feat(session-flow): add Matt's retro findings to /session-flow:retro session mode #12122

Parent #7575 decides "Retro absorbs Matt's retro: "no guardrail for this mistake" and "unwired or broken check" findings, information-access and no-op-in-steering-file categories, writing-for-agents yardstick, pointer after a long search, tool-economy finding (D9, D48, Q524)", and names this as half of the retro-absorb-and-trends-move slice, which was never filed. The owner approved D48 on 2026-10-13 (poteto-review decision row 48): "absorb into /session-flow:retro, adapted (copy where it fits): the "no guardrail for this mistake" and "unwired or broken check" finding (builds decision 9); information-access and no-ops-in-steering-files categories; writing-for-agents as the yardstick for steering files; a pointer after a long search." The owner accepted Q524 on 2026-10-15, which adds the tool-economy line verbatim: "an expensive tool call, or a custom CLI or MCP server whose output costs more tokens than its use needs". The source is Matt Pocock's retro skill at mattpocock/skills@49dd158d, skills/engineering/retro/SKILL.md lines 11 and 17-23. On main, session mode (plugins/session-flow/skills/retro/context/session.md) looks for none of these. The "no guardrail" finding exists only in codify mode (context/codify.md:54-56). The Phase 3 catalog session mode loads already routes three of them to owner skills (reference/ecosystem-improvement-catalog.md:76-79: no-ops to /harness-config:audit-instructions and /harness-config:unhobble, token-costly tools to /developer-experience:build-cli, misleading steering files to /docs-hygiene:write-for-agents); build on those rows rather than adding a second route. The retro row of docs/upstream/mattpocock-skills.md still says ours "took nothing".

Done when

• plugins/session-flow/skills/retro/context/session.md names each of the following as a finding session mode looks for, in Phase 2 or Phase 3, with the route each one takes:
  • No guardrail: a recurring mistake that meets the "Guardrail criterion" section of reference/enforcement-ladder.md but that no check covers. It is reported as a "no guardrail" finding naming that section, as context/codify.md already does.
  • Unwired or broken check: an existing check that should have caught the mistake but is not wired into CI or pre-commit, or is broken. This finding is reported before any new check is proposed.
  • Information access: information the session needed but could not reach, such as dev-server logs not written to a file the agent can read, or a read-only service it had no way to query.
  • No-op in a steering file: a line in CLAUDE.md, AGENTS.md or a rules file that the model would follow without it. It is routed to /harness-config:unhobble or /harness-config:audit-instructions.
  • Pointer after a long search: when the session searched at length to find something, the finding proposes a navigation pointer in the steering file that would have led there.
• Section 2E, "Efficiency assessment", in context/session.md adds a tool-economy bullet that contains the text "an expensive tool call, or a custom CLI or MCP server whose output costs more tokens than its use needs".
• context/session.md states that a finding about a steering file (CLAUDE.md, AGENTS.md, rules) is judged against /docs-hygiene:write-for-agents, read through the Skill tool when it is among the available skills. Without it, the report says the yardstick was not available. The skill is cited by its /plugin:skill name, never by path.
• The "When to propose a check" paragraph in plugins/session-flow/skills/retro/reference/ecosystem-improvement-catalog.md says that an unwired or broken existing check is the finding before any new check is proposed, and that a repository with no check to wire one into is itself a finding. The existing clause "or the repo has nothing to wire a check into" (:85 on main) stays, and the added text says that this case is still reported as a finding, so the change only adds text.
• docs/upstream/mattpocock-skills.md no longer says ours took nothing from Matt's retro:
  • the retro row (:57 on main) and the per-skill summary entry (:319 on main) list the items adopted, citing D48 and Q524;
  • the "Pruning: no-ops" row (:131 on main) says retro now carries the no-op category.
• plugins/session-flow/evals/ gains at least one case whose case.yaml uses the same schema_version as re-codify-routes-check-asks-first/case.yaml:
  • its prompt runs /session-flow:retro session on a session where a lint script that would have caught the mistake exists but no CI or pre-commit step runs it;
  • its expected_outcome says the report names the unwired check as the finding before proposing any new check;
  • its graders include a no-Write and no-Edit grader, because Phase 4 asks before acting.
• The diff edits neither plugins/session-flow/skills/retro/SKILL.md nor plugins/session-flow/skills/retro/evals/evals.json, which slice S2 changes.
• A .changes/session-flow/*.md fragment with bump: minor describes the new findings.
• The PR's CI is green.

Out of scope

• Moving trends to /session-flow:audit-sessions (slice S2).
• The chain-end retro offer and the findings buffer (D50 item 10).
• Building any check: retro proposes, and the owner skill picks the mechanism.

Related

• feat(session-flow): retro routes mechanical findings to check owners and proposes checks by occurrence #6944: routing a mechanical finding to a hook, lint or test.

Parent: #7575