dsh-longrun-suite

DSH LongRun Suite — long-running mission tooling for DeepSeek Harness

This project explores how AI can run long-horizon research and project work. Deliverables are aimed to be as reliable as possible. It is under active development; please open an issue if you find problems — the author checks them regularly. If you find it useful, a star would be appreciated. Thank you!

One repository for long-horizon autonomous work: mission control, the Long-Run Captain preset, a strictly reviewed LLM-as-a-Verifier, and a self-wake timer. Clone and install.

中文 README

Problem

DSH's built-in goal / todo / subagent tools are great for short tasks, but long, hard missions fail in predictable ways:

1. Plan once, then stop. A rejected direction does not trigger a new plan or new tasks.
2. Fail and halt. Verification failures stop the run instead of fixing the gap and pivoting.
3. Human wake-ups required. Long experiments are not re-checked automatically.
4. Weak completion criteria. An empty task list is treated as "done" even when the real goal is unverified.
5. Subagent overhead. Durable continuable children are enumerated repeatedly, wasting memory and CPU.
6. Single-point summary loss. A central agent re-stating worker results drops conditions, numbers, and details.
7. Primitive worker coordination. Waiting/chatting leads to duplicated work, double writes, and unclear ownership.
8. Lazy/downgraded work. The model writes simplified replacements, ignores existing implementations, or silently replaces a standard method with a weaker one.
9. Experience loss. Cross-mission lessons are not searchable and old pitfalls repeat.
10. Self-assessment inflation. Producers review their own work without independent/no-memory evaluation.
11. Unrecoverable long tasks/reminders. Reminders are lost, fork cold-start fails, and stale reminders fire repeatedly.

This suite fixes that with file-backed mission state, a tiny generic task lifecycle, data-driven verification plans, independent review, forced replanning on rejection, Claim Pool/Lease, Blackboard artifact communication, WorkReceipts, no-memory blind review, LLM Wiki memory, method provenance cards, and timer-based self-wake — without hardcoding any domain workflow. Model upgrades, mathematical research, and software projects are all just different verificationPlan values.

Components

The two presets

The suite ships two Long-Run Captain presets with the same mission capabilities, differing only in first-turn system prompt shape:

Rules of thumb:

• For DeepSeek V4 Flash family, use Long-Run Captain Router: lighter first turn, full mission tools / subagents / timers, and role constraints travel via the Role Card in dispatch prompts.
• For other models or when you want the heavier full system prompt, use Long-Run Captain.
• Both presets require dsh-mission-control and share the same task lifecycle, review and replanning rules.

DSH Store submission

This repository is a monorepo of three independent DSH plugins. Each plugin must be submitted with an explicit package path:

Every subpackage declares the exact DSH compatibility matrix in package.json. See STORE_SUBMISSION.md.

Quick start

In a Long-Run Captain or Long-Run Captain Router session:

Start a mission: build a CLI tool that recursively scans a directory of
Markdown files and generates an index.md with heading levels, relative paths,
and last-modified times.
termination_policy: success
budget: { maxRounds: 6, maxHours: 4 }
Success criteria:
- The CLI recursively scans a directory and writes index.md
- The index is organized by heading levels and includes relative paths + mtimes
- 3 test cases are provided and all pass
- A README explains installation and usage
- An independent reviewer verifies the index contents

Expected flow:

mission_start
→ heavy early web research + wiki_search for prior experience
→ mission_add_tasks (acceptance + verificationPlan + capabilities)
→ workers claim by capability with mission_claim; long tasks mission_heartbeat; release when unable
→ method-card before non-trivial implementation; reuse first, no speculation
→ long experiments in background + schedule_reminder self-wake
→ workers exchange typed artifacts via mission_publish_artifact / mission_consume_artifacts
→ mission_submit + WorkReceipt, then independent reviewer
→ rejected tasks -> mission_replan + replaces -> new direction
→ verify_track monitors direction health
→ wiki_write for durable lessons; mission_blind_review + wiki_lint before final
→ mission_final_audit maps every success criterion to evidence
→ mission_complete

Key mechanisms

• Only five task statuses: open → active → needs_review → accepted / rejected
• Rejections require a follow-up task via replaces=...; otherwise completion is refused
• Default terminationPolicy=success: tasks mapped to success criteria must have outcome=success
• mission_check validates evidence honesty, not domain correctness
• socratic-self-audit makes producers attack their own work before submission
• verify_track uses the reference implementation's strict calibration prompt and refuses to trust agent narration; pivot selection and rollout criteria also follow the reference implementation
• WorkReceipt: evidence files are SHA-256 hashed on submit; passing review writes an immutable work receipt
• Swarm-style worker pool: tasks can declare capabilities; workers claim only matching tasks; leases are extended by mission_heartbeat, released by mission_release, expired automatically, and blocked after repeated reclaims
• Blackboard artifact communication: mission_publish_artifact / mission_consume_artifacts; workers exchange typed artifacts instead of chatting
• No-memory blind review: mission_blind_review writes blind_review.md + calibration_gap and is a hard gate for substantive deliverable missions
• LLM Wiki memory: wiki_write / wiki_search / wiki_lint maintain a searchable .memory/ across missions

Overall structure

Host / plugin layer
├── dsh-mission-control
│   ├── lib/core.js          # pure mission state machine (no DSH dependency)
│   ├── lib/index.js         # mission_* / wiki_* / artifact tools
│   ├── bin/mission_check.mjs
│   └── preset/              # Captain presets and protocol skills
├── dsh-timer-scheduler-ui    # reminders + auto-cancel + same-session cold resume
└── dsh-plugin-llm-verifier   # LLM-as-a-Verifier

Agent / preset layer
├── long-run-captain/         # full system-prompt edition
└── long-run-router/          # router-standard minimal edition (DeepSeek V4 Flash tuned)

Skills layer
├── mission-protocol          # decomposition, dispatch, reuse-first, tool map
├── task-profile              # Input/Decision/Output/Core Challenge/No-Lazy
├── plan-critique             # plan critique + A/B/C + module contracts
├── adaptive-verification     # minimum validation package
├── method-card               # method provenance / downgrade detector
├── wiki-memory               # LLM Wiki: ingest/query/lint
├── lessons                   # intra-mission lessons + mission-legacy + mission-cases
└── report-protocol / socratic-self-audit / etc.

Runtime data layer
├── .mission/<id>/mission.json   # mission/task/attempt/receipt/blindReview/artifacts
├── .memory/                     # cross-mission LLM Wiki, capability vocab, worker resumes
├── .mission-cases/              # lightweight case cards
└── timer-reminders.json         # persisted reminders

References & inspiration

This project borrows mechanism-level designs from several public agent systems:

We adapt, not copy: DSH genericity is preserved and all domain content remains owned by mission/task data.

Repository layout

dsh-longrun-suite/
├── README.md                 # Chinese entry
├── README.en.md              # English
├── install.sh
├── packages/
│   ├── dsh-mission-control/
│   ├── dsh-plugin-llm-verifier/
│   └── dsh-timer-scheduler-ui/
└── preset/
    ├── long-run-captain/
    └── long-run-router/

Install

Requirements: Node 20+, DSH 0.1.0-rc.8+, a configured LLM provider.

Option A: one line (plugins only)

dsh plugin --profile web add github:GMH13552/dsh-longrun-suite

This installs all three plugins at once. Then install the presets (both, choose per session):

# A1: from a clone
git clone https://github.com/GMH13552/dsh-longrun-suite.git
cp -R dsh-longrun-suite/preset/long-run-captain ~/.dsh/.agent-presets/long-run-captain
cp -R dsh-longrun-suite/preset/long-run-router   ~/.dsh/.agent-presets/long-run-router

# A2: from the installed package
cp -R ~/.dsh/profiles/web/node_modules/dsh-longrun-suite/preset/long-run-captain ~/.dsh/.agent-presets/long-run-captain
cp -R ~/.dsh/profiles/web/node_modules/dsh-longrun-suite/preset/long-run-router   ~/.dsh/.agent-presets/long-run-router

Option B: clone + one script (recommended; plugins and preset together)

git clone https://github.com/GMH13552/dsh-longrun-suite.git
cd dsh-longrun-suite
./install.sh            # installs into the web profile
# ./install.sh tui      # another profile

The script:

1. Adds the three plugins to your profile.
2. Copies both long-run-captain and long-run-router into $DSH_HOME/.agent-presets/.
3. Prints restart instructions.

Then restart DSH:

dsh web

and start a new session with Long-Run Captain (full system prompt) or Long-Run Captain Router (DeepSeek V4 Flash optimized minimal edition).

Manual install

dsh plugin --profile web add ./packages/dsh-mission-control
dsh plugin --profile web add ./packages/dsh-plugin-llm-verifier
dsh plugin --profile web add ./packages/dsh-timer-scheduler-ui

mkdir -p "$HOME/.dsh/.agent-presets"
cp -R preset/long-run-captain "$HOME/.dsh/.agent-presets/long-run-captain"
cp -R preset/long-run-router   "$HOME/.dsh/.agent-presets/long-run-router"

dsh-plugin-llm-verifier defaults to provider: deepseek-official and model: deepseek-v4-flash-vision-exp. Edit the llm-verifier row in your profile cordis.patch.yml if your routing differs, or edit packages/dsh-plugin-llm-verifier/cordis.patch.yml before installing.

Known limits

• DSH streams do not expose logprobs, so the verifier approximates the paper's logit expectation with repeated sampling
• schedule_reminder currently wakes only live sessions; cold resume is a future direction
• Independent review is a procedural boundary, not a sandbox
• Claim Pool / Lease is file-based (mission.json + timestamps), not a service; expired leases are reclaimed automatically, and 3 reclaims block the task.
• Capability Matching is tag set matching with no built-in ontology; each workspace maintains .memory/_capabilities.md.
• Blackboard artifacts store metadata only; real artifact files remain on disk.
• mission_blind_review records external ratings and writes blind_review.md; it does not generate ratings itself.
• LLM Wiki is Markdown + grep, not a vector DB; wiki_lint is text-level only.
• method-card / wiki_* are mostly soft protocol; only blind review is a hard gate for substantive deliverable missions.

Credits

• The verifier is based on the LLM-as-a-Verifier paper and dsh-plugin-llm-verifier, with stricter review and corrections
• Self-wake timer based on dsh-timer-scheduler

License

MIT