Job 98194fa8d215 showed every extract call hitting the 32k cap with only
~20k chars visible despite reasoning effort=low - Gemini 2.5 Pro still
burned ~25k thinking tokens per sheet.
- EXTRACT_MAX_TOKENS default 32768 -> 65536 (model output ceiling)
- new EXTRACT_REASONING_MAX_TOKENS (default 2048): OpenRouter reasoning
max_tokens / Gemini thinking_budget; takes precedence over effort
- log per-call reasoning token counts (usage.completion_tokens_details)
and include thinking count in the finish_reason=length marker
- llm.py: set_model_overrides(vision, text) replaces the single job override;
UI picks still beat per-call agent model args, but never name the hybrid
local model (avoids main's hybrid footgun); local->cloud fallback uses the
text pick.
- jobs.py: timestamped line-split tee (job_log.py), in-memory log + log_tail
polls, full log on terminal states (done/error/needs_review/finalization_error),
log-only disk recovery, error email links to the run log, and failed runs now
append the full traceback to job.log. Keeps pipeline_mode, job.json, and the
review gate.
- models.py: vision/text split via architecture modalities, pricing kept;
/models returns {vision, text, defaults}; /check takes vision_model/text_model
(replacing model); /health adds text_model. models_catalog.py dropped.
- UI: two priced dropdowns (OpenRouter compute only) + live run-log panel.
- Tests updated for dual overrides and the /models shape; new coverage for
traceback capture and local-model immunity.
- Job logs: each job's stdout/stderr is teed into outputs/<id>/job.log
(survives restarts) and served at GET /jobs/{id}/log as text/plain, so
full run logs can be shared for debugging and refinement.
- Model picker: GET /models proxies OpenRouter's public model list with
per-1M-token pricing (1h cache, 502 on failure); the UI shows a model
dropdown with costs when OpenRouter compute is selected, and the pick
overrides vision+text models for that job (Classic and Agent modes).
- Conflicts in the report view are grouped by discipline pair
(collapsible sections, severity-ordered within groups) instead of one
flat severity-only list.
APP_BASE_URL default (config, .env.example, both compose files) is now
https://conchecker.scoutitsystems.com with no port, so review-required
and final-report email links use the public site. CI bakes the short
commit SHA into the image as APP_BUILD via a Docker build-arg; /health
returns version+build and the site header shows the build so it's easy
to confirm which image is deployed. Local runs default to 'dev'.
Agent web jobs now stop after Brain consolidation and enter needs_review
with a persisted review queue (blocking: high-severity, low-confidence,
sensitive-category findings; audit sample of clean clusters). Humans
decide confirm/reject/unsure/needs_clarification via new review API and
frontend queue; a finalizer applies decisions (rejections suppressed with
reason codes), performs bounded targeted reruns for clarifications,
drafts RFIs only for kept issues, and only then marks the job done and
sends the final email. Two-phase email (review-required, then final
report), per-decision feedback labels with redacted aggregate metrics,
restart recovery from job artifacts, and CLI --no-review bypass.
Classic pipeline unchanged. 65 non-LLM tests.
woogi
John Wilganowski