jpmschweitzer/odysseus
1
0
Fork 0
mirror of https://github.com/pewdiepie-archdaemon/odysseus.git synced 2026-10-07 07:22:21 +02:00
Code Issues Packages Projects Releases Wiki Activity

Record confirmed named-search failure and pending deployment

Browse Source
This commit is contained in:
pewdiepie-archdaemon
2026-09-17 21:31:09 +00:00
parent a80a090d37
commit 6be1db32de
1 changed files with 2 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/search-quality-audit-20260917.md
+2
View File
@@ -75,6 +75,8 @@ Further provider inspection found that the news-to-general fallback dropped the
### Additional informal/multi-part live checks
Canonical-system confirmation `reports/search-tool-choice-probe-1789680586200.json`: auto and required supplied queries for both prompts; named search choice omitted query in both (and typo prompt emitted `command`). The same compact schema and model were used. Implemented forced-search dispatch as one offered web_search schema with required choice, preserving the forced-tool intent and original schema. Other tool choices remain unchanged. 1,230 routing/runtime regressions pass. **Not deployed yet:** the pre-change 23-conversation sweep remains active (nine conversations complete at this checkpoint); wait for its terminal state before restart and paired replay. This is a demonstrated argument-generation difference, not yet an end-to-end quality/speed win.
Full 23-conversation regression launched on `6000b718`/current deployed harness: `reports/clean-v3-search-quality-2026-09-17T21-27-27-686Z.json`. Active handle recorded in session; do not restart based on elapsed observation time.
Read-only tool-choice control `reports/search-tool-choice-probe-1789680509169.json` uses the exact compact web_search schema, a short system prompt, identical user prompts/temperature/model, and never executes emitted calls. For both weekly-news and typo-news prompts, auto/required emitted nonempty queries. Forced named mode emitted an extraneous `command` field in both; typo-news omitted query entirely. Six calls are preliminary evidence of tool-choice/schema behavior, not proof of a universal backend defect or a production fix. Next test should use the canonical harness system/history before changing dispatch. Probe script saves full schemas and public emitted calls for reproducibility.