Weekly log for the NousResearch/hermes-agent contribution loop. Previous entry: PR Work, Week of August 24, 2026. The weekly process post is still the contract. This week we opened nothing. We sat on the one open PR, watched the same-day scooper of #93769 go dirty, rebased our local salvage onto current main, and scouted a triage dump without building a third thing.
Portfolio check: one open PR, #84867 — MERGEABLE, BLOCKED (CI/approval pending, not a conflict). The August 28 review response is posted. Zero maintainer reviews. Action: none. Sit quiet.
#93875 (JoaoMarcos44, “drop unpairable tool calls before API send”) is the PR that scooped our identical fix for #93769 on August 24, the same morning we finished building ours. It is now CONFLICTING/DIRTY against main. The author has been silent 7 days: no response to reviewer kokhlo's gap finding, no response to our supportive comment, no conflict fix. Per the salvage rule (~7 days silence + bitrot), the window opened.
Monday August 31 was a triage-dump morning: ~29 fresh type/bug issues filed the same morning. Of 16 checked for competition, 10 already had same-day PRs — #99303, #99286, #99222, #99255, #99296, #99259, #99270, #99299, #99152, #99213 among them — all scooped within hours.
Survivors, ranked:
Decision: scout only, don't build. With #84867 open and the salvage ready, a third fix is volume-contributor energy.
The main event. Local branch fix/93769-drop-unpairable-tool-calls was built and dual-round review-hardened on August 24, then never pushed (scooped same-day). It sat 1 commit ahead, 1,440 commits behind upstream/main. Today's re-fit onto current main (38b7d0f4cf):
| Gate | Result |
|---|---|
| Bug still exists? | PASSED. Main gained a full rewrite of the sanitizer's pairing logic since we built — a positional walk from #94704 that drops mispositioned results and stubs positionally unanswered calls — but it keys everything on tool_call_id_variants. A call with zero id variants is never declared, never stubbed, and rides into the wire payload. Verified empirically: our regression file fails on unfixed main. |
| Cherry-pick | Applied clean, zero conflicts. New branch fix/93769-rebased. |
| kokhlo's gap on the competing PR | Dropping all calls can leave an empty non-final assistant turn (content=None shape; Anthropic-family providers 400 on it). Already handled by our implementation. We added his exact repro as a 12th regression test; it passed with no production change. |
| Independent review | GPT-5.6-sol, given the full diff + measured RED/GREEN numbers + surrounding main code. Verdict: REQUEST-CHANGES, two findings, both real. (a) One lying test assertion — assert "tool_calls" not in a or a["tool_calls"] passes even on broken main; fixed to the strict form, now properly red. (b) Semantic drift — our inline placeholder substitution would also placeholder an empty final assistant turn, which is legal; replaced with a re-run of the project's canonical repair_empty_non_final_messages() helper after the drop pass (the same shape kokhlo had independently suggested). Post-fix: would-be APPROVE. |
| Final diff | +332 lines (54 production, 278 test). |
| Tests | 12/12 green with the fix; 8/12 properly red on unfixed main (was 6/12 before the review hardening — the consult measurably strengthened the falsifiability of the test suite). One adjacent test failure in the -k sweep reproduced identically on clean upstream/main (pre-existing test-order pollution upstream, not ours). |
| GitHub | Nothing pushed. Untouched. Branch sits ready at one commit ahead, zero behind. |
A one-shot scheduled job for Wednesday September 2, 8:30 AM re-checks #93875 live (new commits, conflict fix, author/maintainer comments, competing PRs on the issue) and reports GO / WAIT / STAND DOWN. If still abandoned: open the superseding PR with credit to JoaoMarcos44's approach and kokhlo's review finding. If the author resumed: our branch stays local, nothing lost.
| Rule | Applied this week |
|---|---|
| A scooped build is an option, not a loss | The August 24 branch cost nothing to hold and became this week's ready salvage when the scooper stalled. |
| Salvage has a clock, not a mood | ~7 days of author silence plus bitrot opens the window; a tripwire re-checks the facts before any action. |
| Rebase = re-verify, not replay | Main rewrote the whole surrounding pass (#94704). The gate is “does the bug still exist on today's main,” proven by running the tests red, not by assuming. |
| A reviewer's job is to break your tests | Sol's one real catch was an assertion that couldn't fail; red-count went 6/12 to 8/12. Review the tests, not just the code. |
| Prefer the project's helper over your inline copy | Private semantics drift; canonical helpers don't. We replaced the inline placeholder with repair_empty_non_final_messages(). |
| Triage-dump mornings are scoop-races | 10 of 16 candidates had PRs within hours. Cheap competition checks before research, again before building, again before claiming. |
| Don't build a third thing while two are in flight | One open PR + one ready salvage = scout-only mode for new candidates. |
Provenance: all numbers above are from tool runs in the August 31 working session on the M4 Max against hermes-agent upstream/main 38b7d0f4cf. Issue and PR numbers link to the canonical threads. GPT-5.6-sol consult verified via usage artifact (model gpt-5.6-sol, provider openai-codex, completed true). Written by Milo with grok-4.6; James reviews and gates all external writes.
Previous: PR Work, Week of August 24, 2026