dsh-verify

When the agent says "I fixed it", this runs it for real.

The umbrella package: it strings dsh-quota / dsh-crashlab / dsh-probe / dsh-timelab into one verification = run + assert + a readable report, and injects a "verify before declaring completion" rule into the system prompt.

Install

dsh plugin --profile "$env:DSH_PROFILE" add @alexzeng666/dsh-verify

The other four are declared in this package's dependencies and land in the profile automatically. Do not install them separately — this package's patch inserts them, and mounting them twice reports "tool already registered".

Tools

Tool	In one line
verify_run	Run a command; get exit code, stdout/stderr, and on a crash the stack and source lines
verify_assert_exit	Assert the exit code — a number, 0, or nonzero
verify_assert_stdout	Assert the output contains or lacks text; /regex/ supported
verify_assert_file	Assert existence, content, or sha256
verify_assert_json	Assert the value at a JSON pointer
verify_assert_screenshot	Pixel comparison against a baseline PNG — decodes PNG itself, no vision dependency
verify_report	Fold .verify/fragments.jsonl into one readable report

Assertions target the most recent verify_run unless you pin one with run_id (like V-001).

Flow

verify_run(cmd="npm test")
  └─ crash_helper.exe run           the command runs under the debugger
       ├─ exit code + stdout/stderr → files
       └─ crashed → minidump + StackWalk64 → call stack / module / source line
verify_assert_*                     read the latest run's fragment back and assert on it
verify_report                       fold fragments.jsonl into one report

verify_run goes through crash_helper rather than spawning the process itself, so that "run + attribution" come out of a single step. Otherwise the agent would have to start the process and then race to attach — and those few hundred milliseconds are enough for the target to die.

Return

Everything returns VerifyFragment. verify_report's detail carries total / by_result / by_source / recent.

System prompt

The plugin injects a block of verification discipline via ctx.systemPrompt.section({ name: 'verify:policy', order: 3040, ... }): run verify_* before declaring completion, state conclusions alongside evidence, and know the difference between error and fail. Turn it off with promptInjection: false in the plugin config.