Skip to content

One /go front door: skill ecosystem consolidation, model routing, real-gate alignment (A/B-proven 76 vs 44) - #7

Draft
anotherben wants to merge 10 commits into
mainfrom
skill-refactor
Draft

One /go front door: skill ecosystem consolidation, model routing, real-gate alignment (A/B-proven 76 vs 44)#7
anotherben wants to merge 10 commits into
mainfrom
skill-refactor

Conversation

@anotherben

@anotherben anotherben commented Jul 3, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner

What changed

  • NEW /go dispatcher (skills/go/SKILL.md): universal entry point — classify → depth gate (ANSWER/QUICK/STANDARD/DEEP/DEBUG/OPS, DEBUG+OPS categorical) → model routing (opus/sonnet/haiku/codex) → universal SHIP stage.
  • NEW skills/go/GATES.md: single source of truth for the real merge gates (30-PR mining + memories), incl. operator-authorization rule (dev autonomous, main human-gated).
  • NEW /advisor: consult-only multi-model second opinion (opus + codex, one recommendation with confidence/reversibility).
  • NEW /incident: prod incident response / rollback / post-deploy watch (top coverage-matrix gap).
  • Enterprise family: Claude-native rewrite of an unrunnable Codex/GPT-5.5 harness port, 17 → 9 skills; explicit artifact handoffs (docs/designs/*-{discover,tdd,plan}.md → docs/verify/-receipts.jsonl).
  • Entry points: full-cycle ×4 + vault-process superseded by /go; vault-status→vault-sweep; vault-triage→triage; bug-factory = canonical SHIP stage.
  • Guards: patch-or-fix + zoom-out → /diagnose modes (with FIX/PARTIAL/PATCH-class roll-up contract); scope-check → session-heartbeat; sql-guard gained migration-operations rules (55P03 lock-cascade trap); integration-guard gained the unmapped-tender fail-loudly trap (bug #3618 class).
  • 20 skills archived to skills/.archive/ (recoverable). Net ~−4,600 lines.

Proof (docs/ab-proof/2026-07-04-skill-refactor/)

  1. Blind A/B, opus judge: new 76/90 vs old 44/90; ship-stage gate fidelity 10–0; old scored 0 on model routing in all 8 scenarios (judge-scores.md).
  2. Codex adversarial review of the diff: 3 HIGH findings, all triaged (2 confirmed + fixed, 1 partially accepted) (codex-adversarial-review.md).
  3. Per-skill test→fix→retest to convergence (skill-test-ledger.md): every one of 33 skills cold-read tested by fresh agents against realistic scenarios; every fix retested; all CONVERGED at zero HIGH+MED findings (3 cosmetic LOWs recorded). 14 test/audit/judge agents total. Notable catches: 3 enterprise artifact-handoff holes, /go's DEBUG/OPS routing ambiguity, a docs/plans-vs-GATES collision, diagnose↔bug-factory verdict-vocabulary contract, one retest false positive (live launchd infra) correctly kept.
  4. Coverage matrix (coverage-matrix.md): 14 lifecycle phases × 10 situation types; genuine gaps closed by /advisor + /incident + sql-guard rollout section; perf-regression skill deferred.

Deployment

Synced to live ~/.claude/skills (50 → 33 skills incl. 2 new); removed set backed up at ~/.claude/skills/.removed-backup-2026-07-04/; live ~/.codex daily-sweep copy given the same __dirname fix (syntax-checked).

Follow-ups

  • Reference sweep across the 51 repo-only (never-installed) skills.
  • A/B-3 live worker race (Claude vs Codex on a real helpdesk bug).
  • Perf-regression skill; 3 recorded LOWs.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud

Ben Van Der Veen and others added 10 commits July 4, 2026 09:15
Ben Van Der Veen @claude Ben Van Der Veen
sync: reconcile 14 live-edited skills from ~/.claude/skills into repo
6717aa9 
Live versions are newer (edited in place after last sync).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Ben Van Der Veen @claude Ben Van Der Veen
refactor(skills): one /go front door, real-gate GATES.md, 20 skills c…
f0262e4 
…onsolidated/archived

- NEW skills/go: universal dispatcher (classify -> depth gate -> model routing
  opus/sonnet/haiku/codex -> universal SHIP stage) + GATES.md as the single
  source of truth for the real helpdesk merge gates (from 30-PR mining + memories).
- Enterprise family: Claude-native rewrite (was an unrunnable Codex/GPT-5.5
  harness port), 17 -> 9 skills: plan+contract merged, review+forge+verify+harness
  merged, dev/stack-review/next archived, ~1,500 duplicated ceremony lines removed.
- Entry points: full-cycle x4 and vault-process superseded by /go (+/go queue);
  vault-status -> vault-sweep; vault-triage -> triage; vault-context hook-marker
  hack made explicit; bug-factory kept load-bearing, cross-referenced as the SHIP stage.
- Guards: patch-or-fix + zoom-out -> /diagnose modes; scope-check -> session-heartbeat;
  grill-me -> grill-with-docs; senior-architect/deploy-checklist/run-verification
  archived; blast-radius trimmed 25% (docs/blast-radius/ noted as LIVE CI input).
- Archived skills moved to skills/.archive/ (recoverable), stale cross-references fixed.

Based on 5 parallel audits (enterprise, entry-points, guards, codex plugin, PR-reality
mining of 30 merged PRs).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Ben Van Der Veen @claude Ben Van Der Veen
fix(skills): fold in A/B judge + codex adversarial findings; attach p…
0a205c3 
…roof artifacts

- /go: OPS routing row (-> /promote etc), intake normalize + yield-check-before-claim,
  depth tie-breaker (commit low, escalate; risk surfaces force higher)
- GATES.md: operator-authorization rule (dev autonomous, main human-gated)
- bug-factory: 5-header -> 6-section PR body (matches real gate; codex HIGH finding)
- proof-chain: patch-or-fix -> /diagnose post-fix mode (codex HIGH finding)
- docs/ab-proof/2026-07-04-skill-refactor/: blind-judge A/B (new 76/90 vs old 44/90,
  ship fidelity 10-0), codex adversarial review, scenarios, design doc

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Ben Van Der Veen @claude Ben Van Der Veen
fix(enterprise): close chain-test findings — artifact handoff holes, …
e82e20d 
…GATES divergences, dead stages

- plan/build/discover now name explicit artifact paths (docs/plans/*-plan.md,
  docs/verify/<slug>-receipts.jsonl, docs/designs/*-discover.md) — fixes the 3
  handoff holes left where .codex state files were removed
- enterprise-pr-review: treadmill mechanics + gh run rerun re-fire + main-merge
  operator-authorization per GATES.md (HIGH: premature-merge risk)
- dead forge/verify/harness stage names purged from debug/pr-review/compound
- diagnose <-> enterprise-debug boundary defined in both skills
- dead Codex-harness support files deleted (discover refs/scripts, build script,
  pr-review scripts, 8 stale evals.json)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Ben Van Der Veen @claude Ben Van Der Veen
fix(skills): diagnose <-> enterprise-debug boundary notes (chain-test…
696d115 
… T3)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Ben Van Der Veen @claude Ben Van Der Veen
feat(skills): /incident + /advisor skills, coverage-matrix fixes (wav…
b355151 
…e-1 findings)

- NEW /incident: prod incident response, rollback (pinned deploys, migration
  caveats), post-deploy watch, operator gating — closes top-2 coverage gaps
- NEW /advisor: consult-only multi-model second opinion (opus+codex), one
  recommendation with confidence/reversibility; wired into /go ANSWER lane
- sql-guard: migration OPERATIONS rules (lock-cascade trap, migrationLockTimeout
  owner, out-of-band-run ban) from prod 55P03 postmortem
- pr-schema-audit + code-variable-audit: hardcoded ~/.codex paths -> skill dir
- triage: decision/note items get a Done/Reference route (was unroutable)
- vault-init: dead vault-status pointer -> vault-sweep
- coverage matrix committed to proof dir; perf-regression skill deferred (noted)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Ben Van Der Veen @claude Ben Van Der Veen
fix(enterprise): retest-2 MEDs — plan artifact to docs/designs/, purg…
5f30351 
…e residual forge/verify enum names

Enterprise chain retest verdict: 3 prior HIGH holes closed, GATES divergences
converged, diagnose/debug boundary converged. Remaining 2 MEDs fixed here:
- docs/plans/ collided with GATES dead-plans rule -> docs/designs/<slug>-plan.md
- 5 enum-list lines still said forge lens / verify command -> review lens /
  release-proof check (grep-verified zero residuals)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Ben Van Der Veen @claude Ben Van Der Veen
fix(guards): diagnose verdict roll-up + mapping-only bypass; integrat…
26a6cf8 
…ion-guard unmapped-value trap

Guards retest-2: blast-radius + sql-guard CONVERGED. Fixes here:
- diagnose post-fix mode: explicit FIX/PARTIAL/PATCH-class roll-up so bug-factory's
  exit gate contract holds while keeping the fine-grained 8-value vocabulary (HIGH)
- diagnose system-mapping mode: mapping-only requests skip Phases 1-2 (HIGH)
- integration-guard: unmapped external tender/enum fail-loudly line — would have
  caught bug #3618 gift-card poison class (MED) + vault-MCP fallback note (LOW)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Ben Van Der Veen @claude Ben Van Der Veen
fix(skills): final-wave fixes — go DEBUG/OPS ranking + SHIP dedup, tr…
a518d4e 
…iage folder creation, misc

- go: DEBUG/OPS declared categorical (not on the QUICK<STANDARD<DEEP ladder) — #3618
  scenario now deterministically routes DEBUG; SHIP bullet no longer restates GATES
  mechanics (was violating its own sole-source rule); acceptance-criteria timing for
  DEBUG-bound work; heading cite fix
- triage: create missing destination folders (vault-capture only makes 00-03)
- advisor: small-decision short-circuit hoisted to step 1
- handover-writer: invalid vault type issue -> note
- pr-schema-audit: daily-sweep AUDIT path now __dirname-relative (live .codex copy
  updated identically); secret path annotated as canonical live infra (retest false
  positive); both audit skills' guardrail names ~/.claude/skills as canonical
verdicts this wave: advisor/incident/session-heartbeat/grill-with-docs/
contract-manager/plan-360-audit/vault-init/code-variable-audit CONVERGED

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Ben Van Der Veen @claude Ben Van Der Veen
docs(proof): final skill-test ledger — all 33 units CONVERGED
b358492 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkcPLtpdXQJDpYawhMtYud
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@anotherben