Hiro development journal

Hiro and local Qwen enter a verified maintenance shutdown

2026.08.09 at 15:40:39 America/Los_Angeles Offline by user request Machine-readable JSON

Executive summary

At the user's request, Hiro and its local Qwen inference service were shut down for a planned maintenance interval.

The shutdown covered Hiro's API process tree, Daylab, local graph helper, Telegram lifecycle, benchmark tunnel, and all Hiro scheduled tasks.

Qwen was unloaded through LM Studio's supported lifecycle and its local inference server was stopped.

Post-shutdown checks found no listeners on Hiro's application ports or Qwen's inference port, no verified Hiro processes, and zero loaded LM Studio models.

Stage 6 remains fail-closed: its scheduler is disabled, its persistent sentinel is present, no promotion is active, and the frozen replacement candidate remains available for a later restart.

Work completed

Hiro runtime shutdown

Complete

Identified the exact API parent and child processes, Daylab process tree, local graph helper, and benchmark tunnel before termination.
Stopped Telegram through its checked-in lifecycle helper so intentional downtime cannot generate service alerts.
Stopped Daylab through its checked-in lifecycle helper and terminated only the remaining verified descendants.
Disabled the nightly preflight, Stage 6 evidence promotion, and legacy Hiro scheduled tasks so they cannot relaunch services during maintenance.

Qwen shutdown

Complete

Verified that Qwen was served by LM Studio on the configured local inference port.
Unloaded the active Qwen model and stopped the LM Studio local server using the supported command-line lifecycle.
Left the LM Studio desktop application itself outside Hiro's runtime scope; no model remains loaded and no inference listener remains active.

Stage 6 preservation

Fail-closed

Kept the Stage 6 scheduled task disabled and verified the persistent DISABLED sentinel remains present.
No promotion or probation was active at shutdown.
The frozen replacement evidence packet is durable on disk and does not depend on Hiro remaining online.

Decisions and reasoning

Use checked-in lifecycle helpers and supported model-server commands before resorting to exact process termination.
Disable every Hiro scheduled task during maintenance to prevent an unexpected automatic restart.
Do not start or consume a new Stage 6 authorization window while Hiro is offline.
Preserve the frozen replacement candidate and require new health, metrics, and exact-identity checks after restart.

Validation and evidence

Check	Status	Result
Hiro network endpoints	passed	No listener remained on the API, dashboard, local helper, benchmark tunnel, or configured inference ports.
Hiro process inventory	passed	No previously verified Hiro API, Daylab, graph-helper, or benchmark-tunnel process remained.
Qwen model state	passed	LM Studio reported an empty loaded-model list after unloading Qwen and stopping its local server.
Scheduled relaunch prevention	passed	All three discovered Hiro scheduled tasks reported the disabled state.
Stage 6 external kill switch	passed	The persistent Stage 6 sentinel remained present after shutdown.

Current state

Hiro is offline.
Qwen is unloaded and its local inference server is offline.
Daylab, Telegram, the local graph helper, and the benchmark tunnel are offline.
All discovered Hiro scheduled tasks are disabled.
Stage 6 has no active promotion or probation, and the external sentinel remains present.
The corrected replacement candidate remains frozen for a future, separately approved activation.

Next steps

Leave Hiro and Qwen offline until the user requests restart.
On restart, bring Qwen online first, verify exact model identity, then start Hiro and confirm its local health and metrics endpoints.
Restart only the auxiliary services the user wants, and deliberately re-enable any required scheduled tasks.
After runtime health is stable, request exact approval for the corrected Stage 6 revision and replacement candidate before opening a fresh 72-hour window.