dsh-completion-gate
以证据为依据的 DeepSeek Harness 生产就绪完成门槛，配备持久化且可由操作员配置的控制中心。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:AGSQ11/dsh-completion-gate#ab75d3bf9c7ee8874dd91e6f10639dee1327cc73Completion Gate 以 dsh-completion-gate 发布，当前版本为 0.1.3。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Completion Gate is a DeepSeek Harness plugin that prevents a coding agent from successfully ending a changed task until there is evidence that the work is actually ready.
It combines two related safeguards:
The goal is not to make the model say that tests passed. The goal is to make DSH execute the available checks, bind the result to the current repository state, and reject stale or incomplete completion evidence.
LLM coding agents frequently stop in one of these states:
"Now I need to inspect the call site..."
→ tool call
→ reasoning
→ model naturally stops
or:
"Implemented. Looks good."
without having actually run the repository's tests, build, lint, typecheck, or reviewed every changed file.
Completion Gate adds a host-side boundary around those failure modes.
Agent works
│
├── model stops while clearly unfinished
│ └── Premature Stop Guard → continue
│
└── model appears genuinely finished
│
└── Completion Gate
├── repository changed?
├── tests
├── build
├── lint
├── typecheck
├── TODO/FIXME regression scan
├── security regression tripwires
├── changed-file review
└── acceptance criteria → evidence
│
PASS ─┴─ BLOCK
DSH exposes agent/turn-stopping when a model has naturally stopped and no tool call or queued continuation is still keeping the turn alive.
Completion Gate inspects the latest turn for strong evidence that the model itself said work remained, for example:
Now I need to inspect...Next I will run...I still need to...Let me trace......and in parallel...When detected, the plugin queues a waking agent.followup() for a fresh turn such as:
PREMATURE STOP RECOVERY:
Continue the unfinished task.
Execute the pending investigation or implementation now.
Do not summarize or claim completion yet.
The recovery chain is bounded by prematureStopMaxContinuations, so repeated premature/empty recovery turns cannot create an infinite follow-up loop.
Generic phrases such as Let me know if you want changes are not treated as unfinished work.
By default:
gateSubagents: false
Only the root agent owns final production readiness. Spawned DSH subagents are allowed to complete their delegated jobs normally.
This is the recommended setting when using autonomous orchestration, Phoenix/model failover, AI Council, critics, or large subagent teams. The root agent remains responsible for integrating the work and passing the final gate.
You can explicitly enable subagent gating from the control center if a project really needs it.
Completion evidence is tied to a SHA-256 fingerprint of the current repository state.
The fingerprint accounts for:
This matters because a clean worktree does not prove the agent made no changes: the agent may already have committed them.
If the code changes after a successful check or attestation, the fingerprint changes and the old evidence is no longer valid.
Completion Gate can auto-detect and execute common project checks.
| Ecosystem | Detection / checks |
|---|---|
| Node.js | package scripts such as test, build, lint, typecheck |
| npm / pnpm / Yarn / Bun | package manager selected from project lockfiles |
| Python | pytest when tests/config are present; Ruff and Mypy when configured |
| Go | go test ./..., go vet ./... |
| Rust | cargo test --all-targets, cargo check --all-targets |
| Composer / PHP | composer test, composer lint when scripts exist |
| Any project | operator-defined custom checks |
When requireTests is enabled, a project with no executable test command detected is considered blocked rather than silently treated as tested.
Successful machine results are cached for the exact workspace fingerprint. This avoids rerunning a ten-minute suite merely because the agent needs one extra turn to submit its final review.
The gate can inspect newly added lines for:
TODO
FIXME
HACK
XXX
and a deliberately narrow set of obvious security-regression patterns such as:
eval(...) / dynamic Function construction;chmod 777;This is a regression tripwire, not a replacement for dedicated SAST/security tooling.
Machine checks cannot prove every requirement. Before final completion, the agent can be required to call the completion_gate tool with an attestation bound to the current workspace fingerprint.
Example:
{
"action": "attest",
"reviewed_files": [
"src/engine.ts",
"tests/engine.test.ts"
],
"review_summary": "Reviewed the complete diff, error paths, state transitions and backwards compatibility.",
"acceptance_criteria": [
{
"criterion": "Risk sizing uses ATR",
"evidence": "Unit tests cover ATR-derived risk sizing and the full test suite passes."
},
{
"criterion": "Existing behavior remains compatible",
"evidence": "Regression tests and typecheck pass; changed call sites were reviewed."
}
],
"unresolved_issues": []
}
In strict mode the gate verifies that:
reviewed_files;unresolved_issues is empty;Open:
Settings → Completion Gate
The control center is both a runtime dashboard and a persistent configuration editor.
Strict: a failed gate steers the agent back and prevents normal completion.Advisory: evidence is collected, but failure does not force another step.The UI can add arbitrary operator-defined checks:
Name: integration-tests
Category: test
Command: npm run test:integration
Required: yes
Custom commands execute in the project's workspace on the DSH host. Only configure commands you trust.
Save settings applies changes immediately and persists them across dsh web restarts.
Configuration precedence is:
built-in defaults
↓
profile / cordis.patch.yml configuration
↓
settings saved from the Completion Gate UI
Whenever gate policy changes, cached machine evidence, attestations, and operator overrides are invalidated. A previous PASS is never reused under a newly changed policy.
Reset to profile defaults removes the saved UI overlay and immediately returns to the plugin configuration supplied by the active DSH profile.
The host-side settings file is intentionally not exposed in the browser UI or gate reports.
/gateRun the gate immediately for the active root session.
/gate
Useful when you want to inspect readiness before the model tries to finish.
/gate-statusShow the latest report for the active session.
/gate-status
/gate-resetDiscard cached machine evidence, attestation, and operator override for the active session.
/gate-reset
/gate-override <reason>Explicit operator override for the exact current workspace fingerprint.
/gate-override accepted temporary compatibility risk
If any file or commit changes afterward, the override no longer applies.
This is intended for the human operator, not as an escape hatch for the agent.
The plugin registers:
completion_gate
Actions:
run — execute readiness checks now;status — return the latest/current report;attest — submit the final changed-file and acceptance-criteria review.Completion Gate also adds a system-prompt section telling coding agents that the gate is authoritative before they claim completion.
The conversation header shows a compact status:
Gate · ready
Gate · 1 blocked
Gate · off
This is only an operator summary. The authoritative result remains the workspace-bound gate report.
The bundled profile patch starts with:
- insert:
- id: completion-gate
name: dsh-completion-gate
config:
enabled: true
mode: strict
gateSubagents: false
preventPrematureStops: true
prematureStopMaxContinuations: 3
autoDetectChecks: true
requireTests: true
requireBuildWhenAvailable: true
requireLintWhenAvailable: true
requireTypecheckWhenAvailable: true
blockNewTodos: true
securityScan: true
requireAttestation: true
gateOnlyChangedWorkspaces: true
commandTimeoutMs: 600000
maxOutputChars: 12000
maxChangedFiles: 500
customChecks: []
Everything above can be changed from the control center. Profile configuration remains the baseline that Reset to profile defaults returns to.
Example from the parent directory containing the plugin source:
dsh plugin --profile web add .\dsh-completion-gate
dsh web
Then open Settings → Completion Gate and review the defaults before relying on strict mode for a production repository.
No dependency installation is required by this plugin.
Replace the plugin source with the newer source revision and restart dsh web.
UI-saved Completion Gate settings are stored separately from the plugin source, so replacing the source directory does not intentionally reset operator preferences. Use Reset to profile defaults if you want the new profile defaults instead.
Completion Gate is session/workspace scoped, not model scoped.
If a model fails:
Model A
↓ 429 / timeout / provider failure
Phoenix / DSH failover
↓
Model B continues same session
↓
Completion Gate still owns the same baseline and workspace evidence
A replacement model can continue remediation without restarting the task. If the replacement changes code, the workspace fingerprint changes and stale gate evidence is invalidated naturally.
gateSubagents: false is recommended with autonomous failover/team orchestration so child jobs are not unnecessarily blocked by the root production-readiness policy.
In strict mode DSH receives a steering message similar to:
COMPLETION GATE BLOCKED. Do not claim completion.
- Node test: test failed.
- Changed files not reviewed: src/engine.ts
- Acceptance-criteria evidence is missing.
Fix the failures, review every changed file, then call completion_gate
with action=attest and concrete acceptance-criteria evidence.
The agent continues from the existing task/workspace state.
The Premature Stop Guard intentionally uses agent.followup() rather than agent.steer(). A follow-up lets the naturally stopped turn close and immediately wakes a new turn with the unfinished-work instruction. This composes better with provider retry/failover and avoids same-turn recovery steps that can visibly inject context and then fall idle.
The actual Production Readiness Gate still uses agent.steer() because blocking a failed readiness check must veto the current natural completion boundary.
DSH can record an assistant/message before agent/turn-stopping runs. Therefore Completion Gate can prevent the turn from closing and force another step, but it cannot retroactively erase premature prose already emitted by a non-compliant model.
The Premature Stop Guard and system-prompt instructions are intended to make the normal path continue before the task is treated as complete.
Passing tests do not prove correctness. The gate proves that configured checks passed for a specific repository fingerprint.
The gate verifies completeness and freshness of the attestation structure; it cannot mathematically prove that an LLM's semantic review was insightful. Independent review plugins, Council, security tooling, and human review still add value.
The built-in scanner catches obvious newly added risky patterns. It is not a comprehensive vulnerability scanner.
A custom check is operator-controlled code execution on the DSH host. Treat it with the same care as adding a command to a CI configuration.
enabled: true
mode: strict
gateSubagents: false
preventPrematureStops: true
prematureStopMaxContinuations: 3
autoDetectChecks: true
requireTests: true
requireBuildWhenAvailable: true
requireLintWhenAvailable: true
requireTypecheckWhenAvailable: true
blockNewTodos: true
securityScan: true
requireAttestation: true
gateOnlyChangedWorkspaces: true
This leaves implementation freedom to subagents and model failover while keeping one hard readiness boundary at the root agent.
completion_gate attestation tool.See SECURITY.md for the concise security model.
MIT. See LICENSE.