fix(compression): disable compression on background-review fork (#38727)

[teknium1]

Summary

The background-review fork can no longer fork the parent session it shares a session_id with — closing the cross-turn gap left by the per-session compression lock (#38727).

Root cause: the per-session lock only serialises a same-window concurrent compression race. If review_agent won that race, its new child session was never adopted by the gateway (the fork is single-lifecycle — one run_conversation, then dies). The next foreground turn started from the stale parent and compressed it again, leaving the same parent with two sibling children.

Changes

• agent/background_review.py: set review_agent.compression_enabled = False on the fork. Both compression trigger sites in conversation_loop.py (preflight + mid-loop) gate on compression_enabled before calling _compress_context, so a fork with the flag cleared never reaches the rotation path. Review needs full context anyway — compressing would degrade the memory/skill summary.
• tests/agent/test_compression_concurrent_fork.py: regression test driving the real _run_review_in_thread worker and asserting the fork inherits the parent session_id but has compression_enabled == False.

The per-session lock is kept as defense-in-depth for any future shared-session path.

Validation

	Without fix	With fix
new regression test	FAILS (fork compresses → orphan child)	PASSES
full file (4 tests)	—	4 passed

Implements the fix proposed by @lgYanami in #38727.

Closes #38727

Infographic

[liuhao1024]

Positive verification — clean security review

Reviewed the diff, test, and rationale for the background-review fork compression guard. The fix is correct and well-scoped:

1. Root cause is clearly documented — the fork shares the parent's session_id; winning a compression race would rotate the parent into an orphan child that the gateway never adopts, then the foreground turn compresses the stale parent again, producing two sibling children.

2. Fix is minimal and sufficient — setting review_agent.compression_enabled = False short-circuits both compression trigger sites in conversation_loop.py (preflight and mid-loop), both of which gate on this flag before calling _compress_context.

3. Test pins the contract at the source — _run_review_in_thread is called synchronously with AIAgent.run_conversation patched (no LLM call), the constructed agent is captured, and the assertion checks compression_enabled is False with a clear regression message.

4. No side effects — the review fork still inherits session_id and session_start for logging correlation; only the compression flag is cleared.

No issues found.

[github-actions]

🔎 Lint report: fix/compression-review-fork-38727 vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 10032 on HEAD, 10031 on base (🆕 +1)

🆕 New issues (1):

Rule	Count
unresolved-attribute	1

First entries

agent/background_review.py:462: [unresolved-attribute] unresolved-attribute: Unresolved attribute `compression_enabled` on type `AIAgent`

✅ Fixed issues: none

Unchanged: 5200 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.