fix(anthropic): tool_use/tool_result adjacency + OAuth claude-code/ UA (salvage #52145 #51948)

[hashbender]

Summary

Batch salvage of two complementary agent/anthropic_adapter.py fixes, rebased onto current main. Both bugs confirmed still-present; disjoint regions, no conflict.

Fixes

NousResearch#52145 (@fsaad1984) — enforce tool_use/tool_result adjacency in _strip_orphaned_tool_blocks
The old logic collected tool_result ids across all user messages and kept any assistant tool_use whose id appeared anywhere. Anthropic requires each tool_use to have its tool_result in the immediately following user message — a stale match elsewhere in the transcript would keep a genuinely-orphaned tool_use and 400 the request. Rewritten to adjacency-checked two-pass logic: pass 1 strips a tool_use whose result isn't in the next user message; pass 2 removes results orphaned by pass 1.

NousResearch#51948 (@DhivinX) — OAuth UA claude-cli/ → claude-code/
Anthropic 404s the OAuth token endpoint for the claude-cli/ UA prefix (NousResearch#48534). Switched all three OAuth UA sites — build_anthropic_client (:820), refresh_anthropic_oauth_pure (:1048), run_hermes_oauth_login_pure (:1491) — to claude-code/.

Review follow-ups (mine)

A structured review of the salvage surfaced three test-quality issues, all fixed:

• fix(adapter): enforce tool_use/tool_result adjacency in _strip_orphaned_tool_blocks NousResearch/hermes-agent#52145 shipped no behavior test. Added test_strips_tool_use_when_result_not_immediately_adjacent (the exact non-adjacent case the old global match got wrong) + an adjacent-pair control. Mutation-checked: reverting to a global match fails the non-adjacent test.
• test_token_exchange_ua_prefix false-failed on a comment (naive whole-function substring). Scoped it to actual User-Agent header lines, matching its sibling.
• test_token_refresh_ua_prefix was vacuous — it bound to a wrapper (_refresh_oauth_token) with no HTTP call, so its assert never ran. Retargeted at the real refresh site refresh_anthropic_oauth_pure. Mutation-checked: reverting :1048 now fails it.

Verification

• All UA + adjacency + existing-orphan tests pass (10/10 for the touched surface).
• Note: test_anthropic_adapter.py has ~15 pre-existing failures on current main in unrelated OAuth-token-resolution classes (TestResolveAnthropicToken / TestRefreshOauthToken / TestRunOauthSetupToken) — verified identical on clean origin/main with none of these changes applied. Not introduced by this PR.

Closes NousResearch#52145. Closes NousResearch#51948.

Co-authored-by: fsaad1984 38867992+fsaad1984@users.noreply.github.com
Co-authored-by: DhivinX 20087092+DhivinX@users.noreply.github.com

---

Mirror-of: NousResearch#56263
NousResearch#56263