dsh-plugin-stalegreen
English | 中文
The stalegreen freshness gate as a DeepSeek Harness plugin. Verification runs (pytest, pnpm test, tsc, eslint, cargo test, next build and about eighty others) become receipts, file edits become edit events, and when the agent ends a turn with "all tests pass", "tsc is clean" or "the build succeeds", the claim is checked against the latest receipt of that category:
| Evidence | Verdict | What happens |
|---|
| Passing run, no edits since | FRESH | the turn ends |
| Passing run, files changed since | STALE | one more step, with the receipt and the edited files |
| Failing run | FAILED | one more step |
| Run whose result was hidden by a pipe or a suffix | MASKED | one more step, asking for an unmasked rerun |
| No run at all | NONE | the turn ends (strict mode: one more step) |
DeepSeek Harness's own goal system says it has no independent evaluator. This plugin is one: deterministic, local, and citing a receipt for every verdict. Zero tokens, zero network, zero telemetry.
Install
dsh plugin add dsh-plugin-stalegreen
The bundle inserts one plugin row, stalegreen, into the profile. Options go on that row in the profile's cordis.patch.yml:
- id: stalegreen
config:
policy: advisory # record verdicts, never steer
mode: strict # deny masked verification commands
The receipts, edit events and verdicts live under ~/.stalegreen/sessions/<session id>/, the same store the stalegreen CLI reads: npx stalegreen check shows the last session's claims and evidence, npx stalegreen receipt r-0017 a run's receipt and log tail.
How it hooks in
tools/pre-execute: a verification command gets a pending receipt; in strict mode a masked one (| tail -5, || true, 2>/dev/null) is denied with the instruction to run it without the pipe. The harness's pre-execute waterfall cannot rewrite a call, so unlike the Claude Code and Codex hooks nothing is rewritten here; the bash tool keeps the full output of a long run in a file, which limits what a pipe can hide.tools/post-execute: the bash tool's output becomes a receipt, using the runner's own summary line and the [exit code: N] marker; edit, write and str_replace_editor calls become edit events.agent/turn-stopping: the last assistant message is read for green claims. A stale, failed or masked claim steers one more step with a message such as:
stalegreen: "all tests pass" is stale. Receipt r-0017 (`pnpm test`, 41 passed, 14:02:11) predates 3 later edits:
src/routes/pay.ts (14:05:40), src/lib/hold.ts (14:06:02), src/lib/hold.test.ts (14:06:31)
Rerun `pnpm test` and report the result, or state explicitly that the tests were not rerun after these edits.
A category is steered at most once per turn, so the agent can never be trapped.
What it is not
Not a judge of intent: it checks whether the evidence behind a claim is fresh and complete. Not a replacement for the runner's judgement: a pass is the runner's own summary, never an inference from silence alone.
License
MIT. Copyright (c) 2026 Pavan Gupta.
