# 0.1.3: review process and bounded recovery

Published on 2026-10-08 to [official npm](https://www.npmjs.com/package/dsh-task-supervisor/v/0.1.3), with latest at 0.1.3. The public package is byte-identical to the frozen artifact. [GitHub Release](https://github.com/gosomea/dsh-task-supervisor/releases/tag/v0.1.3) provides the same tarball. Clean public installation passed on official DSH 0.2.0-rc.2.

## Changes

The primary Session now displays Supervisor's native review conversation at the durable intervention position, with review kind, node attempt and review attempt. Active reviews expand; ended reviews collapse while retaining their outcome and next step. Historical expansion retains the reviewer Session only while open. Primary Agent replies are unchanged, with no reviewer composer. Replies, tool groups, errors, duration and reasoning follow native presentation settings.

The DAG shows task progress; the sidebar shows requirements, check scope and history. The process appears once in the primary conversation. Current activity is separate from previous decisions, with measured duration, recent records, tools and deadlines.

Plan submissions, node reports, final acceptance and manual retries return quickly before the controller runs reviews. Internal timeouts and recognized temporary request failures retry once by default, retaining the job, Session, cutoff, requirements version, node attempt and snapshot. Protocol supplements have a separate counter and do not extend an attempt deadline.

Successful recovery continues only with a valid earlier approval and matching versions. User pause, close, Host restart, needs-user and older records without permission await manual recovery. Initial plans retain their approval policy; recovery grants no new authorization. Review control records use version 7. Successful native tool results restore read eligibility; failures, index summaries and model claims do not.

Changes are confined to the plugin, without altering DSH's main loop, PTC timeout or native sandbox. See the [README](../../README.md) and [review process and recovery](../review-experience.md).

## Configuration and upgrade

| Configuration | Default | Range |
| --- | --- | --- |
| reviewFaultRetryAttempts | 1 | 0–3 fault retries per job |
| reviewFaultRetryDelayMs | 3000 | 0–60000 milliseconds |
| resumeAfterReviewRecovery | true | Reuse valid earlier approval only |

Install through the official DSH command:

```sh
dsh plugin add --profile <your-profile> dsh-task-supervisor@0.1.3
```

New tasks keep native modes without Supervisor presets. Older records remain readable. Early reviewer Sessions without native parent/child descriptors cannot embed the native child conversation; the compatibility reader remains available without rewriting history. Older jobs without a stored permission do not gain automatic continuation. Restart your instance when idle to upgrade; this work did not restart in-flight tasks on port 3080.

## Acceptance results

| Check | Observed result |
| --- | --- |
| Focused kernel and Host regressions | 28 files, 405 passed, 13 skipped |
| Strict Host/Client types and build | Passed |
| Long review across PTC boundary | A controlled review completed after 125 real seconds, beyond 120 seconds; the submitting tool returned earlier |
| Log mode with real models | Two tasks in one Session; one approval each; plan, node and final reviews passed |
| Independent mode with real models | Three artifacts read and Docker checks executed; 6 node commands and 12 final commands |
| Request fault recovery | One injected HTTP 503 on the first node request; one retry in the same job/Session, then continuation |
| Actual review deadline recovery | First final attempt exceeded 300 seconds; same Session and prior checks retained for one successful retry |
| History and refresh | Six unique log jobs and three independent jobs; collapsed by default; history viewing added 0 model requests |
| Expansion and release | Native tools, errors and replies visible; 0 mounted native transcripts after collapse; no visible reviewer composer |
| Layout | 760×900 viewport, long records, sidebar navigation and history browsing checked; no forced scroll to bottom |

The 13 conditional Docker tests skipped because the generic kernel command had no Docker parameters: 5 gateway tests, 7 command runner tests and 1 controlled defect scenario. The real independent case separately used a pinned Docker image; that does not turn skipped tests into passes. Windows, other browsers and other DSH versions were not exercised in this release cycle.

Actual request headers for the primary Sessions and every reviewer Session matched the daily CodeBuddy `deepseek-v4.1-flash` route. Native reported tokens across all Sessions were **1,828,258** for log cases (including a route-health turn) and **4,783,475** for the independent fault case. Counts include cached input and do not represent paid tokens; failed requests without usage remain unknown. Executed check counts include failed commands subsequently corrected, not a coverage percentage.

A local forwarding proxy injected exactly one 503. Native llm-retry was disabled only in that fixture so the failure reached Supervisor. No rescue prompt, hidden test or extra approval was supplied. The final review timeout was not injected. Each recovery consumed its own job's single retry allowance. The input hash remained unchanged; probe failures and revisions remain in native history.

Sanitized machine evidence: [0.1.3-checks.json](0.1.3-checks.json). Screenshots:

![Live review in the primary conversation](0.1.3-assets/live-review.jpg)

![503 and recovery retained in one job](0.1.3-assets/fault-recovery.jpg)

![Native process and final sidebar result](0.1.3-assets/native-review.jpg)

## Release installation

Publication uses the same frozen npm tarball checked through a clean official DSH home, with actual Host/Client file and loading checks. Machine evidence records the public version, latest tag, digest, Git tag and GitHub Release.

These cases test usability and recovery correctness, not a public benchmark comparison. Correction, false pause and false acceptance rates without blind annotation remain `null`. No independent browser executor was added, and no superiority over Goal, Plan or Agent Team is claimed.

[中文](0.1.3.zh.md)