Hiro development journal

Why Hiro's active improvement queue appeared empty

2026.08.11 at 20:43:41 America/Los_Angeles Diagnosis completed; corrective changes not yet applied Machine-readable JSON

Executive summary

The improvement inputs are enabled, but the visible ranked backlog can reach zero because the dashboard separates the single item being processed from the waiting queue.

At diagnosis time Hiro had one actionable arithmetic-reasoning audit finding in candidate construction, even though the ranked waiting count was zero.

External discovery is functioning. A recent eligible fetch reviewed 71 items across Moltbook, GitHub, BBC Technology, Wired AI, and Ars Technica and formed three mechanism-level hypotheses. arXiv was correctly waiting for its daily interval, and Reddit remains disabled.

The more important issue is lifecycle design: external observations collapse into a small number of mechanism identifiers, previously rejected identifiers cannot re-enter the queue when new supporting evidence arrives, and candidate failures immediately become terminal rejections rather than revised attempts.

Conversation intake reads Hiro's own response-quality database, not conversations with Codex. It only queues reproducible failure signatures, so a user conversation that does not trigger those signals contributes nothing to the backlog.

No Hiro source files or runtime records were changed during this diagnostic session.

Work completed

Live queue reconciliation

Diagnosed

The queue ledger contained 39 total records: three implemented, twenty-five rejected, ten superseded, and one actionable interaction-audit record.
The actionable record was in candidate construction. The Observatory's ranked-queue count excludes current work, which made the queue appear empty while processing was still underway.
The current worker lease and local-model connection showed that candidate work had started rather than waiting for authorization.

External-source intake

Running with a re-entry gap

Moltbook, GitHub, BBC Technology, Wired AI, and Ars Technica are enabled on two-hour minimum fetch intervals; arXiv is enabled on a twenty-four-hour interval.
The latest eligible source pass reviewed 71 items and reduced them to three bounded mechanism hypotheses. The following pass reviewed zero items because every source was still within its configured minimum interval.
The three latest hypotheses reused lineage identifiers that already had terminal rejected records. Queue insertion therefore did not create a new actionable record despite additional supporting evidence.
The source packet's created-idea count describes hypotheses emitted into the packet, not the number actually admitted into the active queue, making the discovery summary look healthier than the queue handoff was.

Candidate throughput

Over-terminal

The worker can advance up to three transitions each minute, so a small newly admitted backlog is consumed quickly.
External candidates that do not clear the first isolated construction and evaluation attempt move directly to rejected status.
The public queue showed repeated candidate-failed outcomes across Moltbook, arXiv, GitHub, and news-derived ideas without a bounded revision attempt, which explains the high rejection count and short-lived backlog.

Conversation and proactive-audit intake

Partially effective

Hiro's production response-quality harvester had scanned through the latest available local quality-log row.
Codex conversations are not part of Hiro's response-quality database and therefore do not become improvement incidents automatically.
The rotating everyday-question audit produced two failures in its first batch. One first candidate was rejected and the other remained in active candidate construction during diagnosis.
The audit and conversation paths are producing findings, but terminal rejection and dashboard presentation can make that activity look like an empty queue soon afterward.

Decisions and reasoning

Treat zero waiting items and zero total work as different states; current candidate, testing, and canary work must be included in the user-facing backlog picture.
Do not interpret frequent source checks with zero reviewed items as a source outage when the source-specific minimum fetch interval has not elapsed.
Count external hypotheses as admitted only after queue insertion succeeds.
A rejected mechanism should be eligible for a new version when materially new source support or a changed local baseline exists, while exact duplicate evidence should remain suppressed.
A candidate-generation failure should support a small bounded revision budget before terminal rejection; safety-gate failures should remain immediately terminal.
Keep Hiro conversation capture and proactive audit capture distinct from Codex discussion so private task contexts are not silently mixed.

Validation and evidence

Check	Status	Result
Live service health	passed	Hiro reported healthy with the configured local model connected.
Queue ledger inspection	passed with findings	One actionable interaction-audit item was present in current candidate work while the waiting ranked list was empty.
External discovery history	passed with findings	The latest eligible cycle reviewed 71 items and emitted three hypotheses; the next cycle correctly found all enabled sources not due.
Lineage re-entry comparison	failed	The three recently emitted mechanism hypotheses reused identifiers belonging to terminal rejected records and did not restore actionable work.
Conversation intake watermark	passed	The production failure harvester had scanned through the newest response-quality row available at diagnosis time.

Current state

Hiro is healthy and the continuous scheduler is enabled.
Moltbook, GitHub, arXiv, BBC Technology, Wired AI, and Ars Technica remain configured sources; Reddit remains disabled.
One proactive arithmetic-reasoning finding was actively being processed when the live ledger was inspected.
The ranked waiting list was zero because current work is shown separately, not because every input path was inactive.
External-source evidence is being fetched, but rejected mechanism lineages cannot presently re-enter as revised hypotheses.
No production code or queue records were modified during this diagnosis.

Next steps

Change the Observatory to show available backlog, current work, deferred work, and source intake separately rather than calling only waiting records the queue.
Version external mechanism hypotheses by evidence set and baseline revision so materially new support can re-enter after an earlier rejection.
Report both emitted hypotheses and successfully admitted queue records for every discovery cycle.
Add a bounded candidate-revision budget for construction or test failures while keeping safety and invariant failures terminal.
Introduce a configurable backlog floor so discovery can refill before the final waiting item is consumed.
Keep expanding the proactive interaction suite and make its failed and in-progress counts visible alongside external ideas.