dsh-mission

A long-horizon autonomous mission runtime for DeepSeek Harness — lets an
agent drive a long-term objective over days: a cross-session, independently
verifiable, replan-on-failure, crash-recoverable autonomous control loop.
The dependency DAG is one of its driving mechanisms, not the runtime itself.

English | 中文

What it is

dsh-mission is a pure-TypeScript cordis plugin for DeepSeek Harness that
provides the long-horizon autonomous mission runtime dsh lacks: a durable,
verifiable, crash-recoverable control loop that breaks a long-term objective
into a dependency DAG of tasks and advances it autonomously — instead of
finishing everything in one conversation.

Core principle: agents propose, the environment adjudicates, the runtime commits.

• Agent plans, executions, and claims are never authoritative;
• For every work item: task/report is a declaration about the environment
  (not authoritative); task/verify independently aligns the declaration
  with environment feedback (authoritative);
• Workers cannot self-certify — only state changes that pass deterministic
  validation and commit atomically as a single session event are authoritative.

Why dsh needs it

Features

• Event-sourced persistence: mission/* session events; one event = one
  atomic commit; no separate database;
• Version guard: revision CAS (mirrors dsh-goal's GoalRef pattern);
• Autonomous driver host (mission-driver): listens to mission/changed
  and runs claim → subagent execute → report → verify → commit automatically;
• Lazy lease reclamation: an expired, unreported claim can be reclaimed and
  redone; an in-flight task guard prevents duplicate dispatch;
• Replan on failure: a stuck mission gets a new revision while DONE tasks
  with unchanged specs carry over;
• Optional model strategies: llmDecider (pick the next task) /
  llmReplanner (replan) / userApprovalGate (approval gate);
• UI data plane: a mission session-projection unit;
• Trigger companion (mission-trigger): a soft-guidance prompt + /mission
  slash command that routes trigger phrases into mission_create/mission_plan;
• Tolerant companion: mission-invariant registers its audit whenever the
  invariants service exists, and is a no-op otherwise (production profiles
  load without extra configuration).

Event vocabulary (mission/*)

States: mission CREATED / RUNNING / COMPLETED / FAILED / CANCELLED;
tasks PENDING / CLAIMED / DONE / FAILED + derived READY / BLOCKED.

Install & use

# Install from npm (published; inserts a 6-row bundle patch into the composition)
dsh plugin --profile <name> add @qiaoy01/mission

Install from source (for contributors)

The published npm package ships prebuilt (lib/), so consumers only need the
one command above. Build from source when you want to modify or review the code:

git clone https://github.com/qiaoy01/dsh-mission.git
cd dsh-mission   # the bundle lives at this repo root
npm install                                # build deps, pinned to the running-harness peer versions
npm run build                              # typecheck + emit lib/
dsh plugin --profile <name> add file:.

file:. points at the repo root you just cd'd into — whose package.json
declares dsh.bundle.patch → cordis.patch.yml.
After changing source, re-run npm run build and re-add the package.

In a session, the agent drives the mission through four model tools:
mission_create (create a mission), mission_plan (commit a task DAG; revision
must equal current + 1), mission_status (read the current state), and
mission_cancel (cancel the mission).

After the model commits a plan, the runtime takes over: mission-driver
claims tasks, dispatches real subagents, and verifies independently. The model
should not execute tasks itself.

Quick start (30 seconds)

# 1. install the plugin into a dsh profile
dsh plugin --profile web add @qiaoy01/mission

# 2. in a session, ask the agent:
#    "mission: build a small web game. Use mission_create, then mission_plan,
#     then stop — the runtime will execute the tasks."
#
# The model creates and plans; mission-driver claims tasks, dispatches real
# subagents, and independently verifies each one until the mission is COMPLETED.

Defaults are conservative (deterministic task selection, no automatic replan —
zero model cost). Enable model strategies via the profile row config:

- id: mission-driver
  config:
    decider: llm     # the model picks the next task
    replan: true     # the model replans when the mission is stuck

Build & test

npm run build          # auto-discovers a local tsc; typecheck + emit to lib/
# Six test suites, all run in isolation:
node --experimental-strip-types tests/domain.test.mjs    # pure domain model
node tests/service.test.mjs                               # runtime (real cordis/session/agent)
node tests/driver.test.mjs                                # driver loop (mock executor)
node tests/driver-host.test.mjs                           # driver host (stubbed subagents)
node tests/trigger.test.mjs                               # trigger companion (soft guidance + /mission)
node --experimental-strip-types tests/projection.test.mjs # session-projection unit

Status

• Stage A (domain model) ✅ · Stage B (service + tools + companion) ✅ ·
  Stage C (driver loop) ✅ · Stage D (replan / lease reclamation / model
  decisions / approval gate) ✅ · Stage E (projection ✅, browser UI pending);
• Verified in a real composition: a 7-task DAG driven end-to-end by real
  subagents to COMPLETED, every task independently verified, zero failures and
  zero duplicate dispatches.

License

MIT — see LICENSE.