dsh-pr-watch

dsh-pr-watch

Reports what changed in your authored pull requests since your last check — merged, closed, or gone stale.

dsh-pr-watch is a dependency-free DeepSeek Harness (dsh) plugin that exposes one agent tool, pr_watch. It tracks every pull request you authored, across all repositories — including ones you have never cloned — and tells you only what changed since you last looked.

Status: pre-release. The implementation is complete: the plugin registers exactly one tool, pr_watch, and the whole pipeline is covered by the test suite. Nothing is published to npm yet, so the install command below does not work today. It is included so the intended path is clear, not because it is ready.

• Design: docs/superpowers/specs/2026-09-10-dsh-pr-watch-design.md
• Plan: docs/superpowers/plans/2026-09-10-dsh-pr-watch.md
• What is left: Roadmap

What works today

Installing this plugin today registers pr_watch. What is still missing is the release itself, not the tool.

Test counts are deliberately not listed here. They went stale within one pull request, and CI already reports them per commit — a number in prose is a claim nobody re-checks.

Why

An agent has no memory of your last check, so every status question gets re-derived from scratch. That is fine for one pull request and useless for thirty. The friction is not seeing your pull requests — it is re-checking the same ones over and over and having to remember what the state was last time.

pr_watch closes that gap by storing a snapshot and reporting deltas.

Features

Implemented, pending release. Everything below is built and covered by the test suite. It is not on npm yet, so there is no way to install it other than from a checkout — see Installation and What works today.

How it works

The naive approach — fetch open pull requests, diff against the snapshot — cannot work. The moment a pull request merges it leaves the open list, so a bare diff cannot tell "merged" from "closed unmerged", and an empty result looks identical to a failed gh call.

So pr_watch fetches in two phases:

1. Enumerate every open pull request you authored, in one call.
2. Resolve each snapshot entry that left that set, individually, to learn its terminal state.

Only pull requests that actually changed incur a second call — normally zero or one per check. An entry that has left the open set but cannot be resolved is retried on each check until its outcome is known, so it adds one call per check in the meantime.

Installation

dsh plugin --profile PROFILE add dsh-pr-watch

Replace PROFILE with the dsh profile you actually run — the tool only appears in that one. The Desktop app manages a profile of its own, so from its terminal omit the flag entirely and let it target the active profile:

dsh plugin add dsh-pr-watch

Not yet published, so neither command works today. They are the intended path once the package is released.

Until then, install from a checkout. dsh plugin forwards its arguments to pnpm in the profile directory and anchors a relative path spec to the directory you ran it from, so a local link works on every platform:

pnpm install && pnpm run build
dsh plugin --profile PROFILE add link:.

Run that from the repository root, and use the same profile you would have used above.

Requires the GitHub CLI (gh) on your PATH, authenticated:

gh auth login

dsh-pr-watch stores no credentials of its own. It relies entirely on gh's own keyring-backed authentication.

Usage

Ask the agent to check your pull requests, or call the tool directly.

{
  "staleDays": 14,
  "all": false
}

Typical output:

✅ Merged (2):
  octo/repo#41 — Add Russian locale (last activity 1 day ago)
  octo/repo#38 — Fix broken link (last activity 3 days ago)

⏳ Became stale (1):
  octo/repo#29 — docs: clarify install steps (last activity 21 days ago)

🆕 Newly noticed (1):
  octo/repo#44 — Add MCP server entry (last activity today)

Run it a second time and only genuine changes appear. A merge reported today is never reported again.

Parameters

Configuration

v1 exposes everything through the tool's two parameters above. It reads no settings file, so there is nothing else to configure yet.

Two further keys appear in the design doc — ignoreRepos and a user-settable snapshotPath — but they are deferred to v2 and not implemented. Honouring them needs a settings loader that does not exist yet, and ignoreRepos additionally needs a rule to evict ignored entries from the snapshot: an entry dropped from enumeration but kept in the snapshot would never resolve, so every check would report it as unresolved forever. See the plan's refinements section for the full reasoning.

Behaviour notes

• Staleness fires on transition, and only once. A pull request that has been stale for weeks is reported the first time it crosses the threshold, then stays quiet. If it sees new activity, the staleness clock resets and it can be reported again later.
• A failed enumeration writes nothing. If the one gh search call fails, or its output cannot be read, the snapshot is left untouched so pending changes are not silently marked as seen. A failure in the resolve phase is narrower: every outcome that was determined is still recorded, and the entry that could not be resolved stays OPEN and is retried on the next check — without being announced again in the meantime.
• A reopened pull request is announced again. Closing is not final: if a closed pull request is open again at the next check, it is reported as newly noticed and resumes normal tracking. A merge is treated as final, since it cannot be undone.
• The snapshot is written atomically (temporary file, then rename), so an interrupted write can never leave a truncated file that reads as "everything vanished".
• A corrupt snapshot is quarantined, never discarded. It is moved to snapshot.json.corrupt-<n> and the tool says so in its output, including when the check fails afterwards — the move is irreversible, so that message is the only chance to report it.

Known limitations

A pull request opened and merged between two checks is invisible. It is absent from the open set (it already merged) and absent from the snapshot (it did not exist when the snapshot was taken), so neither phase sees it. This is a direct consequence of enumerating by current state rather than by activity window, and it only becomes likely if you check infrequently. The fix — an activity-window query — is recorded in the design doc as a v2 candidate.

New comments and reviews are not reported. An outcome like "merged" is not the same as "a maintainer replied asking for changes", and the latter is often what should change your next action. It is deferred rather than overlooked: it needs a separate data source and a per-entry comment cursor.

Reopening is taken at the enumeration's word. Whether a closed pull request is open again is decided from the same gh search listing as everything else. If that listing is stale and names a pull request that is still closed, it is announced as newly noticed, and when it clears, reported closed a second time. Confirming every reappearance with gh pr view would cost a lookup per candidate to guard against a rare disagreement, so the listing is trusted instead.

A pruned outcome is remembered for 180 days, then forgotten for good. Entries are dropped 90 days after reaching a terminal state, so the snapshot stays bounded — but the record being dropped is also the only evidence the outcome was reported. A pruned entry therefore leaves a two-field memory behind (the state and when it was pruned), and a departure that later resolves to that same state is not announced again while the memory lasts. That is what keeps "a merge is never reported twice" true across pruning. Once the memory ages out the key is genuinely gone: if the enumeration then lists it as open (a stale index can), it is announced as newly noticed, and a later departure can report the outcome a second time.

CI status is not reported.

One snapshot per GitHub identity. Running under a second account would share a snapshot and produce spurious "newly noticed" entries.

Roadmap

The design is complete and the work is broken into fifteen tasks in the implementation plan.

The remaining work is a release, not a feature: the package has not been published, so dsh plugin --profile web add dsh-pr-watch cannot resolve yet.

Development

pnpm install
pnpm run build        # tsdown → lib/
pnpm test             # vitest — no network access
pnpm run verify:live  # opt-in: drives the real gh against GitHub
pnpm run lint         # oxlint
pnpm run typecheck    # tsc --noEmit
pnpm run format       # prettier --write .

The test suite makes no network requests and never runs gh. delta.ts and tools/watch.ts are pure — no filesystem, no network, no clock — and every IO module takes its dependencies as arguments, so the whole pipeline is driven by fixtures and injected executors rather than mocks.

That isolation has one blind spot, and it is not theoretical: a fixture written with the same wrong assumption as the code cannot catch it. gh pr view was asked for a field it does not have, so every resolve failed against the real CLI while every mocked test passed. pnpm run verify:live closes that gap. It needs gh authenticated, makes real API calls, and checks what fixtures cannot: that the argument lists are accepted, that live payloads map, and that a full check writes a snapshot and stays quiet on the second run. It is a stretch to run before a release, not on every commit.

To check the built artifact rather than the sources, run pnpm run build and then load lib/index.js: it must export name, inject, apply, and a tools array of exactly one entry. pnpm pack should produce a tarball containing only lib/index.js, cordis.patch.yml, package.json, README.md, LICENSE, and SECURITY.md.

On Windows, pnpm organises node_modules with symlinks, which needs Developer Mode or an elevated shell; without it, installs fail with EPERM. The npm cache and store are redirected inside the repository by .npmrc, so nothing is written outside the checkout.

License

MIT