Hiro development journal

Retire exhausted queue entries and refill from eligible evidence

Implemented, tested and runtime verified Machine-readable JSON

Executive summary

Changed reservoir selection to filter exhausted, explicitly incomplete, and policy-excluded evidence before applying the 100-entry limit.

Unusable waiting work is retired with history preserved. Passing unrelated controls no longer triggers repeated attempts or a claim of hypothesis disproof.

Work completed

Admission and retirement

Implemented, tested and active
  • Use all checksum-verified historical source packets as selection input, apply eligibility first, then limit the live reservoir to 100.
  • Preserve source content fingerprints in grouped queue records, so unchanged evidence cannot be recycled merely by changing the baseline revision. New content can be reconsidered.
  • Retire unowned queued, observing, or artifact-blocked external records excluded by policy or explicitly marked as incomplete claims. Preserve active work and supervisor-owned retries.
  • Route insufficient external evidence to superseded history without claiming the idea was disproven. Transient unavailable evidence retains bounded retries.
  • A depleted eligible reservoir requests replenishment every configured minimum discovery interval, instead of waiting for unrelated evidence or the two-hour idle window. Provider fetch limits remain unchanged.
  • Live verification exposed three legacy routing leads with no concrete approach. Added the same specificity requirement used by queue grouping to reservoir selection, preventing unavailable work from being shown as available.

Dashboard

Implemented, tested and active
  • Show eligible source and grouped-approach counts, historical exclusion reasons, and waiting eligibility.
  • Correct runtime-probation display to executed checks, consistent with the deployed policy.

Decisions and reasoning

Validation and evidence

CheckStatusResult
Focused tests passed Final focused tests: 95 passed in 12.56 seconds, including queue, reservoir, duplicate protection and active discovery scheduling.
Dashboard JavaScript syntax passed node --check passed on extracted inline scripts.
Read-only copy of live queue passed Preview retired ten stranded waiting records. Selection inspected 136 historical source entries: 131 excluded (42 policy, 35 explicitly incomplete, 54 processed evidence), five unprocessed routing leads remained.
First full regression failed_then_fixed 6 failed, 1043 passed, 11 skipped in 496.48 seconds. All six failures were UTF-8 decoding errors caused by three newly inserted Windows-encoded middle dots in the dashboard. Corrected encoding; six page tests then passed in 0.11 seconds. Full suite is being rerun.
Final line-ending normalization passed Final fa89f43 differs from full-suite source46552dd only in line endings, verified using git diff --ignore-space-at-eol. Six dashboard checks passed again in 0.10 seconds. Exact-revision qualification is running on fa89f43.
Final pipeline qualification passed fa89f43d5798284e0fec885d57a4cc94e3e3bac1 passed all five frozen scenario groups: candidate evaluation, promotion/rollback, autonomous recovery, runtime isolation/state and assistant product path.
Full regression rerun passed 1052 passed, 11 skipped, 18 warnings in 482.74 seconds in WSL/Bubblewrap at46552dd. Finalfa89f43 differs only in verified line endings and separately passed exact-revision qualification.
Production integration passed Backed up live queue; retired ten stranded entries; second retirement made no changes. Fast-forwarded production and restarted with existing launcher. Loaded revision and checkout bothfa89f43d5798284e0fec885d57a4cc94e3e3bac1, Qwen connected.
Final routing specificity guard passed 102 focused queue, scheduling, duplicate-protection and dashboard tests passed in 13.28 seconds. Runtime code delta fromfa89f43 is the two-line routing-specificity guard; exact qualification underway.
Canonical controller after first deployment passed The scheduler resumed without manual candidate selection: reactivated a bounded incident, completed investigation, entered candidate construction, and admitted a newly exposed external routing approach. API showed one active candidate and two waiting entries. The dashboard served valid UTF-8 and the updated executed-check probation policy.
Routing guard final qualification and deployment passed All five qualification groups passed at3513da2084df74b3616f23f7d743740f1b41f60b. The final runtime delta fromfa89f43 is a two-line routing-specificity guard, separately covered by102 focused tests. Existing launcher restarted Hiro after construction reached an idle boundary; loaded revision equals checkout3513da2, Qwen connected.
Final live queue and reservoir passed Verified qualification current and audit enabled. Reservoir:2 eligible source entries representing1 grouped approach;134 excluded from136 historical entries. Queue snapshot:3 waiting and1 retrying; the attempted candidate scheduled its next repair for2026-09-14 22:18:39 PDT. No new autonomous promotion is claimed.

Current state

Next steps