dsh-plannotator

Browser review for the DeepSeek Harness terminal profile: read a plan, a document, or a code change in Plannotator's interface, then send what you think back to the agent — without leaving the TUI.

/plannotator-plan      review this session's plan
/plannotator-review    review the worktree or a pull request
/plannotator-annotate  annotate a file, a folder, or a URL
/plannotator-last      annotate the agent's last answer

Each command opens the browser, waits for your decision, prints one line in the terminal, and forwards your words to the model as a steering message. The agent can then act on them in the same session.

Why commands instead of plan-mode integration

Plan mode keeps its own review path. This plugin adds a second, on-demand one, because that path is not always reachable: a session on a remote host without a forwarded port cannot open a browser at all, and a reader often wants to review something that is not the plan they just wrote. Nothing here replaces or patches the shipped plan-mode review; if the profile cannot use this plugin, plan mode still works exactly as before.

Requirements

• DeepSeek Harness 0.1.5-rc.1 or newer, 0.1.6 or older.
• The plannotator CLI on PATH, or pointed at by configuration. Install it the way its own documentation describes; this plugin never downloads it.
• A browser on the machine that runs the CLI. For a remote host, either forward the port or set tailscale: true and reach the session over the tailnet.

Install

dsh plugin --profile tui add @sagmans/dsh-plannotator

The profile loads every bundle listed in its package.json; confirm the bundle is listed there after the install. From a checkout, install the directory instead:

dsh plugin --profile tui add link:/path/to/dsh-plannotator

Commands

| Command | Runs | Notes | | --- | --- | --- | | /plannotator-plan [file.md] | annotate <plan> --gate --json --require-approval | With no argument, the newest plan in the session is snapshotted and reviewed. | | /plannotator-review [options] [PR_URL] | review <options> --json | Reviews the worktree by default. | | /plannotator-annotate <file\|folder\|URL> | annotate <target> --gate --json --require-approval | The target must exist, or be an http(s) URL. | | /plannotator-last | annotate-last --stdin --gate --json | The message is piped on stdin and never written to disk. |

review accepts --git, --gitbutler, --base <ref>, --diff-type <type>, --local, --no-local, --no-git-remote-check, and --tailscale. annotate accepts --markdown, --no-jina, --app, --static, --render-html, and --tailscale. Any other flag is refused with usage text instead of being passed to the CLI.

How a decision comes back

The plugin reads the CLI's own verdict, never the wording of a message:

| Surface | Verdict source | Feedback field | | --- | --- | --- | | annotate | exit 0 approved, 1 not approved, 2 gate failure | feedback | | review | decision in the JSON record | message | | annotate-last | decision in the JSON record | feedback |

An approval is reported as an approval. Notes attached to an approval are forwarded as non-blocking guidance, which is how Plannotator frames them; set approveNotes: ignore to keep them out of the conversation, or steerFeedback: false to keep every review local.

Configuration

Every key is optional. Restate the ones you keep when you patch the row, because an id-targeted patch replaces the whole config block.

| Key | Default | Meaning | | --- | --- | --- | | binary | '' | Absolute path to the executable. Empty means $PLANNOTATOR_BIN, then PATH. | | dataDir | '' | Where plan snapshots are written. Empty means $PLANNOTATOR_DATA_DIR, then ~/.plannotator. | | reviewTimeoutMs | 0 | Hard ceiling on one review. 0 waits as long as the browser session lives. | | tailscale | false | Publish each review over the tailnet. | | steerFeedback | true | Forward reviewer feedback to the model. | | approveNotes | steer | steer or ignore notes that arrive with an approval. | | allowUntrustedBinary | false | Accept an executable this user does not own, or that others can rewrite. | | maxInputBytes | 1048576 | Largest plan or transcript handed to a review. |

# ~/.dsh/profiles/tui/cordis.patch.yml
- id: plannotator
  config:
    reviewTimeoutMs: 900000
    tailscale: true

Security

• No shell. The executable is spawned with an argv array and shell: false, so no path, branch name, or reviewer's word is ever parsed as syntax.
• Vetted executable. The resolved binary must be executable, owned by this user, and not writable by others; a file that fails is refused by name unless allowUntrustedBinary is set.
• Private plans. A plan snapshot is written under <dataDir>/dsh/plans/<session-id>.md with mode 0600, its directory with 0700, and it is renamed into place so a reader never sees a partial file. Plans stay out of your repository.
• Allowlisted flags. Each subcommand has its own flag list. An unknown flag fails the command; it is never forwarded.
• Bounded capture. Each stream is capped, and a cancelled or over-long review is terminated with SIGTERM and then SIGKILL after two seconds.
• One review at a time. A second command is refused while a review is open, and unloading the plugin stops the one that is running.
• No runtime dependencies. The bundle uses harness packages it receives by injection and Node builtins only.

Development

pnpm install
pnpm run check     # typecheck, unit tests, build, built-artefact tests, pack smoke
pnpm run drive     # drive a real TUI in a PTY (see tools/pty-drive.mjs --help)

tests/fixtures/fake-plannotator.mjs stands in for the CLI. It implements the documented contract — argv, stdin, JSON records, strict exit codes — so the suite covers the whole spawn path without a browser.

End-to-end check with a real browser

The suite stops at the CLI boundary. To prove the whole loop, mount the bundle in a scratch profile, make the CLI publish its URL instead of opening a browser, and drive that page:

Copy the profile you already work in — it carries the bundles and the patch that register your model route and credentials — then add this plugin to the copy. Dropping those rows leaves the profile without an adapter for the model its settings name, and the surface then answers from whichever provider it does have, which is a failure that looks like a plugin problem and is not one.

PROFILE=tui-plannotator-e2e
cp -R ~/.dsh/profiles/tui ~/.dsh/profiles/$PROFILE
cd ~/.dsh/profiles/$PROFILE
# Add the plugin to dsh.profile.bundles and to dependencies, then:
dsh plugin --profile $PROFILE install

# Drive the surface, and let the CLI report where its review page lives.
PLANNOTATOR_SKIP_BROWSER_OPEN=1 \
PLANNOTATOR_READY_FILE=/tmp/plannotator-ready.json \
PLANNOTATOR_DATA_DIR=/tmp/plannotator-data \
  pnpm run drive -- --profile tui-plannotator-e2e --cwd /tmp/scratch \
  --prelude "/plannotator-annotate notes.md" --seconds 120 \
  --expect "annotation feedback sent to the agent"

Open the URL from /tmp/plannotator-ready.json, annotate, and send feedback. The terminal shows one notice, and the reviewer's words appear in the transcript as a [plannotator] user message — which is what the model then answers.

Symptom to recognise: a turn that fails with Insufficient Balance or no adapter registered for provider "<route>" right after the notice means the profile in use cannot reach the model its settings name. It is a profile composition fault, not a review fault: the reviewer's words did arrive.

License

MIT