“Change the third step” is vague. A comment attached to the exact sentence
preserves the context the agent needs to revise the plan correctly.

Select exact text → comment on several risks → send one review → approve when ready.

🎯 Why dsh-plannotator

Coding agents are good at producing plans, but a binary Approve / Reject
decision is too coarse for serious work. Architecture migrations, API changes,
security fixes, and rollout plans often need several independent corrections
before implementation begins.

dsh-plannotator turns DSH's native Plan Review into a compact gate plus a
responsive companion panel. On a wide screen, the conversation and review
occupy separate columns, so opening the panel never covers chat text. Collapse
it to a slim edge rail and reopen it without settling the request. Your comments
still travel through DSH's existing response flow as structured Markdown, so the
agent can revise the proposal in plan mode and ask for review again.

This is an unofficial integration inspired by
Plannotator.

🧰 Features

Comment on the exact claim—not “somewhere in the plan”

Drag over text for a precise annotation, or double-click a paragraph, list
item, heading, bold phrase, or code fragment. The review keeps the quote and
comment together within the current plan revision.

Review the whole plan in one pass

Collect multiple comments across compatibility, security, rollback, and tests;
add overall feedback; click an annotation in the review panel to return to its
source; then send one coherent review. This keeps the review compact and less
ambiguous than a sequence of detached chat messages.

Collapse the review without losing your place

The review panel can shrink to a blue edge rail while the compact composer gate
remains visible. Reopen either control to continue with the same comments and
overall feedback.

Return actionable feedback to the agent

Send feedback answers the real exit_plan_mode interaction. DSH records the
quoted plan text, each requested change, and the overall feedback in the tool
result and Session Log. The agent remains in plan mode and can immediately
produce a revised proposal.

Ask AI about the plan

Select plan text and choose ✦ Ask AI to stage the excerpt, or just type
into the Ask AI sidebar that sits permanently to the left of the plan
preview. The question travels with the plan text, the quoted
excerpt, and your earlier Q&A to a one-shot subagent of the reviewed session —
a fork that inherits the same model, workspace, and earlier conversation
turns (the plan itself always rides along verbatim, since it was submitted in
the still-open review turn) — that can inspect the repository with read-only
tools (read, grep, glob, web_search, web_fetch, probed against the
session's actual tool set). The answer renders inline as Markdown; follow-up
questions keep the thread's context, and Stop cancels a slow answer. The
thread survives navigating into the answering subagent and page reloads (it is
kept locally like annotation drafts), so an in-flight answer still lands when
you come back. Stop silently drops the pending question, while a host-side
cancellation stays visible as an error entry with Retry. The answering child
never modifies files, never rewrites the plan, and cannot delegate further.

Protect unfinished reviews

Unsent comments are saved locally in the browser, isolated by Session, pending
request, and plan revision. If you try to approve while feedback is still
pending, the plugin requires an explicit second confirmation instead of
silently discarding your work.

📦 Install

Install the GitHub bundle into the DSH Web profile, then restart dsh web:

dsh plugin --profile web add github:titanwings/dsh-plannotator#v0.1.4

The repository ships its built Host and Web bundles, so installation runs no
package build script and needs no allowBuilds entry. Pin a reviewed commit SHA
instead of the release tag when you need an exact source revision.

pnpm install
pnpm check

cd /path/to/deepseek-harness
pnpm dsh plugin --profile web add /path/to/dsh-plannotator

🔄 How it works

1. Ask the coding agent to create a plan in DSH Plan mode.
2. When exit_plan_mode reaches Plan Review, DSH shows a compact gate. On wide
   screens the review opens beside the conversation; narrower screens keep it
   closed until you choose Open review.
3. Select the exact text that needs work. Collapse and reopen the panel at any
   time without settling the review.
4. Add as many targeted comments as necessary, plus optional overall feedback.
5. Choose Send feedback. The agent receives one structured review and stays
   in plan mode.
6. Review the revision and choose Approve when it is ready to implement.

Chat about it dismisses the gate and returns to the ordinary composer.
Removing the plugin restores DSH's built-in Plan Review automatically.

Built for real coding plans

The screenshots above use a production-style authentication migration example,
not placeholder copy. The same workflow is useful whenever several plan details
must be correct before the first edit:

🧩 Compatibility and boundaries

• Designed for the DeepSeek Harness Web client and Node.js 22.19+.
• Claims only a valid, single-question DSH plan-review interaction. Other
  questions fall through to the built-in renderer.
• Reviews Markdown plans; it is not a general document editor, Git diff viewer,
  PR publisher, file tree, or the full standalone Plannotator SPA.
• Drafts live in the current browser's local storage. They are not cloud-synced
  and are intentionally rejected when the plan revision changes.
• At 1480px and above, the plugin reserves a 440–560px companion column beside
  DSH, so the panel and conversation never overlap. Narrower desktops use an
  on-demand drawer; phones use a compact bottom sheet.
• The companion panel is plugin-owned, not DSH's core details panel. It uses
  the stable Web #root mount boundary to reserve space and lets AppFrame
  reflow normally; it does not register in or rewrite the core details grid.
• The Ask AI channel is a loopback/trusted-host RPC channel (/dsh-plannotator)
  on DSH's shared Connection transport. Each question runs as a one-shot
  subagent (labelled plan-ask, visible in the session's subagent list) that
  forks the reviewed session — inheriting its model, composition, and completed
  conversation turns — under a read-only tool filter (a composition without the
  fork provider, or whose fork lacks the requested tool-filter/persona
  capabilities, falls back to a fresh child). Answers are unary (no streaming
  yet), capped at 32k characters so they always fit a follow-up's context, and
  the Q&A thread is kept in local storage per review — it survives panel
  navigation and reloads, unlike annotation drafts which reject stale plans.
• Feedback travels through DSH's existing response channel. No third-party
  service or telemetry is used.

🛠️ Development

pnpm typecheck
pnpm test
pnpm build

The browser bundle follows DSH's window.__ModuleLoader__ contract and treats
React, ReactDOM, and DSH UI primitives as platform modules, preserving one React
runtime.