bchanot/claude
1
0
Fork 1
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(agents): pin dispatched judgment agents to opus — Fable = inline reflection only (BDR-076)

Browse Source 
Reverses the BDR-066 rejected alternative (opus pins on audit agents):
session default is now Fable, so inherit burned Fable quota on every
dispatched audit/challenge. analyzer, plan-challenger, seo/geo/
validator-analyzer pinned model: opus; onboard's 6 general-purpose
audit dispatches carry model="opus"; tour Phase B repointed.
interviewer + client-handover-writer stay unpinned (inline-load only,
a pin there is inert). settings.json default: claude-fable-5[1m].
Census flipped: model-routing §3 + new §11 (61 pass), loops-light 35,
full make test green.
This commit is contained in:
Bastien Chanot
2026-07-19 17:38:55 +02:00
parent 9bc6ab7e07
commit 354ff2644f
10 changed files with 41 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
agents/analyzer.md
+1
View File
@@ -2,6 +2,7 @@
name: analyzer
description: Analyze code, codebase, or problem before any modification. Produces a factual report without proposing solutions. Use proactively before any refactoring, design, or implementation.
tools: Read, Grep, Glob, Bash
model: opus
memory: project
---
agents/geo-analyzer.md
+1
View File
@@ -2,6 +2,7 @@
name: geo-analyzer
description: GEO audit agent for AI search engines — dispatched by /geo and /seo. Audits AI crawlers, llms.txt, entity signals, Schema.org; emits a fix bundle (dispatcher applies), scored report. Classical SEO → seo-analyzer agent.
tools: Read, Edit, Write, Bash, Grep, Glob, WebFetch, WebSearch
model: opus
---
# GEO — Generative Engine Optimization audit, fix & strategy
agents/plan-challenger.md
+5 -2
View File
@@ -2,6 +2,7 @@
name: plan-challenger
description: Fresh independent plan challenger — reads a PLAN file from disk and adversarially attacks it through ONE assigned lens (correctness | robustness | simplicity), then renders structured findings + a verdict. Report-only, never fixes, never implements. Dispatched fresh; blind to the other lenses.
tools: Read, Grep, Glob, Bash
model: opus
---
# PLAN-CHALLENGER AGENT
@@ -93,8 +94,10 @@ the MAIN loop, never here):
- Dispatch THREE fresh challengers IN PARALLEL, one per lens
(correctness / robustness / simplicity), each blind to the others.
- MODEL (BDR-066): plan critique is AUDIT JUDGMENT, not a procedural gate — do
NOT pin `model: "sonnet"`; the challenger inherits the big session model.
- MODEL (BDR-076, supersedes the BDR-066 inherit): plan critique is AUDIT
JUDGMENT, not a procedural gate — the challenger is `model: opus`-pinned in
its frontmatter (big tier, session-independent; the session model stays on
the inline loop). Never `model: "sonnet"` — a silent judgment downgrade.
(Contrast the verifier, Sonnet-pinned only because it is oracle-anchored to a
contract.)
- FAIL-SAFE — never fail open: a malformed/empty verdict, a missing `PROOF`, or
agents/seo-analyzer.md
+1
View File
@@ -2,6 +2,7 @@
name: seo-analyzer
description: 'Classical SEO audit agent (Google, Bing) — dispatched from /seo. Live audit: Core Web Vitals, on-page, technical, local SEO, legal (FR). Emits a fix bundle (dispatcher applies) + scored report. AI/GEO → geo-analyzer agent.'
tools: Read, Edit, Write, Bash, Grep, Glob, WebFetch, WebSearch
model: opus
---
# SEO — Classical Search Engines audit, fix & strategy
agents/validator-analyzer.md
+1
View File
@@ -2,6 +2,7 @@
name: validator-analyzer
description: Web standards audit agent — W3C HTML validity (validator.nu), W3C CSS validity (jigsaw.w3.org), WCAG 2.1 accessibility (axe-core, pa11y, WAVE). Dispatched from /web-validate. Produces scored .claude/audits/VALIDATE.md report with concrete diffs for auto-fixable issues and user actions for judgment-required fixes. Complementary to /harden (security), /seo (indexability), /geo (AI extraction).
tools: Read, Edit, Write, Bash, Grep, Glob, WebFetch
model: opus
---
# Validator — W3C + WCAG audit
lib/challenge-plan.md
+5 -3
View File
@@ -42,9 +42,11 @@ Agent(subagent_type="plan-challenger", description="challenge:<lens>", prompt=""
""")
```
**MODEL (BDR-066):** plan critique is AUDIT JUDGMENT — do NOT pin
`model: "sonnet"`; the challengers inherit the big session model. (The executor
gates stay sonnet; the challenger does not.)
**MODEL (BDR-076, supersedes the BDR-066 inherit):** plan critique is AUDIT
JUDGMENT — the challengers are `model: opus`-pinned in their frontmatter: a big
tier, session-independent, off the session model. The session model (Fable)
keeps only this loop — synthesis, RE-THINK, gate. Never sonnet: that would
silently downgrade the judgment. (The executor gates stay sonnet.)
**Lens framing by `KIND`** (the agent's three lenses, read against the artifact):
- `build-plan` — will it WORK / will it BREAK / is it needlessly COMPLEX.
lib/tests/model-routing.test.sh
+15 -8
View File
@@ -1,5 +1,5 @@
#!/usr/bin/env bash
# lib/tests/model-routing.test.sh — census: gate wiring + pins + executor shape (BDR-066)
# lib/tests/model-routing.test.sh — census: gate wiring + pins + executor shape (BDR-066, BDR-076)
set -u
R="$(cd "$(dirname "$0")/../.." && pwd)"
pass=0; fail=0
@@ -22,7 +22,7 @@ has "agents/feater.md" 'model: sonnet'
has "agents/hotfixer.md" 'model: sonnet'
has "agents/verifier.md" 'model: sonnet'
has "agents/security-auditor.md" 'model: sonnet'
fm_lacks "agents/analyzer.md" 'model:'
has "agents/analyzer.md" 'model: opus'
# 4) /feat executor shape
has "skills/feat/SKILL.md" 'subagent_type="feater"'
has "skills/feat/SKILL.md" 'verify-secure-loop.md'
@@ -54,17 +54,24 @@ lacks "agents/handover-doc-writer.md" 'AskUserQuestion'
lacks "agents/handover-doc-writer.md" 'Agent('
has "agents/client-handover-writer.md" 'subagent_type="handover-doc-writer"'
# 10) post-merge edge fixes (ronde): F1 feater applier carve-out, F2 /refactor
# dispatch + pin, F3 /analyze gated (in loop 1), F4 interviewer un-pinned,
# F5 audit agents' ABSENT pin locked (a stray sonnet pin would silently
# downgrade a live audit even though the skill's gate passed)
# dispatch + pin, F3 /analyze gated (in loop 1)
has "agents/feater.md" 'Applier path'
has "skills/refactor/SKILL.md" 'subagent_type="refactorer"'
has "agents/refactorer.md" 'model: sonnet'
fm_lacks "agents/seo-analyzer.md" 'model:'
fm_lacks "agents/geo-analyzer.md" 'model:'
fm_lacks "agents/validator-analyzer.md" 'model:'
# 11) BDR-076 — session model (Fable) = orchestration + inline reflection ONLY.
# Dispatched judgment agents pinned OPUS (big tier, session-independent;
# never sonnet — that would silently downgrade a live audit). Inline-load-
# only agents (interviewer, client-handover-writer) STAY unpinned: they run
# IN the main loop, a frontmatter pin there is inert and misleads.
has "agents/seo-analyzer.md" 'model: opus'
has "agents/geo-analyzer.md" 'model: opus'
has "agents/validator-analyzer.md" 'model: opus'
has "agents/plan-challenger.md" 'model: opus'
fm_lacks "agents/client-handover-writer.md" 'model:'
fm_lacks "agents/interviewer.md" 'model:'
has "skills/onboard/SKILL.md" 'model="opus"'
has "skills/tour/SKILL.md" 'model="opus"'
has "lib/challenge-plan.md" 'BDR-076'
printf 'model-routing census: %d pass, %d fail\n' "$pass" "$fail"
[ "$fail" -eq 0 ]
settings.json
+1 -1
View File
@@ -250,7 +250,7 @@
"disableBypassPermissionsMode": "disable",
"additionalDirectories": []
},
"model": "opus[1m]",
"model": "claude-fable-5[1m]",
"hooks": {
"SessionStart": [
{
skills/onboard/SKILL.md
+8 -2
View File
@@ -354,7 +354,7 @@ Lire le bloc `audit_stack:` du fichier `~/.claude/lib/project-archetypes/<archet
| Entry | Action | Livraison |
|---|---|---|
| `analyze` | Déjà fait en STEP 5 | L3a |
| `code-clean` | Spawn subagent `general-purpose` (audit-only, inherits session = big model) | L3a |
| `code-clean` | Spawn subagent `general-purpose` (audit-only, `model="opus"` — BDR-076: dispatched audits off the session model) | L3a |
| `cso` | Si gstack ON → Skill(cso). Sinon → Agent general-purpose avec checklist OWASP + deps audit | L3a |
| `doc` | Spawn subagent `doc-syncer` (auto-mode OFF, report-only) | L3a |
| `seo` | Subagents seo-analyzer + geo-analyzer en parallèle | L3b |
@@ -370,7 +370,8 @@ Lancer EN PARALLÈLE (un seul message, plusieurs Agent calls) les audits corresp
```
Agent(
subagent_type="general-purpose",
description="Onboard — code-clean audit only (read-only, big session model)",
model="opus",
description="Onboard — code-clean audit only (read-only, opus)",
prompt="""
AUDIT-ONLY mode — NO fixes, NO refactoring, NO file modifications.
Target: <PROJECT_ROOT>. ARCHETYPE: <archetype>.
@@ -406,6 +407,7 @@ bash $HOME/.claude/lib/toggle-external.sh list 2>/dev/null | grep -E "^gstack\s+
```
Agent(
subagent_type="general-purpose",
model="opus",
description="Onboard — security audit fallback (archetype-adaptive)",
prompt="""
READ-ONLY security audit. No file modifications.
@@ -647,6 +649,7 @@ Si le skill ne supporte pas `--output`, capturer la sortie et écrire à la main
```
Agent(
subagent_type="general-purpose",
model="opus",
description="Onboard — static design review fallback",
prompt="""
AUDIT-ONLY mode — NO edits. Static design review du code UI.
@@ -690,6 +693,7 @@ Puis parser le JSON Lighthouse (scores perf/a11y/bp/seo/pwa + top opportunities)
```
Agent(
subagent_type="general-purpose",
model="opus",
description="Onboard — static perf audit",
prompt="""
AUDIT-ONLY mode — NO edits.
@@ -732,6 +736,7 @@ Parser axe-core résultats (violations, incomplete, inapplicable, passes) → `.
```
Agent(
subagent_type="general-purpose",
model="opus",
description="Onboard — static a11y audit",
prompt="""
AUDIT-ONLY mode — NO edits.
@@ -777,6 +782,7 @@ Spawn un subagent synthétiseur (isolé, chargé uniquement du contenu de `.onbo
```
Agent(
subagent_type="general-purpose",
model="opus",
description="Onboard — synthèse vers .claude/audits/",
prompt="""
Lire tous les fichiers de <PROJECT_ROOT>/.onboard-audit/ :
skills/tour/SKILL.md
+3 -3
View File
@@ -111,9 +111,9 @@ honestly in the summary. Never loop past 3.
### Phase B — CLEAN
1. Dispatch a read-only cleanup audit (analyzer or general-purpose —
inherits the big session model; NOT the sonnet code-cleaner, which is
now a fix executor): dead code, unused imports/exports,
1. Dispatch a read-only cleanup audit (analyzer — opus-pinned, BDR-076 —
or general-purpose with `model="opus"`; NOT the sonnet code-cleaner,
which is now a fix executor): dead code, unused imports/exports,
commented-out blocks, stale flags, norm violations. Findings as
`id | file:line | finding | proposed fix`.
2. Apply **behavior-preserving** fixes only. A finding that would change