emrg: open-source task must never stash/reset host's uncommitted work — dirty tree runs read-only (rant 2026-08-20T11:58:27)

[argszero]

open-source task must never stash/reset the host's uncommitted work (rant 2026-08-20T11:58:27)

Bug (host data-loss report)

The open-source task type (emrg/server/open_source_prompt.md) ran on the host's live working directory (e.g. aitokenpool-opensource-task on the aitokenpool repo) and its §0.3 "Source sync" phase instructed agents to git stash the host's uncommitted local changes. This silently hid — and effectively lost — the host's live edits twice on 2026-08-20 11:15-11:20 (files reset to HEAD with no reflog trace). The host had to disable the task (~/.emrg/tasks.yml enabled: false) to protect their work.

Fix

• §0.3 Source sync rewritten: a dirty working tree is now treated as NORMAL (the source dir is the host's working directory, not a dedicated clone):
  • Never run git stash / git checkout . / git restore . / git clean / git reset --hard or anything that hides/discards uncommitted changes
  • Never create branches / commit / push / open PRs while the tree is dirty
  • Dirty tree → the cycle runs read-only (scan / review / issue discussion / state-file update), records "dirty working tree — read-only cycle", and finishes without any git write
  • git pull --rebase only when the tree is clean; if dirty + behind → skip the pull
• Error-handling table: git pull conflicts now use git rebase --abort (the tree was clean before the pull, so abort restores it) instead of stashing host work
• Regression testtest_open_source_template_never_stashes_host_work asserts the rendered prompt contains the safety rules and no longer instructs git stash

Files

• emrg/server/open_source_prompt.md — §0.3 + error-handling row
• tests/test_scheduler.py — new regression test
• Agent.md — test count 987 → 988 (doc-count guard)

Verification

• uv run pytest tests/ -q → 987 passed, 1 skipped
• import + CLI checks OK

[argszero]

✅ LGTM — cycle (1/3)

Host data-loss fix (rant 2026-08-20T11:58:27): the open-source task's source-sync phase used to git stash the host's uncommitted work in the live working directory — silently losing edits (no reflog trace). Rewritten:

• dirty working tree → read-only cycle (never stash/checkout ./restore ./reset --hard; no branch/commit/push/PR while dirty)
• git pull --rebase gated on a clean tree; dirty + behind → skip
• pull conflicts → git rebase --abort (tree was clean pre-pull) instead of stashing
• regression test asserts the rendered prompt contains the safety rules

CI: test + test-windows PASS (32330596704). pytest 987 passed + 1 skipped locally.

[argszero]

✅ LGTM — cycle (2/3)

Fresh-cycle re-review of the full diff (head 25f4785, 3 files +58/−5):

• open_source_prompt.md §0.3 rewritten: dirty working tree is now explicitly NORMAL (source dir = host's working dir, not a dedicated clone) → cycle runs read-only. Prohibitions cover stash/checkout ./restore ./clean/reset --hard; no branch/commit/push/PR while dirty. pull --rebase gated on clean tree; dirty+behind → skip.
• Error-handling table: pull conflicts → git rebase --abort (tree was clean pre-pull, abort restores it) — never stashes host work.
• Regression test test_open_source_template_never_stashes_host_work asserts the rendered prompt contains the safety rules and no longer instructs stash; Agent.md count synced (987→988) for the doc-count guard.

CI: test + test-windows PASS (32330596704). Local pytest 987 passed + 1 skipped.

[argszero]

✅ LGTM — cycle (3/3)

Fresh-cycle re-verification: head 25f4785 unchanged since 2/3; mergeState CLEAN; test + test-windows SUCCESS (32330596704). The safety fix is complete and correct — open-source task now treats the host's dirty working tree as normal and runs read-only instead of stashing/resetting host work. Merging.