Add explicit Curated writer and reviewer request profiles #116

Activity

1. R055LE commented on Oct 29, 2026

   

   R055LE

   on Oct 29, 2026

   OwnerAuthor

   More actions

   Include the cache-policy distinction from #132 in the explicit request profiles: response-cache bypass alone does not disable provider prompt caching. Verify the intended prompt-cache control at the serialized request boundary and retain allowlisted usage details so the evaluated boundary can be checked later. Coordinate pricing with #117; do not silently inherit the evaluation's blanket cache-disabled claim.
2. R055LE commented on Oct 29, 2026

   

   R055LE

   on Oct 29, 2026

   OwnerAuthor

   More actions

   Starting the authorized production-profile implementation on issue-116-curated-profiles, based on PR #133. A single writer owns the worktree; integration review and reservation design are read-only.

   The change will add an opt-in profile for the evaluated story writer/reviewer, preserve the existing default path, verify the serialized request and classified response handling, and exercise the no-send path offline. Request controls include the cache distinction from #132. #117 follows as a separate durable-reservation change.

   The coordinator will inspect the final diff, run the full checks, and prepare the PR. No paid preview, production model switch, budget increase, merge, deployment or posting is included in this implementation step. Live acceptance remains a separate concrete preview after both changes are ready.
3. R055LE commented on Oct 29, 2026

   

   R055LE

   on Oct 29, 2026

   OwnerAuthor

   More actions

   The approved current-Scout preview completed through the production wrapper at combined #134/#135 head 6fd942d18635bfb47ce87fb96d5bab8c52ea5e7c.

   • One workflow selected a Mistral Large 4 weights-release brief from five eligible sources.
   • Four calls (writer, reviewer, bounded revision, reviewer). Every reservation settled within the approved allowance; no new unknown exposure or overrun.
   • Exact requested profiles and returned route-match checks passed. All four responses reported zero cache-read and cache-write tokens; this is provider-reported evidence, not independent proof of absent caching.
   • The initial review rejected only the decision reason's citation support. Revision added the earlier-preview evidence; public wording was unchanged. Final review accepted.
   • The no-send path left story, decision, observation, delivery and seen tables unchanged in the isolated scratch database. No Discord delivery, merge, deployment or production configuration change occurred.

   Independent assessment passes this as a narrow, grounded availability brief. The adjective “improved” still adopts vendor framing without attribution, and the commentary is minimal; that remaining editorial refinement belongs with #97. Empty-history acceptance does not prove live story continuity, duplicate suppression, scheduler behavior, permissions or delivery. Human review of the copy and production activation remain separate decisions.
4. R055LE commented on Oct 30, 2026

   

   R055LE

   on Oct 30, 2026 – with ChatGPT Codex Connector

   Last edited by R055LE

   OwnerAuthor

   More actions

   Read-only recurrence check, October 30:

   • The October 27–30 scheduled runs all reported unusable model response: response is not JSON; skipped.
   • The running release is still 8f01ccc, with developing stories disabled and the legacy Curated chain. The selected profile and reservation PRs remain open. The successful October 29 isolated preview did not change production.
   • Legacy Curated requests ordinary text with a 900-token output limit. Its shared parser emits the same error for malformed JSON, empty text, refusal text and truncated output. Daily aggregate usage lacks role, returned model and finish reason, so these historical failures cannot identify the exact provider response defect.
   • Offline fake-response probes reproduced that ambiguity. The selected profile adds strict response schemas, role-specific headroom and classified response failures. Merely merging with legacy defaults would leave the observed path unchanged.

   Concrete remediation proposal, pending owner approval:

   1. Land the existing reviewed stack in dependency order: Constrain story output and bound validation retries #109, Add isolated story model evaluation #113, Preserve material distinctions in rendered stories #118, Ground story stages through revision #120, Account for prompt caching in story evaluations #133, Add evaluated Curated writer and reviewer profiles #134, Reserve Curated spend before provider attempts #135. Preserve required CI, verify final tree against the tested combined head, and reconcile the committed Compose additions.
   2. Enable CURATED_REQUEST_PROFILE=sol-astra-v1 and CURATED_DEVELOPING_STORIES=true.
   3. Initial daily capacity: DAILY_TOKENS_CURATED=200000, DAILY_USD_CURATED=2.14, DAILY_ATTEMPTS_CURATED=4, DAILY_WORKFLOWS_CURATED=1. These are admission ceilings, not expected expenditure. The configured values passed offline Settings validation and cover the conservative two-writer/two-reviewer allowance.
   4. Use the existing digest destination and daily schedule for the first production check. Manual previews share the same allowance. No extra paid trial or immediate manual post is proposed.
   5. Before activation, recheck current endpoint compatibility and account capacity without inference, verify live history/delivery compatibility, and preserve a rollback copy of existing non-secret configuration. Verify release identity, effective settings, health, and the next scheduled outcome. If activation fails, pause automatic Curated work by clearing its destination before any image/config rollback. Retain ledger history; legacy mode does not enforce the selected-profile uncertainty/overrun holds, so do not resume it while those remain unresolved. Pausing requires container reconciliation, not only editing the environment file; avoid manual Curated previews during rollback/reconciliation. Recheck actual daily usage immediately before activation.

   This is the initial repair cutover; higher-frequency stream settings remain follow-up under #98. Approval must cover landing, deployment and profile/capacity activation. No production mutation or provider completion was performed during this diagnosis.
5. R055LE commented on Oct 30, 2026

   

   R055LE

   on Oct 30, 2026 – with ChatGPT Codex Connector

   OwnerAuthor

   More actions

   Approved production cutover completed October 30 at 489695194fcc75df272a0f4e30ebaa037e0d62be.

   • Landed Constrain story output and bound validation retries #109, Add isolated story model evaluation #113, Preserve material distinctions in rendered stories #118, Ground story stages through revision #120, Account for prompt caching in story evaluations #133, Add evaluated Curated writer and reviewer profiles #134 and Reserve Curated spend before provider attempts #135. Squash-history conflicts were resolved with exact reviewed-tree equality; independent checks matched every landed tree. Final CI passed all 755 tests on Python 3.12 and 3.14, lint, coverage and dependency audit.
   • Release attempt 2 passed build, vulnerability scan and signing. Host verification passed under the existing workflow identity and issuer, then the documented deploy service succeeded. The first attempt's signing failure and successful bounded retry are tracked in Handle release signing failure on transparency-log duplicate entries #136.
   • Effective settings now select sol-astra-v1 with developing stories enabled, 200,000 daily tokens, $2.14 daily ceiling, four attempts and one workflow. The existing digest destination and 07:00 daily schedule are retained. Other settings and credentials were verified unchanged.
   • The running image is healthy, Discord is ready, channel/permission checks passed, database integrity passed, and the new ledger is present with zero immediate Curated attempts. Legacy daily usage remains included, with no reserved or unknown exposure. Normal deployment polling is restored.
   • Startup's two catalog-membership warnings are confirmed false positives: dated IDs appear in canonical_slug, while preflight checks only id. Exact dated endpoints still resolve correctly. Narrow diagnostic follow-up: Recognize canonical model slugs in startup preflight #137.

   Next scheduled check: October 31 at 07:00 America/Detroit. No immediate public post or additional paid preview was invoked. Activation acceptance is complete; the first scheduled reporting/delivery outcome remains to be observed. Higher-frequency stream settings remain under #98.
6. R055LE

   closed this as completedon Oct 30, 2026