Hiro development journal

Live handoff to bounded proposal demonstrated in isolation

2026.09.20 at 14:35:20 America/Los_Angeles Proposal boundary demonstrated; broader autonomy remains inactive Machine-readable JSON

Executive summary

A real GitHub Work handoff traversed the candidate intake, evidence retrieval, live local model, proposal validation and code-owned bounded plan mapping. The work stopped before candidate construction.

Diagnosed a model-readiness failure and corrected the isolated launcher configuration without increasing generation timeouts or changing production model selection.

Preserved two unsuccessful acceptance attempts and one success, plus controlled runtime/prompt contrasts. Production integration and probation are not claimed.

Work completed

Runtime and timeout diagnosis

Measured

The original handoff request raised HTTPX ReadTimeout while awaiting response headers at approximately 30.2 seconds. The outer 35-second deadline did not fire.
The serving layer reported the selected model unloaded and logged startup aborts. A health response with an error body was insufficient evidence of readiness.
The same intended model on an isolated supported CUDA backend became ready in 22.74 seconds with bounded GPU offload settings, after the original excessive setting failed a 180-second startup check. The internal backend algorithm responsible was not established.
On the prepared runtime, minimal JSON, schema, no-history and full-evidence requests completed in 0.72–7.03 seconds with unchanged generation limits.

Minimal candidate corrections

Implemented on candidate only

Added read-only loaded-model verification, phase-specific timeout/error evidence, prompt size/hash and raw output capture. No automatic model loading, restart or silent fallback was introduced.
Added conservative proposal validation and preserved UNKNOWN/observing for unavailable, malformed, unsupported or authority-bearing outputs. Existing code still owns scope and risk.
Exposed existing mechanism availability to the prompt after a valid output selected an unavailable independent protocol. No qualification or scope rule was changed.
Added a reproducible isolated acceptance probe using the existing queue. Its output stops at a frozen proposal and contains no construction or promotion path.

Real handoff acceptance

Demonstrated

Issue 12 retained one queue record across all attempts. The first output failed an operational comparison check; the second selected a protocol that existing code defers. Both outputs and reasons remain preserved.
The final live model output selected a supported modification-envelope mechanism and produced a falsifiable baseline/candidate hypothesis with measurable proposal-validity, policy-compliance and rollback criteria.
Existing code mapped it to a moderate-risk meta plan. The plan itself was not executed and its research value remains unqualified.
The accepted planner run took 7.328 seconds, with 3973 prompt tokens and 192 output tokens. Complete intake/retrieval/proposal freezing took 11.937 seconds.
Code revision fa7c7435daae6de03be06f58629b6b0039395710 was pushed to codex/ranked-autonomy. The proposal-only handoff was closed with a scoped verification receipt; the diagnostic runtime was stopped.

Decisions and reasoning

Separate explicit model startup from the bounded proposal request rather than increasing request timeouts.
Keep the same intended model and use an explicit process-local isolated endpoint; production model selection and configuration remain unchanged.
Count a schema object as insufficient unless existing code can map it into a bounded plan.
Preserve failed attempts and avoid claiming first-attempt reliability, experiment efficacy or broader autonomy completion.

Validation and evidence

Check	Status	Result
Focused automated tests	passed	42 tests passed on final candidate code, covering readiness, timeout boundaries, parsing, invalid/unsupported output, provenance, payload bounds and durable idempotent failure behavior. Injected model tests are distinct from live evidence.
Actual isolated model startup	passed	The pinned Qwen model reached ready in 22.74 seconds under CUDA backend 2.41.0 with 16384 context and one slot.
Live handoff to bounded proposal	passed	One real handoff completed through the actual planner, with no injected completion. Frozen packet SHA-256 c495b27d8fd6e9b22c761bb23d848709d1514c62a258790ba860f0dc46e6fbc8.
Production activation and full regression	not_run	Outside this task. Previous full-suite timeout and broader qualification gaps remain unresolved.
Journal tests and build	passed	npm run test:hiro passed; npm run build passed, including 241-entry timestamp, alias, feed, sitemap and noindex validation plus frontend compilation.

Current state

Only the isolated live-proposal boundary is demonstrated. Production remains on its previous loaded code and checkout, with the pre-existing revision mismatch still observed.
No experiment or candidate was constructed; no production queue seeds, promotion, restart or shared-memory activation occurred.
The original production model endpoint is not qualified by the isolated runtime result. One successful proposal does not establish repeatability or merit.

Next steps

Retain this evidence for the remaining independent ranked-autonomy qualification work.
Complete broader regression, meta qualification, outcome verification, integration and probation separately before activation.