From 4996f2d919e0c0d36500bc8e0551ebc1711e991f Mon Sep 17 00:00:00 2001 From: claude Date: Mon, 20 Jul 2026 16:03:12 +0200 Subject: [PATCH] Revert "feat(models): move the implementer to opus/high" MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This reverts commit cdbc2d0. The bump was a judgment call without evidence, and the planner design intent argues the other way: the plan Iron Law is "EVERY STEP THAT CHANGES CODE INCLUDES THE CODE" and tasks are sized so a subagent executes one "without judgement calls outside its remit" (planner/SKILL.md) — the planner exists precisely to remove design latitude from the implementer, which points at sonnet, not opus. opus is only justified to the degree real plans leave slack (RED-step invention, tree reconciliation, review repair), and that slack is measurable, not felt. Restore sonnet/high; a future bump should be grounded in the loop metrics (spec re-loop count, quality changes_requested rate, end-verify BLOCKED rate) via postmortem, not a gut call. The four fable gates (b591e75) are untouched. --- docs/agent-template.md | 37 ++++++++------------- implement/agents/implementer.md | 2 +- implement/workflows/compiler-driven-edit.js | 5 +-- implement/workflows/implement-loop.js | 10 +++--- 4 files changed, 19 insertions(+), 35 deletions(-) diff --git a/docs/agent-template.md b/docs/agent-template.md index 9b2d78e..7018e2a 100644 --- a/docs/agent-template.md +++ b/docs/agent-template.md @@ -137,18 +137,9 @@ Assignment rule, in priority order: evidence, the in-loop quality review) — run `opus`. Opus is the documented strength for code review, debugging, and deep reasoning. -3. **Mechanical scope** — recon, extraction, compliance-diffing - run `sonnet` (near-opus quality on tightly-scoped work at - lower latency and cost). Plan execution itself — the - `implementer` — is owner-ratified (2026-07-20) onto `opus`: - authored code is the pipeline's primary output, so it takes - the judgment model, but pairs it with `high` effort, not - `xhigh` — it is the most-dispatched in-loop role, and - volume × wall-clock (rule 4) still governs its thinking - budget. The compiler-driven-edit arm keeps dispatching the - implementer agentType at `sonnet`/`high`: compiler-enumerated - propagation is mechanical, and that arm's policy header pins - its own tier. +3. **Mechanical scope** — tightly-scoped execution of a pre-made + plan, recon, extraction, compliance-diffing run `sonnet` + (near-opus coding quality at lower latency and cost). 4. **Volume × wall-clock** — agents dispatched per-task inside loops or fanned out in parallel swarms multiply their model's latency; they run `sonnet` unless rule 2 overrides (currently @@ -165,9 +156,9 @@ Assignment rule, in priority order: Current distribution: `fable` — architect, debugger, grounding-check, tdd-author. `opus` — bencher, fieldtester, -implementer (at `high`), plan-recon, quality-reviewer, -spec-skeptic. `sonnet` — docwriter, glossary-extractor, -spec-reviewer, synthetic-user, tester. +plan-recon, quality-reviewer, spec-skeptic. `sonnet` — docwriter, +glossary-extractor, implementer, spec-reviewer, synthetic-user, +tester. The same rule binds Workflow scripts: every `agent()` call passes an explicit `model:` option, and `fable` is never a valid value there — @@ -183,16 +174,14 @@ effort, coupling every dispatch's thinking budget to whatever the user happens to be chatting at — the same session-state coupling the `model` pin exists to remove. Effort follows the model split: -- **`xhigh` on every `fable` and `opus` judgment agent.** The - judgement roles are the pipeline's quality floor; their thinking - budget must not silently degrade because the session runs lower. - The fable gates deliberately stay at `xhigh`, not `max` — the +- **`xhigh` on every `fable` and `opus` agent.** The judgement + roles are the pipeline's quality floor; their thinking budget + must not silently degrade because the session runs lower. The + fable gates deliberately stay at `xhigh`, not `max` — the marginal thinking is not worth the extra latency and - scarce-budget burn. One `opus` agent is exempt: the - `implementer` pairs `opus` with `high` (§ model rule 3 — the - most-dispatched in-loop role keeps the in-loop effort tier). -- **`high` on every `sonnet` agent.** Tightly-scoped mechanical - work gains little from `xhigh` but pays its latency on + scarce-budget burn. +- **`high` on every `sonnet` agent.** Tightly-scoped execution of a + pre-made plan gains little from `xhigh` but pays its latency on every dispatch — and these are exactly the roles dispatched per-task inside loops (wall-clock is the pipeline's efficiency metric). Not lower than `high`: a sloppy in-loop step triggers diff --git a/implement/agents/implementer.md b/implement/agents/implementer.md index 31b0d68..e3ec492 100644 --- a/implement/agents/implementer.md +++ b/implement/agents/implementer.md @@ -2,7 +2,7 @@ name: implementer description: Carries out a tightly scoped implementation task. Reads the task extract handed by the controller, implements, builds, tests, reports a structured status with the diff. NOT for architecture decisions, NOT for self-curated scope; this agent executes a plan that has already been made. tools: Read, Edit, Write, Bash, Glob, Grep -model: opus +model: sonnet effort: high --- diff --git a/implement/workflows/compiler-driven-edit.js b/implement/workflows/compiler-driven-edit.js index 148205d..22234ca 100644 --- a/implement/workflows/compiler-driven-edit.js +++ b/implement/workflows/compiler-driven-edit.js @@ -43,10 +43,7 @@ // session-model inherit (which could route to an unintended tier, e.g. fable — // reserved for the four ratified frontmatter gates, never for workflow stages; // agent-template § model rule 1). Both phases here are mechanical -// (compiler-enumerated propagation, build/suite re-run), so both run sonnet — -// including the implementer agentType: its opus frontmatter tier is for plan -// execution, not compiler-enumerated propagation (agent-template § model -// rule 3), so this arm pins sonnet explicitly. +// (compiler-enumerated propagation, build/suite re-run), so both run sonnet. // Effort policy mirrors it: every call pins an explicit effort — the edit // (real code work) runs high, the build/suite verify (schema-bound check) // runs medium. See docs/agent-template.md § effort. diff --git a/implement/workflows/implement-loop.js b/implement/workflows/implement-loop.js index c0fce44..0360dda 100644 --- a/implement/workflows/implement-loop.js +++ b/implement/workflows/implement-loop.js @@ -82,9 +82,7 @@ // session-model inherit (which could route to an unintended tier, e.g. // fable — reserved for the four ratified frontmatter gates, never for // workflow stages; agent-template § model rule 1). Mechanical and -// in-loop check steps run sonnet; the implementer runs opus/high -// (owner-ratified 2026-07-20: judgment model for code authorship, kept at -// the in-loop effort tier); the quality gate runs opus by default (last +// in-loop steps run sonnet; the quality gate runs opus by default (last // correctness review before the end-verify), TIERED to sonnet/high for a // small (<= 25 changed lines), non-contract diff as measured // independently by the spec-reviewer (issue #30) — the end-verify / @@ -794,7 +792,7 @@ async function runTask(task) { 'inner-loop discipline — add the failing test inline even if the task text forgot to script it). ' + 'Build and test must be green before you report DONE. Leave all edits UNSTAGED; never commit.\n\n' + `mode: ${mode}\niter_id: ${iter_id}\n\nTASK ${task.id} — ${task.title}\n${task.text}`, - { agentType: 'implementer', model: 'opus', effort: 'high', label: `impl:${task.id}`, phase: 'Per-task loop', schema: IMPL_SCHEMA }, + { agentType: 'implementer', model: 'sonnet', effort: 'high', label: `impl:${task.id}`, phase: 'Per-task loop', schema: IMPL_SCHEMA }, ) if (!impl) return { id: task.id, title: task.title, outcome: 'BLOCKED', reason: 'implementer agent did not return' } if (impl.status === 'BLOCKED') return { id: task.id, title: task.title, outcome: 'BLOCKED', reason: impl.reason || 'worker-blocked' } @@ -835,7 +833,7 @@ async function runTask(task) { `${STANDING}${ANCHOR}\n\nRepair task ${task.id} to satisfy the spec-compliance review. Address EACH finding; ` + 'do not widen scope. Keep build + tests green; leave edits unstaged.\n\n' + `TASK ${task.id} — ${task.title}\n${task.text}\n\nFINDINGS:\n- ${(spec?.findings || ['(none reported)']).join('\n- ')}`, - { agentType: 'implementer', model: 'opus', effort: 'high', label: `impl-fix-spec:${task.id}`, phase: 'Per-task loop', schema: IMPL_SCHEMA }, + { agentType: 'implementer', model: 'sonnet', effort: 'high', label: `impl-fix-spec:${task.id}`, phase: 'Per-task loop', schema: IMPL_SCHEMA }, ) if (specFix && specFix.applied_changes === true) wroteAnyFile = true } @@ -953,7 +951,7 @@ async function runTask(task) { 'into a deviation the next review flags.\n\n' + `TASK ${task.id} — ${task.title}\n${task.text}\n\nISSUES:\n- ` + (qual?.issues?.length ? qual.issues.map((i) => `[${i.severity}] ${i.text}`) : ['(none reported)']).join('\n- '), - { agentType: 'implementer', model: 'opus', effort: 'high', label: `impl-fix-qual:${task.id}`, phase: 'Per-task loop', schema: IMPL_SCHEMA }, + { agentType: 'implementer', model: 'sonnet', effort: 'high', label: `impl-fix-qual:${task.id}`, phase: 'Per-task loop', schema: IMPL_SCHEMA }, ) if (fix && fix.applied_changes === true) wroteAnyFile = true if (fix && (fix.status === 'BLOCKED' || fix.status === 'NEEDS_CONTEXT')) {