The outcome columns on an execution row (prs_created, issues_created, commits) are derived by substring-matching the command text of Bash tool calls. That counts mentions, not creations — so the numbers on squads board (and every surface reading the ledger) can be pure fiction.
Repro from a real run (2026-07-15)
A lane finished with prs_created: 2, issues_created: 1 on its ledger row. It created zero PRs and zero issues — its git credentials had expired mid-run (#1133), so every push and gh call failed with Bad credentials. The PR that eventually landed its work was opened by hand, hours later.
Where the phantom "2 PRs" came from, verbatim from the transcript:
grep -rln "tool_use\|gh pr create\|git commit" --include="*.py" . — a search for the string.git commit -m "…" whose message documents the patterns gh pr create / gh issue create / git commit.
The lane was implementing the very feature that counts PR creations, so the counter counted its own source code and commit message.
Why it matters
squads board reports "N PRs created · today" as a headline number. Today it says 2; the true count from that run is 0. A metric that is silently wrong is worse than one that is absent — it launders an intent (or a mere mention) into a result.- Failure makes it worse, not better: a lane whose pushes all fail still increments the counter, so the number climbs precisely when nothing is shipping.
- This is the ledger-side capture; the API-side derivation for harness sessions shares the same approach and needs the same fix.
Fix direction
Count the result, not the command. gh pr create prints the PR URL on success and exits non-zero on failure — so parse the tool_result for https://…/pull/<n> (likewise /issues/<n>), and take git commit from the resulting SHA rather than the invocation. That yields verifiable identifiers instead of a guess, makes the numbers dedupable across sources, and drops to 0 when nothing shipped.
If a verified count isn't available for a source, render nothing — not a zero, and not an attempt count dressed as a result.
Acceptance
- A run whose
gh pr create fails (bad credentials, CI reject, anything) reports 0 PRs.
- A run that greps for, echoes, or commits the literal text
gh pr create reports 0 PRs.
- A run that actually opens 2 PRs reports 2, and the row carries their numbers/URLs.
The outcome columns on an execution row (
prs_created,issues_created,commits) are derived by substring-matching the command text of Bash tool calls. That counts mentions, not creations — so the numbers onsquads board(and every surface reading the ledger) can be pure fiction.Repro from a real run (2026-07-15)
A lane finished with
prs_created: 2, issues_created: 1on its ledger row. It created zero PRs and zero issues — its git credentials had expired mid-run (#1133), so every push andghcall failed withBad credentials. The PR that eventually landed its work was opened by hand, hours later.Where the phantom "2 PRs" came from, verbatim from the transcript:
grep -rln "tool_use\|gh pr create\|git commit" --include="*.py" .— a search for the string.git commit -m "…"whose message documents the patternsgh pr create/gh issue create/git commit.The lane was implementing the very feature that counts PR creations, so the counter counted its own source code and commit message.
Why it matters
squads boardreports "N PRs created · today" as a headline number. Today it says 2; the true count from that run is 0. A metric that is silently wrong is worse than one that is absent — it launders an intent (or a mere mention) into a result.Fix direction
Count the result, not the command.
gh pr createprints the PR URL on success and exits non-zero on failure — so parse the tool_result forhttps://…/pull/<n>(likewise/issues/<n>), and takegit commitfrom the resulting SHA rather than the invocation. That yields verifiable identifiers instead of a guess, makes the numbers dedupable across sources, and drops to 0 when nothing shipped.If a verified count isn't available for a source, render nothing — not a zero, and not an attempt count dressed as a result.
Acceptance
gh pr createfails (bad credentials, CI reject, anything) reports 0 PRs.gh pr createreports 0 PRs.