Repository navigation
Conversation
…g-sol61-20261027 # Conflicts: # docs/organisation/systems/clinical-content.json # docs/organisation/systems/delivery.json
|
Important Draft PR not reviewedDraft PRs are not automatically reviewed by default.
To automatically review draft PRs, update your CodeRabbit configuration: reviews:
auto_review:
drafts: true
Comment |
|
Updates to Preview Branch (codex/rag-sol61-20261027) ↗︎
Tasks are run on every commit but only new migration files are pushed.
View logs for this Workflow Run ↗︎. |
|
Review record Branch: Outcome: Keep draft. Three P2 findings exist at the published head; all three are corrected and independently re-reviewed locally at Findings and local disposition:
Checks and evidence:
Remaining gates: control the automatic preview compute trigger before publishing the local commits; complete applicable exact integration/CI gates; establish actual account/model access and a feasible, fully priced, specifically approved comparison before activation. The existing 52-execution paid plan cannot fit its configured reservations. The private reduced proposal is not implemented/approved and its complete cost bound is unresolved. Zero paid OpenAI/model evaluation calls were made. Answer defaults, retrieval, embeddings and indexes remain unchanged; quality, latency and total cost improvements are unproven. Reviewed by Codex using the repository review workflow, with two independent reviewers requested as gpt-6.1-sol/high; actual routing telemetry was unavailable. |
|
Review record Generated by Claude Code |
|
Closing at the owner's request: superseded by #3900, which merged at 4a368ee. #3900 contains all of this PR's commits: three were cherry-picked, and the Edge indexing-model fix is included as file content. Of the 28 files here, 18 are byte-identical in #3900 and the other 10 are supersets. The blank-model fix for Node was also added in #3900. The branch is kept. Generated by Claude Code |
Summary
Prepare
gpt-6.1-solfor both existing answer routes while retaining low/medium reasoning effort. Runtime and example answer defaults remain on GPT-5.6 Terra/Sol until model comparison and owner acceptance; merging this preparation does not activate Sol.Keep classifier, summaries, indexing, vision and embeddings independent of answer-model changes. Align the Edge indexing agent with the Node indexing fallback so an answer-only override cannot discard valid structured summaries and generated labels. Preserve explicit indexing overrides in both environments. No schema migration, re-embedding or re-indexing is introduced.
Add server-only per-attempt generation accounting and a synthetic comparison harness. Unknown failed-call usage remains unknown cost. The component experiment reuses production generation, source fencing and validation; it does not measure full-route retrieval, planning, cache or recovery costs. It uses 12 cases, 48 measured executions and four symmetric route warm-ups, stops on failed/incomplete calls, and never automatically accepts a model.
Areas touched: Answer engine, Knowledge and records, Data platform and access, Source intake and indexing
RAG impact: no retrieval behaviour change — retrieval inputs, scoring, source admission, document permissions, embeddings, citation gates and final-only SSE remain unchanged. Edge model resolution preserves the intended structured indexing model when answer models change.
Verification
c4aa2b76dd2d14f83bb32c7d4a1974ce025450e2with a normal merge. Final candidate5d34d2a7d0a31c5a860dd8cbc20e49428c16b99f.5d34d2a7d0: compilation, TypeScript, static generation and client-bundle secret checks completed with exit 0, using separate 30-second admission and 20-minute execution limits (about 10m29s execution). The earlier ten-minute timeout is superseded by this completed evidence. Full typecheck, focused Edge ESLint, formatting, diff integrity, maintainability limits, upload-limit parity, documentation links and snapshot checks also passed.fc5baaa303468f0f8cc31c5cb0c8d2016dcf64c8: 24 pre-existing unplaced files registered to their actual owning areas, plus the generated ledger date. The development documentation check and 54 focused metadata tests passed; existing advisory warnings remain. This follow-up is NOT pushed, so the current PR does not contain it. No additional push is planned until its preview-compute exposure is controlled.Risk and rollout
Risk: medium — shared model handling and indexing validation require the recorded regression checks; live answer quality remains unproven. Keep draft; do not arm auto-merge. Better answer quality and lower total cost per accepted answer are not established. Production recovery can already make up to four generation calls under its deadline; the one-call component experiment cannot establish full-system operating cost.
Approval for paid evaluation must bind the source, fixtures, configuration, environment, verified pricing and a single absolute output directory. Persisted reservations precede dispatch; SDK/outer retries are zero. The 52-execution plan reserves 832,000 output tokens against the unchanged 400,000 cap, and published input reservations may stop it sooner. A feasible revised scope and specific spending approval are required before a live run. No spending cap was loosened. Offline preparation now proposes a 24-generation pilot (20 measured, four warm-ups) plus at most 24 exact input-count requests, capped at 480,000 generation input and 384,000 output tokens, with no retries. Its known generation reservation is US$6.492024. The input-count endpoint's billing terms are not established, so the complete total and its per-endpoint hard reservation remain blocked; no paid approval is requested or inferred. No request-bound allowance code has been introduced before that basis is complete.
On 27 October 2026 the owner reported automatic preview branching disabled, but the PR created an active Supabase preview. Read-only branch metadata verified preview project
vhuuqhbvarmrktwwygofbelongs to PR #3866 andcodex/rag-sol61-20261027, is non-default/non-persistent, and was created withwith_data:false. The owner explicitly authorised inspection and pausing only this preview. Supabase accepted the pause; follow-up metadata changed fromACTIVE_HEALTHYthroughPAUSINGtoINACTIVE. ProductionsjrfecxgysukkwxsowpyremainedACTIVE_HEALTHYand was not modified. Nothing was deleted. Prior charges and any non-compute billing remain unverified; automatic-branching settings were not inspected or changed. Keep this preview paused. Standard GitHub-hosted CI runs on this public repository; do not dispatch model evaluation or deployment.After accepted evaluation and separate activation/deployment authority: preserve explicit non-answer workload settings in both runtimes, deploy the matching Edge function, then set
OPENAI_ANSWER_MODEL,OPENAI_FAST_ANSWER_MODELandOPENAI_STRONG_ANSWER_MODELtogpt-6.1-sol. Retain low/medium effort andOPENAI_STORE_RESPONSES=false.Rollback: restore the recorded prior answer/indexing configuration and matching application/Edge artifacts. Cache identities separate model configurations. No database or index rebuild is required.
Clinical Governance Preflight
PsychSift Production(sjrfecxgysukkwxsowpy)Evidence / unresolved: new comparison fixtures are synthetic equipment examples. Owner clinical review, representative domain evaluation, provider/data-processing acceptance and applicable deployment classification remain prerequisites to activation. No clinical sign-off is asserted.