Date: 2026-10-20
Scope: R6 of post-phase-42-review-plan.md:
Status: Decided with Emil (2026-10-20).
| Set | Where it lives | Active? |
|---|---|---|
| Emil’s account skills (17 custom, plus built-ins) | Claude’s account skill system, in every Claude, Cowork and Claude Code session | Yes. All are loaded in the Cowork session of 2026-10-20 (two used in it). docs/CLAUDE_SKILLS.md was resynced the same day |
The repo’s plugin, testforge-skills (skills/testing/, skills/engineering/) |
This repo; enabled by .claude/settings.json through the local marketplace (.claude-plugin/marketplace.json, skills/.claude-plugin/plugin.json) |
Wired correctly on paper. The files are valid and the paths exist. Its last proof by real invocation is 2026-09-16. Cowork doesn’t load repo plugins, so TFP-236 re-proves it in a Claude Code session |
Martin’s proposals (skills/proposed/, 16) |
The repo, not in the plugin | Not active by design. Four are adopted below |
No skill points at a file that’s gone: the four code files the plugin’s skills cite all exist. The drift is in judgement, not in broken links.
skills/testing/)| Skill | Verdict | Why |
|---|---|---|
write-playwright-test |
Update (TFP-238) | Still teaches the Phase 17 categories from prompt.ts (the parked OpenAI path, D11). It knows none of what Phases 38 to 44 settled. See TFP-238 |
crawl-and-classify-pages |
Update: method plus pointer (TFP-237; D12) | A frozen copy of dom-snapshot.ts’s rules on 2026-09-24. It missed four changes then, and Phases 38 to 42 changed introspection much more (label names, controls, #/ routes, link inventories). It has never run against a live browser |
honest-coverage |
Keep | Already a run-time rule since Phase 35 Decision 6 (the fixed “not tested” statement), extended by Phase 44’s scenario pairs. Its judgement is unchanged |
dont-stub-what-you-prove |
Keep | Phase 44 Decision 2 (a test must be able to fail) is this skill applied to model-written tests. TFP-197’s tautologies are its worked case. TFP-239 adds that case to its known answers |
failure-mode-tests |
Keep, with a pointer | Overlaps Phase 44’s scenario 6 and Phase 18’s failure injection. A one-line pointer to Phase 44’s harness is added in TFP-238 |
find-the-hidden-platform |
Keep | Unchanged judgement |
regression-safety-net |
Keep | Unchanged judgement. Tonight’s TFP-210 fix is a fresh example (a green CI step that couldn’t fail) |
plan-to-specs |
Keep | A port with its own --check. Nothing since affects it |
skills/engineering/)| Skill | Verdict | Why |
|---|---|---|
code-structure, refactor-audit, observability-check, reliability-audit |
Keep | Generic judgement. Their cited files (page.tsx, useRun.ts, useCrawl.ts) still exist |
skills/proposed/)| Skill | Verdict | Why |
|---|---|---|
sre-testing |
Adopt (TFP-236) | Test strategy: what to test, how much, what only production can test. Closest to TestForge’s own purpose |
sre-agents |
Adopt (TFP-236) | Agents with access to production or destructive tools: the gap writer, scenario tests, Phase 45’s sandbox |
sre-overload |
Adopt (TFP-236) | Retries, timeouts, rate limits and cascading failure: scenario 6 (Phase 44) and load testing (D24) |
sre-release |
Adopt (TFP-236) | Safe shipping and CI: the nightly workflow, TFP-216 and TFP-217. It overlaps Emil’s verified-release-pipeline. The overlap note it already carries stays |
The other 12 (sre, sre-architecture, sre-code, sre-cost, sre-data-integrity, sre-incident, sre-monitoring, sre-postmortem, sre-slo, sre-toil, dora, dora-capabilities) |
Stay proposed | Not needed by anything in the build order. A note asks Martin which, if any, he wants adopted. His own first pick was sre-code (below) |
The skill describes the method:
For the current rules, it points to the code:
dom-snapshot.ts and the introspect modules (src/lib/introspect/);It no longer copies the rules, so it can’t drift from them again. Rejected:
honest-coverage is already one (Phase 35 Decision 6).dont-stub-what-you-prove applied to
model-written tests.dont-stub-what-you-prove over every generated suite was
rejected: the template is reviewed code, and model-written tests get
Phase 44’s stronger check.Hi Martin — Emil and I went through your sixteen SRE/DORA proposals in the skills relook (R6, 2026-10-20). Four join the
testforge-skillsplugin now because the build order needs them:sre-testing,sre-agents(the gap writer and the Docker sandbox),sre-overload(retry scenarios and load testing) andsre-release(the nightly CI). The other twelve stay inskills/proposed/, untouched. Your own first pick wassre-code; if you’d like it or any of the others adopted, say which and why it earns its place, and it’ll go to Emil.
skillsclaude plugin list (or /plugin) shows testforge-skills enabled,
with every skill. One testing and one engineering skill are invoked
through the Skill tool, as at Phase 21 (TFP-051).sre-testing, sre-agents, sre-overload and
sre-release move to skills/reliability/, which is added to
plugin.json’s skill paths.
proposed/ become plain mentions
with that path, so no link breaks.proposed/sre/ is copied with them.skills/README.md and skills/ROADMAP.md, and
skills/proposed/README.md’s list.skills/, shown by a link check.skillscrawl-and-classify-pages/SKILL.md rewritten: the method and its
reasons, with pointers to the current code and phase records for the
rules;write-playwright-test brought up to Phases 38 to 44skillsfailure-mode-tests gets a pointer to Phase 44’s scenarios.skills, testskills/<category>/<skill>/cases/,
using skill-creator’s eval format where it fits;dont-stub-what-you-prove on TFP-197’s tautologies (must flag);regression-safety-net on the TFP-210 pipe (must say the green step
proved nothing);honest-coverage on a Phase 44 scenario section.