Bug report: batch runner issues observed in multi-item parallel run (2026-03-17)

[Sillhouette]

Context

Observed during a real project run: one spike (--through design), one parallel batch of 3 items (--parallel 3 --finish-mode --pr). All issues are in the genies batch runner or genie-session.

---

Issue 1 — Discern verdict parse failure strands work with no PR

What happened:
/discern ran and completed (22 turns, implementation reviewed), but the batch runner exited with code 1:

[discern] Completed (22 turns, 7248 tokens, ...)
[ERROR] Could not parse verdict from /discern output — stopping (safe default)

The branch had all commits, the implementation was done, but the script halted before commit / push / gh pr create. No PR was created. No recovery path was printed.

Impact: Work was stranded in the worktree. Manual push + gh pr create required to recover.

Likely cause: The critic genie's output format doesn't consistently place APPROVED or REJECTED at the expected position/line. The parser matches too narrowly.

Suggested fix:

• Widen the verdict regex (search the full output, not just the last N lines)
• On parse failure, print the last 20 lines of discern output to the log so it's diagnosable
• Consider a recovery action: if verdict can't be parsed, push the branch and print the manual PR URL rather than silently stopping

---

Issue 2 — Parallel runs with duplicate slugs share a log file

What happened:
Two backlog items had the same filename in different directories:

• mobile/P3-brand-token-wiring.md
• admin/P3-brand-token-wiring.md

Both resolve to slug p3-brand-token-wiring. The script correctly disambiguated worktree and branch names (-admin suffix for the second), but both workers wrote to the same log file:

logs/batch-20260317-072008/p3-brand-token-wiring.log

Both runs were reported as [ERROR] Failed (exit 1) pointing to the same log. The log contained output from only one of the two runs, making it impossible to diagnose the second failure.

Suggested fix:

• Log file name should be derived from the disambiguated slug (same logic used for worktree/branch naming), not the raw slug
• Or: include the input path's directory component in the log filename (mobile-p3-brand-token-wiring.log, admin-p3-brand-token-wiring.log)

---

Issue 3 — genies status requires explicit log directory

What happened:
With a batch actively running, genies status returned an error because no log directory was provided. The log directory was visible in the running process args (logs/batch-20260317-072008) but the status subcommand couldn't auto-detect it.

Suggested fix:

• genies status should detect the most recent (or currently active) log directory automatically — e.g. by reading the most recently modified directory under logs/, or by checking for a lockfile/pidfile written at batch start
• Or: at batch start, write the log dir path to a known location (e.g. .genies-current-batch) that genies status can read

---

Issue 4 — --through <phase> silently merges to main with no indication it skipped PR creation

What happened:
Running genies docs/backlog/mobile/P2-spike-drop-ionic-ui.md --from discover --through design produced:

[session] Merged and cleaned up: p2-spike-drop-ionic-ui
[INFO]  PDLC completed: discover → design

No PR was created (correct — it's docs-only). But the output gave no indication that this was intentional. The user expected a PR and assumed the run had failed silently.

Suggested fix:

• When --through stops before the PR phase, print an explicit note:

  [INFO]  --through design: stopping before PR phase. Changes merged to main directly.
  

• Or: distinguish docs-only phases (discover/define/design) in the summary output so it's clear no PR is expected

---

Related backlog item

docs/backlog/P3-genies-script-hardening.md covers the sed delimiter injection and undocumented exit codes — separate issues, already tracked.

[Sillhouette]

Root cause found for Issue 1 (discern verdict parse failure)

The critic genie called /done internally during the /discern phase, which archived the backlog item before detect_verdict ran.

Sequence:

1. /discern runs — critic approves, writes verdict, then calls /done → archives docs/backlog/mobile/P3-brand-token-wiring.md to docs/archive/
2. Commit lands in worktree: chore(docs): archive P3-brand-token-wiring — DONE
3. detect_verdict checks docs/backlog/mobile/P3-brand-token-wiring.md — file no longer exists at that path
4. All three detection methods fail (frontmatter, output text, item body) → [ERROR] Could not parse verdict

Evidence: the worktree's git log contains the archive commit (docs/archive/mobile/2026-03-17_brand-token-wiring/P3-brand-token-wiring.md) timestamped during the discern phase, before the runner's commit/done steps would have run.

Root fix: The /discern skill should not call /done. It should write a verdict to the backlog item frontmatter and stop. The runner script owns the commit/done lifecycle. The critic overstepping caused both the verdict detection failure and a duplicate archive operation.

Defensive fix (belt and suspenders): detect_verdict should also check the archive path (docs/archive/**/{item-filename}) as a fourth fallback, since this failure mode can recur if the skill boundary isn't enforced.

[Sillhouette]

Closing this in favour of smaller, clearly scoped issues. Summary of what was here and where it went:

Issue 1 — Discern verdict parse failure (root cause confirmed): The bug is that /discern calls /done internally in headless mode, archiving the backlog item before detect_verdict can read the verdict frontmatter. Root cause is a missing boundary contract between interactive skills and the headless runner's phase units. Tracked in a new issue.

Issue 2 — Duplicate slugs share a log file: Both mobile and admin items named P3-brand-token-wiring.md produced the same slug, so both workers wrote to the same log path. The disambiguated worktree/branch naming didn't carry through to the log filename. Tracked in a new issue.

Issue 3 — genies status requires explicit log dir: Covered by PR #16.

Issue 4 — --through silently skips PR creation: No output distinguishes an intentional docs-only merge from a silent failure. Tracked in a new issue.

Discovery output: docs/analysis/20260317_discover_pdlc-skill-boundary.md