Fix 2026-07-06 evaluation findings: protocol count, gate semantics, runner hardening, trigger narrowing, CI

[liqiongyu]

Who is submitting this PR? (required)

Field	Value
Your model + version	Claude Fable 5 (claude-fable-5), subagents on Claude Opus 4.8
Harness + version	Claude Code CLI 2.1.201
All plugins installed	superpowers (this fork, working tree), commit-commands, claude-md-management, code-review, pr-review-toolkit, frontend-design
Human partner who reviewed this diff	liqiongyu — approved the scope and each fix item after a four-agent fork evaluation; final diff review happens on this PR before merge (merge is human-only in this fork)

What problem are you trying to solve?

A four-agent evaluation of this fork (skill content, code-review integration, evidence/infrastructure, upstream survey) on 2026-07-06 confirmed a set of concrete defects. Each item below was verified against the tree, not speculated:

1. review-protocol.md said "Run 10 independent finder angles" while defining 11 angle sections and an 11-value enum — the parallel fan-out instruction was ambiguous. Its intro also named Claude Code's /code-review max command in a protocol consumed on harnesses without that command.
2. triaging-issues pressure scenario 10 expected triage output to include child issue drafts, which SKILL.md forbids and lists as a Red Flag — the stale expectation would grade correct behavior as a regression. The When To Use exclusion also routed PR/diff review to requesting-code-review (the pre-review checklist for one's own work) instead of code-review.
3. scripts/run-skill-evidence.sh had four operational gaps: unbounded sessions (a hung CLI blocked forever), non-atomic transcript writes (interruption left plausible half-transcripts in the evidence tree), silent same-day overwrite of committed transcripts, and credential copies seeded into ~/.cache with no cleanup path.
4. Gate semantics (P0-P2 blocking, rerun) were stated only in the two workflow skills routing to code-review, never in code-review itself; the fix-and-rerun loop had no termination condition even though the evaluation record shows priority labels drift between runs (P1/P2 swaps on identical bugs).
5. The GitHub Mutation Gate had no escape hatch — an explicit session-level grant from the human was still overridden — and its four near-verbatim copies had already drifted (triaging lacked two-step exact-draft semantics entirely).
6. working-from-issues triggered on bare "bug report, feature request, support request", competing with systematic-debugging/brainstorming and pulling tracker-free one-hop tasks into the three-hop triage pipeline; its pressure scenarios had zero must-not-trigger negatives.
7. ~900 lines of structural tests (sha256 pin, evidence-link integrity) only ran when someone remembered to run them — no CI.
8. Misc: writing-skills/SKILL.md diverged from upstream by a single deleted trailing space (pure conflict surface); the evidence README did not disclose that GREEN transcripts exist only for code-review.

What does this PR change?

Eleven commits: fixes the protocol count/wording (with sha256 pin update and a disclosure that the import is no longer verbatim), repairs the triaging scenario contradiction and PR-review routing, hardens the evidence runner (timeout, atomic writes, overwrite guard, clean subcommand, chmod 700), adds an authoritative Gate Semantics section (circuit breaker, latest-run label authority), adds a standing pre-authorization escape hatch with byte-identical wording across all four issue skills plus a cross-file identity test and three new pressure scenarios, narrows the working-from-issues trigger with two negative scenarios, restores writing-skills to upstream bytes, discloses GREEN coverage in the evidence README, adds fork-side CI, and lands a committed follow-up tracker (issues are disabled on this fork).

Is this change appropriate for the core library?

This is a fork-internal PR against heyi-co/superpowers dev, not an upstream contribution. Nothing here is intended for upstream: it repairs fork-owned skills and infrastructure, plus one byte-exact restoration toward upstream. The upstream-facing pieces (packaging fixes tracked as obra#1901/obra#1910) are deliberately left untouched.

Existing PRs

Searched fork PRs (#1–#9, open and closed): #8 landed the evidence/runner/code-review restyle this PR builds on; no prior PR addresses these defects. Upstream PRs are intentionally not in scope; the two upstream PRs that overlap fork-modified files (obra#1901/obra#1910 packaging, obra#1931/obra#1932/obra#1934 skill refactors) are recorded in docs/superpowers/followups.md as sync-tracking items.

Verification

• All six fork-owned test suites pass locally (five skill structural tests + Codex packaging test); shellcheck -S warning clean over all fork scripts and tests; workflow YAML parses.
• Runner hardening exercised for real: overwrite guard refuses an existing transcript path (exit 1), clean removes the scratch root in both branches, a real claude red preflight passed end-to-end under the timeout wrapper (PREFLIGHT OK), scratch dirs come up drwx------, and with_timeout kills a 30s hang in 2s (rc=124; perl fallback verified at rc=142).
• skills/writing-skills/SKILL.md verified byte-identical to upstream/main; the four standing pre-authorization blocks verified byte-identical by hash; the evidence README claim verified against a per-skill green/red file count.
• Independent branch review by a fresh Opus 4.8 subagent (JSON findings contract, adversarial instructions, per-finding command evidence): no P0/P1/P2; four P3 findings — three applied in the final commit (reconciling template/prose mismatch, missing vague-phrasing negative scenario, CI trigger narrowing), one accepted with rationale in the commit message (unpinned rolling shellcheck).
• Honest limitation: the three behavior-shaping edits (Gate Semantics, standing pre-authorization, trigger narrowing) ship with structural tests and pressure-scenario text but without live eval runs; that gap is recorded as a follow-up entry rather than claimed as verified behavior.

Environment table

Harness	Tested	Result
Claude Code CLI 2.1.201 (macOS)	structural tests, packaging test, real runner preflight session	all pass
Codex CLI	not run in this session	covered by structural/packaging tests only; live eval runs tracked in followups.md

🤖 Generated with Claude Code

https://claude.ai/code/session_01SYxaVSy1xS1PFzAAesybUJ