dsh-pr-gate
Manifest validDeterministic write gate: scope policy denies out-of-scope write/edit calls before they touch the disk, residue rules scan the written text, and every decision lands in a content-hashed JSONL ledger;
dsh-pr-gate
Deterministic write gate for DeepSeek Harness (dsh) — the CI gate moved to the moment of the write.
Every file the agent writes (write / edit / str_replace_editor) passes
the same scope policy and residue rules that
agent-pr-gate applies to PRs
in CI — deny-before-write in tools/pre-execute, residue scan of the written
text in tools/post-execute, and a durable content-hashed JSONL ledger of
every decision. Zero LLM calls in the trust path: the parts of the gate that
must NOT be an LLM live here.
Why now (checked 2026-10-12)
The review/guard lane in the dsh ecosystem is entirely LLM-shaped:
answer-reviewer plugins re-read turns with a second model, dsh-gungnir
verifies goal completion from exit codes, and the closest action-guard
(cc-safety-net, 1.6k★ in the Claude ecosystem) blocks destructive commands.
Nobody gates what the agent writes deterministically — scope policy,
residue markers, or a durable record the CI can consume. That slot is empty,
and this plugin is the agent-pr-gate
engine living inside the runtime instead of waiting for the PR.
What it does
| Stage | Hook | Deterministic action |
|---|---|---|
| Before the write | tools/pre-execute | Scope policy: deny globs always win; a non-empty allow list means "only these". Denied writes never touch the disk |
| After the write | tools/post-execute | Residue scan of the written text: injection text aimed at reviewers, skipped tests, debug markers. warn (default) logs + ledgers; block also fails the write with corrective feedback |
| Always | JSONL ledger | One line per decision — tool, path, action, sha256 of content, matched pattern, residue kinds. agent-pr-gate CI can consume it as the session-side record |
One policy, two enforcement points
Point policyFile at your repo's agent-pr-gate.yaml and the live session
enforces the same allow/deny the CI gate enforces on PRs — all policy entries
are unioned. Inline allow/deny config works without a file.
Install
dsh plugin --profile default add dsh-pr-gate
From npm — no build step, no allowBuilds authorization; git installs are equally prebuilt (lib/ ships in the repo).
Configure
- id: dsh-pr-gate
name: dsh-pr-gate
config:
tools: [write, edit, str_replace_editor]
allow: [src/**, docs/**] # non-empty = only these
deny: ["**/*.pem", .github/workflows/**]
policyFile: "" # or e.g. agent-pr-gate.yaml
residueMode: warn # warn | block
ledgerPath: .dsh-pr-gate/session.jsonl
Develop
npm install
npm test # syncs the vendored agent-pr-gate core, builds, runs unit tests
The scope glob semantics and residue rules are vendored from the sibling
agent-pr-gate checkout (npm run sync; CI runs sync --check).
Roadmap
bashcommand write-heuristics (redirect/shell-write detection, warn-only)- Ledger → GitHub check-run bridge (session attestations surfaced in CI)
- Per-policy selection by agent preset, matching the CI gate's policy classes
- Claims-vs-diff at turn end (the agent's final summary vs. the session ledger)
Kill criteria
If dsh ships first-party deterministic write gating, or reviewer plugins converge on deterministic enforcement (they are all LLM-based today), this freezes at its last working version with a postmortem here. The engine stays owned by agent-pr-gate either way.
Comments
Loading…
Similar plugins
Commitment write-gate: operator-authored rules enforced before a tool call runs — a deterministic guard tier plus an LLM-judge tier, fail-closed by default, every block written to a contradictions log
★ 0
↓ 54/wk
dsh plugin --profile web add dsh-write-gatePermission mode between workspace-write and full access: shell commands are checked by deterministic rules and a subagent review; irreversible or system-level actions require explicit approval.
★ 0
dsh plugin --profile web add dsh-almost-full-accessby tancheng33
Runtime security gate on the tool pipeline: denies calls naming hosts outside an egress allowlist, redacts credentials from results at the canonical value rather than only the rendered content, and ap
★ 1
↓ 95/wk
MIT
TypeScript
Sep 22, 2026
dsh plugin --profile web add dsh-egress-guardby keyiadiannao
Pre-write reuse firewall for DeepSeek Harness: before the agent writes a new helper/service, surface the existing implementations that already cover that intent. Deterministic retrieval (no LLM) via t
★ 0
MIT
TypeScript
Aug 16, 2026
dsh plugin --profile web add dsh-code-reuse-firewallby better-er
Blocks edit/write writes that contain characters matching configurable regex rules, executed before the tool runs, with configurable messages and same-turn pwsh propagation; default rule matches full-
★ 4
↓ 369/wk
MIT
TypeScript
Oct 2, 2026
dsh plugin --profile web add dsh-write-rule-guardDeepSeek Harness approval guardian: a reviewer model decides permission requests before they reach you, with a System One decision-API backend (TypeSafe Jev, SiliconFlow, self-hosted Laya), a configur
★ 0
dsh plugin --profile web add dsh-auto-review-jev