fix(compression): break the infinite compression loop (salvage of #31730)

[teknium1]

Summary

Compression no longer infinite-loops. Salvage of #31730 (@someaka) with a polish + two invariant-pinning tests on top.

Root cause (from #29335)

Three coupled defects cause an infinite compression loop:

1. _compress_context() rolls the agent forward into a new session, but run_conversation() never put session_id in its result dict — the gateway's guard for "did the session change?" was dead code.
2. Preflight compression compared _preflight_tokens >= threshold_tokens directly, bypassing should_compress() which carries the anti-thrash logic. When the system prompt + tool schemas dominate token count, compression saves <10% but preflight re-fires every turn anyway.
3. Gateway updated session_entry.session_id in memory after compression but never called session_store._save() — the two sibling sites in the same file (hygiene rewrite at L8437, manual /compress at L12122) already persisted; this third site was missed.

Changes

• agent/conversation_loop.py: add "session_id": agent.session_id to the result dict; gate preflight on agent.context_compressor.should_compress(_preflight_tokens) instead of a raw threshold compare. (Contributor commit, authorship preserved.)
• agent/conversation_loop.py (polish): drop the redundant _preflight_tokens >= threshold_tokens clause the contributor PR added next to the new should_compress() call — should_compress() already short-circuits when tokens < threshold_tokens, so the explicit comparison was dead on the True branch.
• gateway/run.py: call self.session_store._save() after the post-compression session_entry.session_id = ... assignment. (Contributor commit.)
• tests/run_agent/test_413_compression.py: one new test pinning that preflight calls should_compress() so anti-thrash has a vote — mocks should_compress to return False even with tokens past the raw threshold and asserts no compression runs.
• tests/gateway/test_compression_session_id_persistence.py: new file, one test. AST scan over gateway/run.py asserts every session_entry.session_id = ... assignment is followed by a session_store._save() call within the same block. Empirically verified the test catches the bug: drop the new _save() line → test reports line 8694 missing _save(), restore it → green.
• scripts/release.py: AUTHOR_MAP entry mapping ed@bebop.crew → someaka.

Validation

	Before	After
200k-token session resume	infinite compress loop	one compress, normal
Anti-thrash on preflight	bypassed	applies
Gateway restart after compression	reverts to old session	survives

tests/run_agent/test_413_compression.py                        16 passed
tests/gateway/test_compression_session_id_persistence.py        1 passed
                                                              17 passed in 5.31s

Notes

• Salvage of fix(compression): 3-line fix for infinite compression loop (#29335) #31730 by @someaka. Contributor's commit is preserved with original authorship; the polish + tests are a separate follow-up so per-commit blame stays clean.
• Closes fix(compression): 3-line fix for infinite compression loop (#29335) #31730, fixes Bug: Context compression creates new session but gateway never sees it — infinite compression loop #29335.
• The AST-scan test pins the contract for all three session_entry.session_id assignments in gateway/run.py — adding a fourth site in the future without _save() will fail this test, not silently break compression a year from now.

Infographic

[github-actions]

🔎 Lint report: salvage/pr31730-compression-loop 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: 9137 on HEAD, 9136 on base (🆕 +1)

🆕 New issues (1):

Rule	Count
not-iterable	1

First entries

tests/gateway/test_compression_session_id_persistence.py:75: [not-iterable] not-iterable: Object of type `object` is not iterable

✅ Fixed issues: none

Unchanged: 4872 pre-existing issues carried over.

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

[alt-glitch]

Salvage of #31730 by @someaka. Competes with #29505 (same 3-part fix for #29335). Both address the infinite compression loop.

[talwayh1]

CI Test Fix: borrowed-source credential pruning

test_auth_remove_copilot_suppresses_all_variants fails because _prune_stale_seeded_entries() (from #31416) trims the gh_cli borrowed-source entry when resolve_copilot_token returns empty in CI.

Root cause: The test writes a gh_cli entry into auth.json, then calls auth_remove_command() which invokes load_pool() → _seed_from_singletons() → resolve_copilot_token() returns ("", "") → gh_cli not in active_sources → _prune_stale_seeded_entries() removes the entry → pool has 0 entries → resolve_target("1") raises SystemExit: No credential #1.

Fix: Added monkeypatch for resolve_copilot_token and get_copilot_api_token so the entry is re-seeded into active_sources and survives pruning.

Fix available at d32cad3fc on talwayh1/hermes-agent:salvage/pr31730-compression-loop:

git fetch https://github.com/talwayh1/hermes-agent.git salvage/pr31730-compression-loop
git cherry-pick d32cad3fc

Verified: pytest tests/hermes_cli/test_auth_commands.py::test_auth_remove_copilot_suppresses_all_variants -xvs → PASSED