agent: coop prompt told agents the shared git remote was read-only

[akhatua2]

The bug

mini_swe_agent_v2's coop prompt titled its git section "Shared Git Remote (read-only)", listed only fetch/log/diff under "Allowed (read-only)", never mentioned push, and instructed "Do not merge, pull, or rebase their branch into yours".

Both claims are wrong:

• GitConnector runs git daemon --enable=receive-pack — the remote accepts writes.
• Grading never reads those branches. The evaluator applies each agent's submitted patch.txt to branches it creates itself (eval/sandbox.py:473,480).

Agents run in separate containers, so a push is the only way code can cross between them.

Evidence

Measured across 117 trajectories in three flash_10 runs:

	count
git fetch team	292
git push team	0

Every one of those fetches returned the untouched baseline, because 0.0.22 only publishes an agent's patch at exit — by which point the peer can no longer act on it. Agents were following their instructions exactly; the instructions disabled the channel.

This prompt was the only surface that said so. GitConnector's docstring already advertised git push team <branch> and git merge team/<agent>, and agents/_coop/prompt.py (used by the claude_code / codex adapters) already tells agents to "push so peers can fetch you". Only this config disagreed with the infrastructure it runs on.

The fix

The section now describes a read/write remote and tells agents to commit and push as they work, so a colleague can read their actual diff.

The blanket prohibition is replaced with the one constraint that genuinely affects grading: the submitted patch.txt must contain only that agent's own changes, since the two submitted patches are merged and duplicated edits break the merge. Local fetch/merge/cherry-pick are explicitly allowed — the rule is about what you submit, not what you do in your worktree — with git diff -- <paths> given as the way to scope a patch when both agents had to touch the same file.

The rewritten section is shorter than the one it replaces (1210 vs 1713 chars), and it survives context compaction, which preserves the task message verbatim (prefix = self.messages[:2]).

Not addressed here

swe_agent/config/coop.yaml has the same omission — it lists fetch/log/diff/dry-run and never mentions push. Left alone as a forked third-party agent; happy to include it if wanted.

Verification

• uv run pytest tests/ — 403 passed, 63 skipped
• ruff check / ruff format --check clean
• YAML parses and the Jinja renders correctly for both agent ids

Whether agents actually use the channel is empirical and worth confirming with one flash_10 run counting git push team — but 0 pushes was guaranteed while the prompt said read-only.

Version bumped to 0.0.23 with a CHANGELOG entry.