dsh-review-squad
中文说明
A parallel multi-role code review squad for DeepSeek Harness.
One /review command dispatches independent reviewer subagents — security, correctness, tests, and style — that gather evidence themselves (read the diff, trace callers, check conventions) in strict read-only mode, and merges their work into one severity-grouped report. Each reviewer can optionally run on its own model: put your strongest model on security and a cheap fast model on style.
Why
Existing dsh git plugins stop at diff viewers and worktree cleanup. Review is the workflow other agent ecosystems keep voting for with their stars (Claude Code's official security review, compound-engineering, Superpowers), and dsh's subagent registry — parallel spawned children, per-child model routing, tool filters — is exactly the right primitive for it.
Install
# into a profile (default profile is `web`)
dsh plugin --profile web add dsh-review-squad
Or straight from a GitHub checkout:
dsh plugin --profile web add github:<owner>/dsh-review-squad
Requires dsh ≥ 0.1.1-rc.2 (the base bundle with spawn subagent provider and tool-str-replace-editor — both ship in @deepseek-ai/dsh-base).
Use
Slash command (you)
/review # full squad on uncommitted changes (git diff HEAD)
/review staged # staged changes only
/review commit abc1234 # one commit
/review path src/lib # one path
/review security tests # only these reviewers
/review --style # everyone except style
/review stop # stop this session's running review jobs
/review watch the retry logic # extra focus text for every reviewer
Tool (your agent)
The model gets a code_review tool and can run the squad on its own work before calling a task done:
| parameter | meaning |
|---|
scope | diff (default) · staged · commit:<sha> · path:<dir-or-file> |
focus | extra concerns for every reviewer |
reviewers | comma-separated keys, --key excludes |
Background execution
/review runs the squad as a background job by default (commandBackground: true): the command returns instantly with a job id, so the chat input is never locked while the reviewers work. When the job settles, the session agent is notified; read the report with job_output (or just ask the model "read the review job result"). Set commandBackground: false (settings.yaml review-squad: section) to wait inline instead. The code_review tool always runs foreground.
The report
# Code review report — uncommitted changes
Reviewers: security, correctness, tests, style · findings: 5
| severity | count |
|---|---|
| CRITICAL | 1 |
## CRITICAL
- **SQL injection in user lookup — `src/db/users.ts:42` *(security)***
## Reviewer verdicts
- **security**: Would block merging: yes, because of the injection.
...
Failed reviewers are reported per-reviewer (stop reason + diagnostic + partial output) instead of dragging the whole run down. Reports are deterministically aggregated — no extra LLM call — and hard-capped (maxReportChars, INFO→LOW→MEDIUM sections dropped first).
How it works
/review ──┐
├──► engine: pick reviewers ──► ctx.subagents.start("spawn", …) × N (parallel)
code_review ┘ │ per reviewer:
│ · prompt = persona + evidence guide + read-only rules + output format
│ · agentOptions = optional per-reviewer provider/model/effort
│ · toolFilter.deny = str_replace_editor, write, edit, subagent, subagent_fork
▼
results collected (timeout + cancel per reviewer) ──► deterministic merge
Reviewers are spawned dsh children: fresh context, their own tools, your workspace. They find the evidence themselves — git diff, reading files, tracing code — which is what makes the review deep instead of a prompt over a diff. Read-only is enforced twice: prompt discipline and a toolFilter deny list.
Configuration
Override the review-squad row from a later patch layer (~/.dsh/profiles/<profile>/cordis.patch.yml or a --patch overlay). Patch rows replace whole configs — restate everything you keep.
| field | default | meaning |
|---|
provider | spawn | ctx.subagents provider used for reviewers |
commandName | review | slash command name |
toolName | code_review | tool name |
toolFilterDeny | ['str_replace_editor','write','edit','subagent','subagent_fork','workflow','ralph','send_message','web_search'] | tools reviewers lose — includes every delegation tool (they must not spawn anything) and web search (cost); unknown names are healed automatically |
timeoutMs | 600000 | per-reviewer wall-clock budget |
maxConcurrent | 2 | reviewers run in batches of this many (gentler on API limits) |
maxReportChars | 24000 | report hard cap |
language | en | descriptive text language (severity/file/line markers stay English so parsing works) |
reviewers | built-in four | the squad (see below) |
Custom squad — the "strong brain, cheap hands" pattern
- id: review-squad
name: dsh-review-squad
config:
language: zh
reviewers:
- key: security
role: Security reviewer
instructions: Hunt for injection, authz gaps, committed secrets, unsafe crypto.
agentOptions:
provider: deepseek-official
model: deepseek-v4-pro
reasoningEffort: max
- key: style
role: Style reviewer
instructions: Naming, duplication, complexity, conventions.
agentOptions:
provider: my-cheap-gateway
model: glm-4.6-flash
agentOptions needs a provider that supports the agentOptions subagent capability (the built-in spawn provider does).
Settings page (Web UI)
The plugin ships its own "Review Squad" section in the dsh web Settings page: edit each reviewer's model route (provider / model / reasoning effort / max tokens), report language, timeout, report budget, and the tool deny list. Saving goes through dsh's settings.update channel into the review-squad: section of settings.yaml and applies live (no restart).
Runtime overrides — settings.yaml (hot-reloaded)
Every field except the structural ones (commandName/toolName) can also be changed without restarting dsh via the review-squad: section of $DSH_HOME/settings.yaml (~/.dsh/settings.yaml). Values there override the bundle config at every invocation:
# ~/.dsh/settings.yaml
review-squad:
language: zh
timeoutMs: 300000
reviewers:
- key: security
role: Security reviewer
instructions: Hunt for injection, authz gaps, committed secrets, unsafe crypto.
agentOptions:
provider: deepseek-official
model: deepseek-v4-pro
reasoningEffort: max
- key: style
role: Style reviewer
instructions: Naming, duplication, complexity, conventions.
agentOptions:
provider: my-cheap-gateway # any provider configured on the Models page
model: glm-4.6-flash
Omit the section to fall back to the bundle config. (The bundle-config reviewers example in the patch layer above does the same thing at load time; prefer settings.yaml for day-to-day tuning.)
Compat notes (verified)
- Verified end-to-end on dsh 0.1.1-rc.2 (Windows,
deepseek-v4-flash): real dispatched reviewer via spawn, per-reviewer agentOptions routing honored by the in-process driver, findings parsed and aggregated from real model output.
- On 0.1.1-rc.2 the
spawn provider does not declare the agentOptions capability (it is honored at the driver level). The plugin pre-checks capabilities.agentOptions and, if a runtime rejects the route, retries the reviewer on the session's model and states that in the report — per-reviewer routing degrades loudly, never breaks the run.
toolFilterDeny names are validated fail-loud by dsh; the defaults all exist in the base bundle on 0.1.1-rc.2.- Overriding the plugin row from a
--patch overlay or profile cordis.patch.yml uses the direct row form (- id: review-squad + config:), not - insert: (inserting a second row with the same id fails the boot).
Development
pnpm install
node --test # 70 unit/integration tests, no dsh runtime needed
Smoke-test the composition without a running server:
dsh plugin --profile dshrs-dev add ./dsh-review-squad
dsh --profile dshrs-dev --dump-config | grep review-squad
Note: while another dsh web instance is running, a second instance can hang at boot (shared $DSH_HOME); stop the first one before booting the dev profile — or point DSH_HOME at a scratch directory for fully isolated runs.
Submit to awesome-dsh-plugin
docs/awesome-entry.yml is the ready-made YAML entry (category git). Fill in your owner, add the dsh-plugin topic to your repo, and follow the one-entry-per-PR rules of awesome-dsh-plugin.
Feedback & contact
Found a bug or want a feature? Open an issue or start a discussion. For direct contact: luomeiy@outlook.com · https://github.com/luomeii
License
MIT
