dsh-completion-supervisor
DeepSeek Harness plugin checking whether an agent's 'done' claim is true, using deterministic evidence plus one batched Jev assessment.
Overview
dsh-completion-supervisor checks whether a coding agent's claim of completion is actually true. Agents sometimes say tests passed when they never ran, create untracked files, summarise results that contradict recorded commands, or gloss over failed builds. At the moment a turn is about to close, the plugin gathers deterministic evidence in code, compresses it into a state, asks Jev seven narrow yes/no questions in one request (requirements satisfied, implementation complete, verification sufficient, blocking issue remaining, evidence matches claim, needs more verification, ready to finish), and runs a pure policy.
Since policy version 3, interventions are driven only by deterministic facts (an unresolved tool error, an unsupported verification claim, a recorded failing test), at most once per task; Jev's probabilities are logged with ground truth for later analysis but decide nothing. The README, mainly in Chinese, documents what Jev does and does not do, reports 566 to 850 ms latency and about $0.000022 per call, and explains why an earlier per-tool-call gate was dropped.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | requirements_satisfied: are the user's requirements satisfied? | yes · no |
| Yes / no | evidence_matches_claim: does the recorded evidence match the completion claim? | yes · no |
| Yes / no | ready_to_finish: is the task ready to finish? | yes · no |
How it works
- At turn-stopping, gather deterministic task facts
- Send one Jev request with seven nouls (logged)
- Pure policy intervenes only on deterministic evidence
- Log the assessment and the user's next action as ground truth
Strengths
- One call per completion claim rather than per tool call
- Honest separation of logged probabilities from enforced rules
- Cheap and fast per the README
Caveats
- Jev outputs currently do not drive decisions
- Documentation primarily Chinese
- Specific to DeepSeek Harness
Getting started
Install as a DeepSeek Harness plugin per the README (English version in README.en.md).At a glance
- Type
- Plugin
- Status
- beta
- Language
- TypeScript
- Runtime
- DeepSeek Harness plugin
- Has code
- yes
Use cases and integrations
How curated lists describe it
Checks whether a coding agent completion claim is actually true: deterministic evidence gathered in code, one batched Jev assessment, and a pure policy. DSH plugin.
hellogumbo/awesome-jev
Frequently asked questions
Does Jev decide whether to intervene?
Not in policy_v3; it is logged, and only deterministic evidence triggers interventions.
How many Jev calls per task?
One to three, at completion claims rather than every tool call.
What does each call cost?
About $0.000022 per the README.
Related projects
Local desktop pal that verifies what a coding agent really did, with optional Laya or Jev double-checks of unclear test results.
ProjectbetaTypeScriptDesktop app running coding agents in parallel git worktrees, with optional Jev judgments of each changed file against the task.
AppbetaLinear-backed work ledger for coding agents where a Jev-compatible model verifies done reports, blockers and plan changes against the task.
ProjectprototypeJavaScriptDeepSeek Harness plugin where Jev prunes long tool output, screens fetched pages for injected instructions and gates completion claims.
Agent skillbetaTypeScriptSingle-agent Jev coprocessor for Pi: semantic pre-edit gates, evidence checks, diff review and append-only telemetry.
ProjectbetaTypeScriptOpen lab on using Jev routing in front of Claude Code correctly: measured bugs and patch, hard fallback, completion gate.
Projectresearch