Prevent session logs from stranding: backstop warning, sweep, hygiene rule

[goldingn]

🤖 Posted by Claude (Opus 4.8), an AI coding agent, from the @goldingn account — not written by a human. Why.

Fixes #58. The prevention half of the session-log stranding problem; #57 committed the logs that had already stranded.

Root cause. The Stop hook writes a per-branch log every turn but never commits it — that's folded into a PR. Turns on main (merges, triage) and post-merge wrap-up turns have no PR to carry their log, so it strands untracked.

Three complementary guards:

1. Backstop warning (log-session.sh). After writing its log, the hook warns to stderr about any other branch's uncommitted log — the leak signal — and stays quiet about the current branch's own log (expected uncommitted until swept), so it doesn't nag every turn. It never fails the hook. Uses git status --untracked-files=all so individual files are listed even when dev/sessions/ has no tracked file yet (that collapse is what a first draft got wrong — now covered by a test).
2. Sweep helper (sweep-session-logs.sh, new). One command to commit outstanding logs onto the current branch; stages only log files, so a stray edit elsewhere is never swept in. Idempotent.
3. Hygiene rule (CLAUDE.md). A "keep them off the floor" section: minimise time on main, commit the branch log before ready, sweep what strands, never hand-edit a log. Plus a Definition-of-Done item and a gotcha. Also corrects the stale "CI fails the PR if it is missing" line — that check is advisory since Make the session-log CI check advisory, not failing #53.

Tests. test-log-session.sh gains sections 7–8 (warns about another branch's stranded log; quiet once committed). 24/24 pass, run in CI by hooks-test.yaml. This PR is already dogfooding the backstop — the hook is warning about other branches' logs sitting in my tree.

How this was produced

• Issue: Prevent session logs from stranding (main-branch and post-merge turns) #58 (follows Session logs for main-branch and post-merge turns get orphaned (never committed) #56/Commit orphaned per-branch session logs #57)
• Core prompt(s): "a recommendation for how to prevent this workflow glitch going forward", then a request to consider keeping agents off main and to open the issue + PR.
• Key decisions / trade-offs:
  • Chose an advisory stderr warning + manual sweep over auto-committing in the hook, which would commit to main directly (against our rules) or create unbid commits at awkward times.
  • The backstop excludes the current branch's own log to avoid nagging every turn; test-first surfaced the untracked-directory collapse, fixed with --untracked-files=all.
  • An alternative approach that was weighed and set aside is recorded in Prevent session logs from stranding (main-branch and post-merge turns) #58 (kept there rather than here, since it is a decision record, not part of this diff).
• New dependencies: none (pure bash + git + jq, already required).
• Session log: committed on this branch before marking ready.

[goldingn]

One change to make. Dwell branch idea is also not relevant in this PR description.

[goldingn]

Redundant information about something that's different exist and has never been mentioned to the agent. Amend.

[goldingn]

🤖 Posted by Claude (Opus 4.8), an AI coding agent, from the @goldingn account — not written by a human. Why.

Fixed in 5cd717c — removed the dwell branch parenthetical. Agreed: CLAUDE.md loads fresh each session and the concept was never introduced there, so mentioning-then-negating it was just noise. The considered-and-rejected rationale stays in #58. Also stripped the dwell-branch content from this PR description per your review.

[goldingn]

All good. Merge.