fix: actual retained-task recovery blocked by provider budgets, queue timeouts and startup head-of-line #433

Merged
karsten merged 1 commit from fix/assessment-provider-boundaries into main 2026-11-13 22:30:59 +00:00
Owner

Actual user report: real persisted tasks did not resume despite prior synthetic restart proofs. Shikamaru ready tasks accumulate invalid_shape (up to67 retries); host emits24–53KiB while old TypeSafe adapter silently capped24KiB. Real synthetic gateway response uses canonical openjev-0.1 for request alias openjev-latest. Che real configured tool-free reviewer took32.69s (including503 retries), beyond30s startup default.

Fixes:

  • Pin protected Plugins101 merge d30bea8:56KiB bounded transport, documented alias/canonical response handling, low confidence=>uncertainty (no threshold weakening/no fallback).
  • Optional trusted provider maxPromptBytes validated1024..56000; host projects observational evidence to BOTH transport and host limits, never original authority/expiry.
  • Startup/Main default execution budget120s, same as step assessment. Explicit operator timeout remains exact.
  • Acquire shared host model capacity before provider execution timer. Recheck stop/expiry/abort after queue; nested admission reuses lease. Non-model transports need no runtime model slot.
  • One persisted policy-stamp read-only reassessment for a retained settled report after provider/model/budget boundary update. Preserve original IDs/intent/evidence, all counters, stop/expiry/ownership/held-legacy/questions; no opaque-effect replay or manual unpause. Same policy on subsequent restart preserves normal backoff.
  • Once channels/stop reconciliation ready, release each durably admitted Main recovery immediately instead of waiting for EVERY unrelated startup judge. Stop cancels queued parent assessors.

Local exact merged-Plugin integration:2385 passed/1 skipped; typecheck/build/diff-check green;8 new focused regressions incl long transport input, immutable mandate, queued timeout, queued expiry, current-policy backoff, same execution63-attempt migration, startup isolation and graceful stop. Existing crash-after-reservation assertions retained; fixture now persists current policy stamp as production does. One unconstrained high-concurrency run hit native process startup test wall-clock contention; canonical concurrency4 full suite passed. Plugins required compatibility CI green/normal merge.

NOT a live completion claim: actual-task shadow now accepts8/10 scopes; two remain privacy-export-held (no privacy waiver). Current TypeSafe5s HTTP budget independently timed out; diagnostic-only longer read-only request returned504 at60.5s. This external latency/export gate needs separate treatment; no model substitution/config rewrite/manual ledger changes performed. Final acceptance requires actual retained task progress on installed source.

Actual user report: real persisted tasks did not resume despite prior synthetic restart proofs. Shikamaru ready tasks accumulate invalid_shape (up to67 retries); host emits24–53KiB while old TypeSafe adapter silently capped24KiB. Real synthetic gateway response uses canonical openjev-0.1 for request alias openjev-latest. Che real configured tool-free reviewer took32.69s (including503 retries), beyond30s startup default. Fixes: - Pin protected Plugins101 merge d30bea8:56KiB bounded transport, documented alias/canonical response handling, low confidence=>uncertainty (no threshold weakening/no fallback). - Optional trusted provider maxPromptBytes validated1024..56000; host projects observational evidence to BOTH transport and host limits, never original authority/expiry. - Startup/Main default execution budget120s, same as step assessment. Explicit operator timeout remains exact. - Acquire shared host model capacity before provider execution timer. Recheck stop/expiry/abort after queue; nested admission reuses lease. Non-model transports need no runtime model slot. - One persisted policy-stamp read-only reassessment for a retained settled report after provider/model/budget boundary update. Preserve original IDs/intent/evidence, all counters, stop/expiry/ownership/held-legacy/questions; no opaque-effect replay or manual unpause. Same policy on subsequent restart preserves normal backoff. - Once channels/stop reconciliation ready, release each durably admitted Main recovery immediately instead of waiting for EVERY unrelated startup judge. Stop cancels queued parent assessors. Local exact merged-Plugin integration:2385 passed/1 skipped; typecheck/build/diff-check green;8 new focused regressions incl long transport input, immutable mandate, queued timeout, queued expiry, current-policy backoff, same execution63-attempt migration, startup isolation and graceful stop. Existing crash-after-reservation assertions retained; fixture now persists current policy stamp as production does. One unconstrained high-concurrency run hit native process startup test wall-clock contention; canonical concurrency4 full suite passed. Plugins required compatibility CI green/normal merge. NOT a live completion claim: actual-task shadow now accepts8/10 scopes; two remain privacy-export-held (no privacy waiver). Current TypeSafe5s HTTP budget independently timed out; diagnostic-only longer read-only request returned504 at60.5s. This external latency/export gate needs separate treatment; no model substitution/config rewrite/manual ledger changes performed. Final acceptance requires actual retained task progress on installed source.
fix: unblock retained-task assessments at provider and startup boundaries
All checks were successful
decision-interface-ci / full-validation (pull_request) Successful in 2m51s
46c65a7df7
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
harness/core!433
No description provided.