fix(review): re-unbreak /review slash command poisoned by PR #226 regression

[tend-agent]

Summary

PR #226 (fix(ci-runner): filter CI polling by run URL, not workflow name, merged 2026-04-11T15:01:22Z) reintroduced the exact poisoned-text regression that PR #234 fixed. Its diff on plugins/tend-ci-runner/skills/review/SKILL.md and plugins/tend-ci-runner/skills/running-in-ci/SKILL.md reverted the CI-polling comment from PR #234's wording back to the pre-#234 wording containing `if \!` adjacent to backticks — which the slash-command preprocessor scans and exec's as bash regardless of markdown code fences, consuming through the next backtick and trying to run the em-dash-led fragment as a command.

Since PR #226 merged, every /review invocation has failed identically. All three tend-review runs in this hour produced the same 5-line trace:

{"type":"queue-operation","operation":"enqueue"}
{"type":"queue-operation","operation":"dequeue"}
{"type":"user","content":"<local-command-caveat>..."}
{"type":"user","content":"<command-name>/review</command-name> <command-args>N</command-args>"}
{"type":"user","content":"<local-command-stderr>Error: Shell command failed for pattern \"\!` — the Bash tool escapes `\":\n/bin/bash: line 1: —: command not found</local-command-stderr>"}

No assistant turn, no Skill loads, no tool calls, no review. The model was never given instructions to follow — the exact same failure signature PR #234 documented.

Evidence (this hour)

Run	PR	Branch	Session size	Review posted
24286232974	#240	hourly/review-24286117723-permalink-sha	5 lines	❌
24286792302	#241	fix/review-gates-decision-point	5 lines	❌
24287054138	#242	fix/review-gates-trim-decision-point	5 lines	❌

Confirmation via gh pr view <N> --json reviews: all three PRs return []. PRs #240 and #241 were merged by max-sixty without bot review because the bot was silently broken.

No earlier broken tend-review run exists between PR #226's merge at 15:01Z and the current window — #240 was the first PR to land after the regression.

Gate assessment

• Evidence level: Critical — wrong outcome, reviews 100% non-functional since 2026-04-11T15:01:22Z
• Classification: Structural — deterministic preprocessor failure; every invocation produces the identical 5-line trace regardless of which PR is being reviewed
• Occurrences this run: 3 (one per tend-review run in the past hour)
• Historical evidence: N/A for the regression itself (reintroduced today by PR fix(ci-runner): filter CI polling by run URL, not workflow name #226), but the underlying preprocessor bug is well-characterized — PR fix(ci-runner): replace if-not polling loop with || to avoid bash escaping #217 first introduced it, PR fix(review): unbreak /review slash command poisoned by \!\` sequence #234 first fixed it, PR fix(ci-runner): filter CI polling by run URL, not workflow name #226 reintroduced it, this PR fixes it again
• Change type: Targeted fix (rephrase two lines in two files)
• Both gates: pass

Fix

Rephrase the CI-polling comment in both files to avoid any \! adjacent to a backtick. Guidance content is unchanged. Matches the exact wording PR #234 merged — this is a straight restoration of that fix.

Before:

# Use `||` rather than `if \!` — the Bash tool escapes `\!` as `\\!`, which
# prevents bash from recognizing the pipeline-negation reserved word and
# leaves the loop stuck until the 10-minute timeout.

After:

# Use `||` rather than `if`-based negation. The Bash tool escapes the
# exclamation mark to a literal backslash-exclamation, which prevents bash
# from recognizing the pipeline-negation reserved word and leaves the loop
# stuck until the 10-minute timeout.

grep -rn '\!' plugins/` after the fix returns zero matches.

Why this happened twice

PR #234 fixed the poison text but didn't add a regression guard. When PR #226 later restructured the surrounding comment (legitimately, for the $GITHUB_WORKFLOW → $GITHUB_RUN_ID filter change), the author rewrote the whole block and put the poisoned phrasing back in. A follow-up PR to add grep -rn '\!' plugins/to the lint step (orpre-commit` hook) would prevent a third reintroduction. Leaving that out of this PR to stay atomic.

Test plan

• Lint, test, and review checks pass on this PR
• Bot's self-review session on this PR executes with an assistant session longer than 5 lines (proving the preprocessor poison is gone)
• grep -rn '\!' plugins/` returns no matches after merge
• First post-merge tend-review run on an unrelated PR posts an actual review

[tend-agent]

Closing as a duplicate of #243. Both PRs were opened from concurrent matrix jobs of the same review-reviewers run (GITHUB_RUN_ID 24287235173 — #243 from the max-sixty/worktrunk job, this PR from the max-sixty/tend job) and carry functionally identical fixes for the same PR #226 regression. #243 was created first (17:11:25Z vs 17:13:42Z) and its CI is already green, so deferring to it.

The max-sixty/tend-side evidence (3 runs on PRs #240/#241/#242, all 5-line traces, all zero reviews) is recorded in the review-reviewers-tracking comment on #133.