DSH Super-Agent Orchestrator
Put auditable admission, observe mode, and human approval in front of autonomous CLI coding work.
- Source repo
- vicfei/dsh-super-agent
- Stars
- ★ 26
- Last updated
- 1d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 67/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code · OpenAI API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer maintaining several repositories wants to review which requests the agent would accept in observe mode before enabling automatic work in stages.
- Not a fit if
- Teams unwilling to self-host a daemon and local database
- Teams whose workers must access the network inside the sandbox
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20+0.6 ★/day over 7 days
- Source review
- 67/100 · Some gaps
What does this agent do, and when should you use it?
This is an independent implementation of the DeepSeek Super-agent orchestration design for CLI coding agents such as Claude Code, Codex, and opencode; it does not require DSH or a DeepSeek model. Triggers from GitHub, Feishu, Slack, WeChat, Telegram, and scheduled routines pass through a deterministic admission gate, with an optional OpenAI-compatible assessment layer for inferred requests. A SQLite engine stores decisions, tickets, source cursors, and an outbox; observe mode records would-accept decisions without starting work. Accepted tickets run through a configurable CLI subprocess in an isolated Git worktree, can be delivered as pull requests, and can be subject to approval, line-count, and optional CI checks before auto-merge. It is deployed as a Node.js CLI and watch loop, with a local panel for pending work and audit records.
When dsh watch or a one-shot poll runs, GitHub and configured IM Sources and routines read new messages or scheduled tasks and pass them through the Fast Path using the admission table, allowlist, and explicit/inferred strength; inferred candidates may be assessed by an OpenAI-compatible /chat/completions endpoint, with errors failing closed. Observe mode writes a shadow decision without executing work; auto mode can create a Ticket. dsh work hands a ticket to a configured CLI Executor in a per-ticket worktree and sandbox, saves a report, and can push a branch and open a PR. The Engine stores state and events in SQLite, the Panel shows items awaiting a person, and the Watchdog can send reminders but cannot admit work. When enabled, the Merge Gate checks approval, head SHA, line limits, and optional CI before merging a PR.
- A developer maintaining several repositories wants to review which requests the agent would accept in observe mode before enabling automatic work in stages.
- An operator running coding agents across machines wants GitHub and chat messages to share one ticket flow, with pending decisions approved from a phone over IM.
- A team building a contribution bot for its own repositories needs isolated worktrees, workers that do not deliver after failures, and approval-gated PR merging.
- A maintainer who needs scheduled reports from agent-session logs can trigger routines and deliver the reports as PRs for human review.
- A team that wants repository conventions to persist across tickets can enable two-tier memory, recording candidate facts in observe mode before auto mode injects them into later prompts.
How do you install or deploy this agent?
Use Node.js 20 or later, Git, and GitHub access credentials for the target repository. The README documents both npm-based startup and installation from a source clone. The init wizard writes policy and executor configuration, with observe mode as the default.
How do you use this agent?
First-time setup and launch with npm:
What are this agent's strengths and limitations?
- Host code enforces admission, so messages and quoted documents cannot grant themselves permission; observe mode lets operators gather decision evidence first.
- The Executor is a subprocess boundary and can run Claude Code, Codex, opencode, aider, or a custom shell command, with a separate worktree per ticket.
- SQLite persists tickets and source state for restart recovery; the PR merge gate can check approval against the current head SHA, changed-line limits, and optional CI.
- Sources include GitHub, Feishu, Slack, WeChat, and Telegram, and inferred-request assessment can use an OpenAI-compatible endpoint.
- Operators must run the Node.js daemon, configure policy and executor files, and provide GitHub credentials; each enabled IM channel also needs its own app or bot setup.
- LLM assessment requires a user-configured OpenAI-compatible endpoint and credentials; the README does not state hosted-model pricing.
- The default sandbox requires bwrap on Linux or sandbox-exec on macOS, denies worker network access by default, and limits visible environment variables.
- Auto-merge depends on GitHub approval identity and PR checks; the README recommends GitHub App identity for a sole operator.
How does this agent compare with similar options?
The README compares this project with Baton, Emdash, Agent Orchestrator, Claude Squad, and the DSH-only official dsh-superagent-prototype. It positions this project as using any CLI subprocess as a Worker, while the official prototype has deeper DSH-session integration. The README says the other orchestrators lack formal admission control and notes that their capabilities can evolve.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| DSH Super-Agent Orchestrator This agent | 67 · Some gaps | CLIFree + model costs | ★ 26 | 1d ago | TypeScript | — |
| Foremerge | 76 · Good | CLIFree | ★ 410 | 8d ago | Rust | Codex · Claude Code |
| Bernstein | 74 · Some gaps | CLIFree + model costs | ★ 1.2k | today | Python | Codex · Claude Code · OpenAI API |
| Superset | 66 · Some gaps | Desktop appFreemium | ★ 16k | today | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Trust has genuine design backing: admission tables, explicit/inferred binary strength, observe→auto ladder, double-key confirmation for destructive acts, and a merge gate requiring an allowed-list non-bot approval pinned to head SHA are all concretely described in README and the design doc, and tests cover digit shape, double-key and expiry semantics, so user_confirmation earns 3. least_privilege is only 2: the executor runs arbitrary CLI subprocesses, sandbox default-on is asserted in README but no sandbox implementation source is in the packet; blockedCommands and a 30-minute process-group timeout are pluses. data_flow_transparency is 2: README describes data flows (GitHub/Feishu/Slack/WeChat/Telegram, SQLite, OpenAI-compatible endpoint) but does not systematically state what data is sent to the external LLM endpoint. sensitive_data_handling is only 1: GITHUB_TOKEN, App private key, WeChat login credentials and IM tokens are involved; README mentions credentials in the engine db and hourly installation tokens, but there is no evidence of encryption, redaction or key handling in logs. dependency_security is 1: dependencies include better-sqlite3, Slack/Feishu SDKs and qrcode, package.json uses ^ ranges, and no audit or vulnerability-mitigation note exists; CI only runs npm ci and tests. external_effects is 2: side effects (open PR, merge, send messages) are gated and auto-merge defaults off, but the surface is broad. rollback is 2: stop is final, PRs are revertable and promote produces a PR rather than a direct merge, but no rollback path for already-merged changes or sent messages is shown. source_attribution is 3: it explicitly states it is an independent community project, not affiliated with DeepSeek, with links to the article and prototype and a Design credits section.
README status table, version (package.json 1.5.0) and design docs are broadly consistent, but the 'Evidence ledger + stream' row appears twice in README and some version labels (v0.6 listed twice, v1.4/v1.5 mixed) are muddled — a pointable inconsistency. dependency_availability is only 1: it needs Node>=20, the native better-sqlite3 module, bubblewrap sandbox and several IM SDKs; README gives no offline or degraded path, and a CI comment notes the lockfile once pointed at a Tencent-internal mirror, showing environment sensitivity in dependency fetching. failure_messages is 2: the CLI has a unified die(), doctor exits 0/1/2, and App auth failures give actionable errors (tests assert /is the App installed/), but there is no full evidence of error text for executor timeouts or LLM failures.
README lists four real use cases (observe-first autonomy, a personal multi-machine fleet, a contribution bot, a scheduled report engine) and targets operators of 'any CLI coding agent'. capability_boundaries is 3: the design doc explicitly lists what IM can and cannot approve (GitHub merge approvals, policy changes and work creation are all excluded) and states non-goals. trigger_precision is 3: explicit/inferred binary, strict digit shape ('1.', 'ok 1', '12' all fail), group digits treated as chat, fail-closed @-mention detection, all covered by tests. environment_fit is 2: multiple IM channels and executors are supported, but it hard-depends on Node 20+, native modules and bwrap; Windows/WSL adaptation is only mentioned in README without concrete evidence.
README is well structured (Why/How/Status/Quick start/channels/design) but very long, with the status table duplicating body text; docs/designs holds design docs — readable but bloated. install_notes is 2: there is an npx init wizard, non-interactive templates, a manual clone path and per-channel one-time setup steps, but no troubleshooting for failed installs. naming_stability is 2: package name @vicfei/dsh-super-agent, bin dsh-super-agent and the command set are stable, but version labels drift in README. examples_and_faq is 2: examples/policy.json, executor.json, a COOKBOOK and demo GIF exist, but no standalone FAQ. known_limitations is 2: README notes WeChat media as placeholders, unreliable group @-parsing and the single-operator App-identity need, but these are not consolidated into a limitations section. license is 3: a complete MIT LICENSE file, consistent with package.json and README. versioning_changelog is 2: semantic versions and 'since' labels exist, but no standalone CHANGELOG file. maintenance_responsibility is only 1: the publisher is unverified, and README states no maintainer commitment, support channel or security-disclosure process beyond CI/publish workflows.
Outputs include gate decision lines (ACCEPT/WOULD-ACCEPT/IGNORE), ticket stages, a panel, PRs and IM receipts — usable in form, but there is no full evidence of structured or machine-readable reports. marginal_value is 2: it fills the 'admission control' gap other orchestrators skipped and the value proposition is concrete, but many orchestrators exist and its value depends on the user already having GitHub/IM infrastructure, so marginal value is moderate. cost_benefit is only 1: deployment needs Node 20+, native modules, bwrap, a GitHub App or PAT, several IM app registrations and a long-running daemon — high operational cost — while the benefit concentrates on admission and audit; README does not quantify cost/benefit.
Most README claims trace to design docs, test files or concrete commands (e.g. dsh events verify, doctor exit codes), but figures like '290 tests' and 'real merged PRs' cannot be independently verified within the packet. cross_source_corroboration is 2: README, design docs, tests and CI workflows corroborate each other (double-key, digit shape, App auth), but key implementation sources for sandbox, memory and ledger are not in the packet and cannot be cross-checked. fact_inference_separation is 2: README largely separates design intent from implemented status (status table with versions), but assertions like 'Every line in this GIF is real output' sit alongside inference, and muddled version labels weaken the separation.
- The executor runs arbitrary CLI subprocesses; sandbox default-on is only asserted in README and no sandbox implementation source is in the packet — review the actual sandbox and blockedCommands constraints before production use.
- GITHUB_TOKEN, GitHub App private key, WeChat login credentials and multiple IM tokens are involved; there is no evidence of encryption, redaction or key handling in logs — verify credential storage and rotation yourself.
- Dependencies include the native better-sqlite3 module and several third-party SDKs, use ^ ranges and have no audit note; a CI comment shows the lockfile once pointed at an internal mirror — check dependency provenance before installing.
- The publisher is not verified by FollowAgents and README states no maintenance commitment or security-disclosure process; assess maintenance sustainability before upgrading.
- README version labels and the status table contain duplication and drift (e.g. the Evidence ledger row repeats, v0.6 appears twice), and some capability claims cannot be independently verified within the packet.
FAQ
Do I need a DeepSeek model or DeepSeek Harness?
/chat/completions endpoint, or be left disabled.Does observe mode start agent work?
Can workers access the network or GitHub credentials?
GITHUB_TOKEN is not passed to a worker. A worker's own auth variables must be listed in passEnv, and secret-looking names are refused.How does auto-merge decide whether a PR can merge?
What runtime and deployment does it require?
Is DSH Super-Agent Orchestrator free?
Is DSH Super-Agent Orchestrator open source?
Is DSH Super-Agent Orchestrator safe to use?
Is DSH Super-Agent Orchestrator still maintained?
Who makes DSH Super-Agent Orchestrator?
Thanks — we use this to improve the page.