fix: dispatch via run-agent.sh; strip inline-code from linkage scan

[johanzander]

Summary

• The backlog skill's Dispatch verb now routes through scripts/run-agent.sh instead of claude --bg. A headless dispatch runs in a disposable container with the dev-role token, a restricted egress allowlist, and permissions skipped inside the container — so it can never stop to ask for a permission nobody is there to answer (the old path's exact hang).
• The merged-PR linkage scan strips inline-code spans before matching. PR fix: link a PR to its issue on any #N reference, not just closing keywords (#652) #679's own body carried `- Blocked by #100 -- part of #409` as a backticked worked example, and that bounced issue Consolidate PredictionSnapshotStore + ScheduleStore into the unified per-day DailyView format #409 to In Verification for work it never did. A real linkage declaration is never in code markup. Genuine Part of #N / tracking #N intermediate PRs still keep their issue In Verification (existing behavior pinned by test_a_merged_intermediate_pr_keeps_the_issue_in_verification).

Test plan

• New regression test test_a_backticked_worked_example_does_not_flip_an_issue — fails on the pre-fix scan, passes now
• .venv/bin/pytest backend/tests/test_backlog_digest.py (69) + test_backlog_rhythm.py/test_backlog_board_init.py (77) — all pass
• black / ruff / mypy on changed files — clean
• Full fast suite: 2261 passed; the 6 test_quality_check_mypy_gate.py failures are sandbox artifacts (pass unsandboxed — the ratchet's git archive|tar runs under /var/folders)
• The quality gate's one remaining error is environmental, not this PR: Black flags untracked scratch_*.py in .claude/worktrees/bench-pwl-everywhere/ (CI checks a clean checkout)

[bess-agent]

@claude-bot review

[bess-manager-claude-bot]

Review

Root-cause match: partial. The PR's stated root cause — "A real linkage declaration is never in code markup" — is correct for the merged-PR scan it patches (scripts/backlog-digest.sh:133, refs: [ (.body // "") | gsub("\[^\`]*`"; "") | scan(...) ]), and matches the actual offending text in merged PR #679's body (- Blocked by #100 -- part of #409inside a backtick span). Verified withgh pr view 679 --json body`.

But the identical bug is left live in the open-PR linkage path, linkage_body/pr_matches_issue (scripts/backlog-digest.sh:253-259):

def linkage_body($body):
  ($body // "")
  | gsub("(?i)(blocked by|depends on|unblocks?(?:ing)?|related to|relationship to|follow[- ]?up to|see also|see|not part of) #[0-9]+"; "");

def pr_matches_issue($p; $n):
  (linkage_body($p.body) | test("(?i)(^|[^0-9])#\($n)\\b"))

No backtick-stripping is applied here, only cross-ref phrase stripping — and part of #N is deliberately not in the stripped-phrase list (it's a legitimate linkage phrase for real PRs). So a code-quoted worked example containing part of #N still links.

This isn't hypothetical — it's live on this PR right now. I ran linkage_body/pr_matches_issue verbatim against PR #684's own current body (which quotes `- Blocked by #100 -- part of #409` while explaining the fix, same as #679 did):

$ jq -n --rawfile body <(gh pr view 684 --json body --jq .body) -f test.jq
{ "matches_409": true, "matches_679": true }

While this PR is open, backlog-digest.sh's open-PR scan links it to issue #409 (and #679) and would place #409 in "In Review" — the exact incident class this PR exists to fix, reproduced by this PR's own description, in the code path the PR didn't touch.

Test coverage: test_a_backticked_worked_example_does_not_flip_an_issue (backend/tests/test_backlog_digest.py) only exercises the merged-PR scan. It would not catch a regression in pr_matches_issue/linkage_body for open PRs — no test constructs an open PR with a backtick-quoted linkage phrase.

Scope: the PR bundles two unrelated fixes — the SKILL.md run-agent.sh dispatch rewiring and the digest regex fix — described separately and each internally minimal, so not a blocker on its own, but worth calling out.

Verdict: REQUEST_CHANGES. Please apply the same code-markup stripping to linkage_body (or a shared helper both scans call) so the fix covers both linkage paths, and add a test that pins an open PR with a backtick-quoted part of #N example not linking. No hard rule violations found otherwise (no Optional, hasattr, or silent-fallback issues — this is bash/jq, not Python).