feat(agents,skills): opus/fable model tiers + skill efficiency pass#278
Conversation
rn-tester, rn-code-explorer, rn-code-reviewer move sonnet → opus; rn-debugger already opus; rn-code-architect moves opus → fable (claude-fable-5 — the top tier, fitting the architect's role as the single deep-reasoning blueprint step in the /rn-feature-dev pipeline, where it runs 1-2× vs the explorers'/reviewers' 2-3× parallel fans). Model-tier prose synced: using-rn-dev-agent agent-map tables + decision tree, docs-site agent pages + index tables. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… stale-claim fixes - rn-feature-development (5,076→3,962 words, −22%): Phase 8 defers to commands/proof-capture.md as the single protocol source (was a ~1,800-word drifting duplicate) with pipeline deltas on top; 8 per-phase evaluator lines collapsed to one core principle; description rewritten trigger-only per the CSO rule (workflow-summarizing descriptions make the body get skipped). - using-rn-dev-agent (2,065→1,825): HELPERS_NOT_INJECTED recovery moved to rn-debugging (its owner) with routing pointers; surface counts fixed (76 tools / 14 commands, were 64/16). - rn-testing: M7 section slimmed to a 5-key table + creating-actions pointer; heading preserved for citations. - rn-best-practices / rn-setup: trigger-only descriptions (dropped the rot-prone 118-rule inventory; added failure-phrase triggers). - Correctness: maestro_run/cdp_run_action params claims updated for #272 (two files said "no -e pass-through"); run-action.md citation fixed ("Maestro Flow Standards" → real heading); smoke-test standardized on cdp_run_action (RunRecord + promotion), plain maestro_run reserved for the on-camera replay; <test-app> placeholder defined. Verified: frontmatter parses on all 8 skills; zero stale-claim greps; all cross-references resolve; confined-subagent retrieval test answered the full Phase 8 chain (skill → proof-capture → creating-actions) concretely — its 4 findings (stale Step-1.4 refs, smoke-test contradiction, Steps-range skipping the diagram, undefined placeholder) are fixed in this commit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1eb5e65c81
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| </example> | ||
| tools: Glob, Grep, LS, Read | ||
| model: opus | ||
| model: fable |
There was a problem hiding this comment.
When /rn-feature-dev reaches Phase 4 and spawns rn-code-architect, this frontmatter asks Claude Code for model: fable. I checked the current Claude Code model/subagent docs: supported aliases are default, best, sonnet, opus, haiku, sonnet[1m], opus[1m], and opusplan; non-alias selections need to be full model names. As written, users with this plugin will hit an invalid/unknown model selection instead of getting the architect blueprint, so this should use a supported alias or the full Fable model ID if Claude Code exposes one.
Useful? React with 👍 / 👎.
1 participant
Summary
Two commits: agent model upgrades, then a skill efficiency pass driven by a duplication/staleness audit.
1. Agent models (
5211903)The architect gets the top tier (
fable→ claude-fable-5): it is the pipeline's single 1–2× deep-reasoning blueprint step, vs the others' 2–3× parallel fan-outs. Model-tier prose synced inusing-rn-dev-agent(agent-map tables + decision tree) and the six docs-site agent pages.2. Skill efficiency pass (
1eb5e65)The audit found the real inefficiencies were drifting duplication and stale claims, not just word counts:
rn-feature-development5,076 → 3,962 words (−22%): Phase 8 was a ~1,800-word copy of the/proof-captureprotocol that had already diverged from it (the command gained video-labeling the skill lacked; both carried a claim falsified by feat(skills): creating-actions — guided authoring of reusable Maestro actions #272). Phase 8 now defers tocommands/proof-capture.mdas the single protocol source and lists only the pipeline deltas (architect's flow table as source, persist-as-action via creating-actions Steps 3–6,cdp_run_actionsmoke-test, Deviations section). Eight repeated per-phase evaluator lines collapsed into one core principle. Description rewritten trigger-only per the CSO rule — a workflow-summarizing description makes the body get skipped.using-rn-dev-agent2,065 → 1,825 (−12%) — loads at the start of every RN conversation, so highest leverage: HELPERS_NOT_INJECTED recovery protocol moved torn-debugging(its owner) with routing pointers; stale surface counts fixed (64→76 MCP tools, 16→14 commands).rn-testing: M7 section slimmed to a 5-key table + creating-actions pointer (the full glossary lives there); heading preserved so existing citations resolve.rn-best-practices/rn-setup: trigger-only descriptions (dropped the rot-prone 118-rule inventory; added concrete failure-phrase triggers).maestro_runhas "no-epass-through" (false since feat(skills): creating-actions — guided authoring of reusable Maestro actions #272 — now showparams={KEY: "VALUE"}); broken citation inrun-action.md("Maestro Flow Standards" → real heading); smoke-test standardized oncdp_run_action(RunRecord + auto-promotion) with plainmaestro_runreserved for the on-camera replay (auto-repair must not mutate a flow mid-recording);<test-app>placeholder defined.Verification
<test-app>placeholder — all fixed in this PR before commit.Wiring
rn-dev-agent-pluginminor (covers both commits)🤖 Generated with Claude Code