haiku/low): status-reporter
- built-ins: Explore explore, Plan judge
- no row: interviewer, client-handover-writer (inline-load), impeccable-*
  (gitignored vendor output, untouched)
Deltas named at the gate: analyzer effort; best-tier raise now also reaches
refactor + the design/vendored reflect skills on a small session
(simplicity 6: the raise is the point of the mod).

## W2-A — mod (contract 2026-10-09-model-router-w2a-1546)
- [ ] A1 `register.ts`: delete the typed-floor code — `slashEffort`,
      `guardedSlash`, `Source` member `'slash'`, `floorSource`'s `typed /`
      branch, the State comments on the typed `/effort-<l>` floor. KEEP
      `EFFORT_SKILL` + `effortBridge` (tool.call bridge, removed in W2-B)
      and `turnFloor`/`higherFloor`/`floorWord` (ultrathink).
- [ ] A2 `DEFAULT_CONFIG`: phases per § Phase table (`write` high, `apply`
      new); `skills` + `agents` per § Row tables; the "Built-ins only"
      comment → "phase = role; one row per routed repo skill/agent (wave 2);
      a project-level agent of the same name shadows its row (A3)".
- [ ] A3 spawn: (a) explicit model = `e.model !== undefined` at spawn (the
      engine does not pre-fill the frontmatter; no tool_use_id map);
      explicit effort as today (`explicitEffort`). (b) `spawnRoute`:
      `frozen` → none; row lookup for every provider; a row is SKIPPED when
      the agent's definition `source`, recorded per `subagentType` from the
      `agent.offer` event, is `projectSettings` or `localSettings` (a
      foreign repo's own `verifier.md`); `userSettings`, `built-in`,
      `plugin` or no record → the row applies (fail-open on routing, as
      W1). Known limit, stated in a comment: the record is keyed by name
      only (an offer fired inside a sub-agent with another cwd overwrites it). (c) `spawnTarget` for a rowed spawn
      without explicit model: resolve WITHIN the tier only, and write only
      an alias ranked ≥ the tier head in `fallback` (agents move UP, never
      below their frontmatter alias): `big` with opus down → fable; opus +
      fable down → no write + one log line (deduped per agent+tier per
      turn) "model-router: <agent> tier <t> down, frontmatter model kept".
      (d) `routedText`/`mainAnswer` always name the id the next main step
      runs on (`model <id> (<why>)`), the sticky/floor note APPENDED, never
      substituted (the gate reads this answer).
- [ ] A4 typed slash: `typedSlash: string | null` = the first token of a
      slash prompt at `prompt.submit` when `origin.kind` ∈ {composer, sdk,
      bridge} (allowlist; floor/default rules stay composer-only), stored
      only when it is a `cfg.skills` key; mid-turn → `pendingSlash`,
      promoted at `endMainTurn`. `prompt.submit` also records
      `st.promptAllowed` (origin in the allowlist) for the turn it opens
      (pending slot for a mid-turn prompt, like the marker). `skill.prompt`
      on main (`skillCalls === 0`): apply the row when `e.skill ===
      typedSlash` (then null it) OR when `promptAllowed && loops.size === 0
      && spawning === 0` (`spawning` = a counter held from spawn-hook entry
      to after `next`); otherwise `next(e)`. Verbose log names which path
      fired (`typed-marker` / `typed-fallback`).
- [ ] A5 `onSkillLoad`: a skill WITHOUT a row leaves `turnMain` untouched
      (find-docs / gstack / plugin skills mid-run no longer clear the
      run's route; correctness 11); a rowed skill replaces it. Same rule
      INSIDE a sub-agent (gated 2026-10-09, feater NEED-DECISION): an
      unrowed skill leaves the agent loop's effort untouched; a rowed one
      writes it (test: `feater` loop at medium, `Skill(find-docs)` with
      that agentId → next step still medium).
- [ ] A6 run slot: new `runMain: Routed | null`. A rowed skill load ON
      MAIN whose phase tier is `best` writes BOTH `turnMain` (source
      `skill`, as today) and `runMain`; a non-best row loaded by the
      model's `Skill` tool writes `turnMain` only (helper skills such as
      `using-git-worktrees` inside SDD never drop the run); a USER-TYPED
      rowed skill (marker path) replaces `runMain` with its row when best,
      drops it when not. Precedence `userMain > turnMain > runMain > floor
      > engine` (per axis, as today); `mainRoute` (statusline, spinner,
      planStep) includes it; `endMainTurn` leaves it; `pushOrchestrate`
      writes `turnMain` when empty exactly as today (derived orchestrate
      overrides the run default for the dispatch span, like a `route
      orchestrate` call); cleared by `/route clear` (text: "run slot
      dropped" when one held), `/route off`, a user `/model` switch
      (`PostModelSwitch` source command|picker|sdk); `route(clear=true)`
      from the model clears `turnMain` only and its answer says "run
      <phase> still holds". A `Skill` call inside a sub-agent never touches
      it. `/route show` prints `run <phase>` when no turn route is in force.
- [ ] A7 (dropped in r3: engine-effort detector, confirmation 4).
- [ ] A8 `mergeTable` accepts `null` in the override's `skills`/`agents`
      to drop a default row (robustness 4).
- [ ] A9 `register.test.ts`: delete the typed `/effort-*` floor tests;
      keep the bridge test; add — typed `/feat` (prompt.submit `/feat x`,
      origin composer, then skill.prompt feat) → `main: skill reflect`,
      `effort high`, `[tier best]`; origin `channel`: not armed AND the
      idle fallback refused (main untouched); skill.prompt `feat` with a
      live sub-agent and no marker → main untouched; same with no live loop
      and an allowed origin → routed; a mid-turn `/status` → pending,
      applied after turn.complete; `Skill(find-docs)` (no row) after
      `route plan` keeps plan — the existing test 'floor: survives a skill
      load' (register.test.ts:341-348) is REWRITTEN to assert orchestrate
      KEPT (A5 behavior, authorized here); run slot: `Skill(feat)`, then
      `route orchestrate`, then `turn.complete` → `/route show` back to
      `run reflect`; `Skill(using-git-worktrees)` (mechanical, model path)
      keeps `run reflect` under a mechanical turn route; a TYPED `/status`
      drops the run slot; `/route clear` drops it; `/route off` drops it;
      PostModelSwitch source `command` drops it; a `Skill(feat)` with an
      agentId leaves `runMain`; `Skill(effort-low)` bridge then
      `turn.complete` → session defaults (not sticky); `feater` spawn (provider
      `{plugin:'engine',tier:'core'}`, no model) → `started.model` = sonnet
      full id and the first `turn.step` of that agentId at medium;
      `plan-challenger` with opus down → fable id; opus AND fable down →
      no write, one log line; explicit `model`/`effort` win; `fork: true`
      untouched; a project-source `agent.offer` record → no write; route
      answer text names the id with a floor in force; override `agents: {
      verifier: null }` drops the row (bottom `fs`/`env` mocked like
      `session.model`; if the kit refuses, the case is dropped and said so).
- Disposition: honors BDR-115 (one writer per axis while on, full ids,
  config tables, closure state; rule 4 closed, rule 5 amended by A5/A6),
  BDR-107/108 (roles + levels preserved except analyzer), BDR-076/077
  (judge rows = opus, off-state floor kept), LRN-205, LRN-206, LRN-207
  (breaker feeds the tier resolution).

## Gate between A and B — live probe (user runs `/reload-plugins`)
Evidence into the W2-B contract: (1) typed `/status` → `/route show` main
`skill mechanical`; (2) a real `plugin-probe` or `feater` dispatch with
verbose on → spawn log line (provider shape, model id written) + `step 0
agent …` effort line = spawn/first-step ordering fact the TODO asks for;
(3) on a SONNET session (`/model sonnet`, then back): typed `/feat` → the
self-check wording of the system prompt + the `route` answer id (gate
witness fact); (4) probe (1) again with a background Explore alive (marker
path vs fallback path in the verbose log). The verbose log line (`typed-marker` /
`typed-fallback`, `spawn … → <id>`, `step 0 agent …`) is the witness for
(1), (2), (4), not `/route show` after the turn (turn-scoped routes are
gone by then). Decision rules: (2) step 0 BEFORE the spawn bookkeeping →
step 0 runs on the frontmatter `effort:` (kept, D2), recorded as a known
limit; (4) `typed-marker` never seen → W2-B blocked, A4 re-planned (the
`prompt.submit` text fact does not hold). Only then W2-B.

## W2-B — repo migration (own contract)
- [ ] B0 mod: delete `EFFORT_SKILL`, `effortBridge`, the Skill-hook branch,
      their test; `agents/`+`skills/` rows unchanged.
- [ ] B1 `lib/effort-shift.md` rewritten (~40 lines): route doctrine, the
      wiring points in `route` terms (D3), judgment built-ins carry
      `effort=`, sticky skill route + `/route clear`, no pairing rule,
      headless OK (hooks run under -p), levers = `ultrathink` / `/route
      effort=max` (builtin `/effort` is not a lever inside a run), last ROWED
      skill loaded wins
      (an unrowed one changes nothing).
- [ ] B2 citers `Skill(effort-*)` → `route`: ship-feature 9, init-project
      5, feat 4, bugfix 4, web-validate 3, seo 3, hotfix 3, geo 3,
      verify-secure-loop 3, harden 2, code-clean 2, audit-delta 2, onboard
      1, client-handover-writer 1; EFFORT SHIFTS header lines in the 13
      orchestrators + tour:33; `effort=` added to every `general-purpose`
      judgment dispatch (`model="opus"`: ship-feature, init-project,
      onboard ×7, tour; `model: "fable"` skill-runners in
      client-handover-writer); STOP/remedy texts (challenge-plan.md:62,
      verify-secure-loop.md:109) → the D4 levers; the 21 prose sites
      stating "sonnet by frontmatter pin" / "`model: opus`-pinned,
      session-independent" (challenge-plan:46, feat:146, bugfix:160,
      hotfix:138, code-clean:175, seo:463, 6 agents…) reworded "routed by
      the model-router row (frontmatter = off-state floor)".
- [ ] B3 `lib/model-gate.md` → D4 (≤ 25 lines, keeps `model: "fable"` for
      skill-runners and the dispatch-tier table); delete
      `lib/model-check.sh`, `lib/tests/model-check.test.sh`.
- [ ] B4 delete `lib/effort-pins.txt`, `lib/effort-pins.sh`,
      `lib/tests/effort-pins.test.sh`; remove the re-apply blocks + comments
      in `install-plugins.sh` (941-943, 1008 comment, 1131-1136) and
      `update-all.sh` (472, 567-572); `lib/tests/higgsfield.test.sh`
      :324 + the `before-pins` check (:330-356) dropped.
- [ ] B5 `skills/effort-*` deleted (5 dirs; `profile`/catalog lists
      grepped). Tracked frontmatter KEPT and ALIGNED: `agents/analyzer.md`
      `effort: xhigh`; the vendored tracked `design-motion-principles`
      keeps `effort: high`; impeccable-* (gitignored) untouched; every
      other `model:`/`effort:` value already equals its row.
- [ ] B6 census: `lib/tests/effort-routing.test.sh` rewritten — every
      tracked `effort:` equals its row's phase effort and every agent
      `model:` alias equals its row's tier head (drift lock, both
      directions, parsed from `register.ts`; haiku rows exempt from the
      effort direction: no effort on haiku); enumeration = tracked
      `SKILL.md` under `skills/` + `skills-external/` (`git ls-files`),
      no-row list = graphify, model-router, find-docs, impeccable, the
      agents interviewer/client-handover-writer/impeccable-*; no
      `Skill(effort-` anywhere in
      skills/agents/lib; D3 wiring markers (orchestrate in the 11, apply
      tail in the 5, escalate ×3 in verify-secure-loop + ship-feature,
      `effort=` on every `model="opus"` general-purpose dispatch); every
      tracked skill/agent (minus the no-row list) has a row in
      `register.ts` with the EXPECTED PHASE (per-name asserts, not mere
      presence); `model-routing.test.sh` :96-97 model-gate locks
      updated (the 18 `model:` locks, loops-light.test.sh:74,79 and
      plan-challenger.test.sh:20 stay valid: frontmatter kept); `CLAUDE.global.md` Design
      work lines 299-301 → "rows in the mod, an unrowed member changes
      nothing".
- [ ] B7 doctrine-citers census; full `make test` once before merge.
- [ ] STEP 6/7 of the B run: doc-syncer audit → README/USAGE/ARCHITECTURE/
      CHANGELOG; BDR-115 amendment (wave 2 closed, D1-D4, A5/A6), LRN
      (typed-slash marker + sticky skill route), EVAL on the run; TODO W2
      lines checked.

## Challenge ledger (r1 → r2)
- robustness 1 BLOCKER (mod off → agents inherit the parent model) →
  D2: `model:` frontmatter kept as off-state floor; A3c row overrides it
  while on.
- correctness 1/2, robustness 2 (marker unbound, ordering unproven) → A4
  name-bound marker + pending slot + loops-size fallback; live probe gate.
- correctness 3 (show text) → A9 asserts `skill reflect`.
- correctness 4 (56) → fixed everywhere.
- correctness 5, simplicity 1/2, robustness 13 (level deltas) → `write`
  high + `apply` phase; one delta left (analyzer), named.
- correctness 6 (raise lasts one turn) → A6 sticky skill route.
- correctness 7, robustness 5, simplicity 8 (gate witness) → D4 route answer.
- correctness 8, robustness 11, simplicity 7 (point 5) → explicit `effort=`.
- correctness 9, robustness 6 (levers) → D4 texts (r2's A7 engine-effort
  detector dropped again in r3).
- correctness 10, robustness 9/10 (census) → B6 per-phase asserts, suites
  listed; no `effort="high"` call-site edits (write = high).
- correctness 11 (unrowed skill clears) → A5.
- correctness 12, robustness 12, simplicity 11 (counts, impeccable) → B5
  from `git ls-files`; impeccable untouched.
- correctness 13 (dead code) → A1 list; contract criterion 5 grep widened.
- correctness 14 (override test needs fs) → A9 bottom mocks or drop, stated.
- correctness 15, robustness 4/8 (provider shape, collisions) → A3b
  source from `agent.offer` + A8 null rows; tests use `engine/core`.
- robustness 3 (haiku via fallback chain) → A3c tier-only + log.
- robustness 7 (live edits mid-migration) → bridge stays until B0; probe
  gate before B.
- robustness 14 (tour:33, install comment) → B2/B4.
- robustness 15 (deploy on haiku) → `apply` rows.
- simplicity 3 (tail) → kept as `route(phase="apply")` (phases only; the
  switch-on cost is the user's opt-in). [kept, reasoned]
- simplicity 4 (W2-C) → folded into W2-B STEP 6/7.
- simplicity 12 (two parsers) → contract oracle = one-shot values; B6 =
  durable per-phase census. [kept, reasoned]

## Confirmation ledger (r2 → r3)
- conf 1 BLOCKER (route calls wipe the sticky slot) → A6 separate `runMain`
  slot, turn writers never touch it.
- conf 2 (low/haiku leaks across turns) → only best-tier rows enter `runMain`.
- conf 3 (bridge sticky in the A→B window) → bridge writes `turnMain` only.
- conf 4 (engine-effort detector) → A7 dropped; builtin `/effort` named as
  a non-lever to the user; levers = `ultrathink`, `/route effort=max`.
- conf 5 (route answer without id) → A3d always names the id, note appended;
  probe (3) on a sonnet session.
- conf 6 (fs/cwd shadow unreliable) → A3b definition source from
  `agent.offer`; fs dropped.
- conf 7 (judge → sonnet in-tier) → A3c rank ≥ tier head, else no write + log.
- conf 8 (off-state quality drop, step-0 ordering) → D2 frontmatter kept
  as floor on both axes, census-locked; probe decision rule written.
- conf 9 (pre-fill premise) → explicit = `e.model !== undefined`, no map.
- conf 10 (origin denylist) → allowlist composer|sdk|bridge.
- conf 11 (preload before trackLoop) → spawning counter; log names the path.
- conf 12 (kit has no fs) → fs dropped. conf 13 → deduped log.
- conf 14 (skill rows shadow) → accepted, named: skill rows affect main
  only; a foreign skill of the same name gets the row for its turn.
- conf 15 (artefact drift) → contract clarification amended; status-reporter
  listed under mechanical.

## Confirmation 2 ledger (r3 → r4)
- conf2 1 (helper skill drops the run) → A6: model-loaded non-best rows
  write turnMain only; typed rowed skills manage runMain.
- conf2 2 (test :347 goes red) → A9 authorizes its rewrite.
- conf2 3 (fallback ignores origin) → A4 `promptAllowed` flag on the fallback.
- conf2 4/5/6 (runMain wiring) → A6 names both slots, mainRoute, clearLoop
  text, pushOrchestrate, `/model` + `/route off` + sub-agent cases in A9.
- conf2 7 (probe witnesses) → gate: verbose lines + decision rules.
- conf2 8 (grep) → contract criterion 5 narrowed.
- conf2 9 (source tokens) → A3b names projectSettings|localSettings + limit.
- conf2 10/11 (census) → B6 enumeration, haiku exemption, cites fixed.