dsh-iris

dsh-iris — progressive capability disclosure for DeepSeek Harness

Start with what matters. Add more when needed.

Iris keeps optional DeepSeek Harness capabilities available without putting all of them in the model context up front.

All four modes start minimal. Capabilities expand when needed.

简体中文

👁️ Why the name Iris?

An eye's iris controls how much light reaches it. dsh-iris does the same for an Agent's capability context: DeepSeek Harness keeps its full capability ceiling, while Iris adjusts how much of that ceiling the model can see in the current step.

That is why the project is called Iris. It does not shrink the interface or remove capabilities. It controls the Agent's capability aperture, revealing Tools, Skills, MCP capabilities, and Plugin-provided extensions only when they are useful.

🎯 Why Iris exists

DeepSeek Harness can combine native Tools, Skills, connected MCP Tools, and capabilities supplied by Plugins. The problem is not that these capabilities exist. The model rarely needs every Tool schema, Skill description, MCP route, and Plugin-specific instruction on the first turn. Exposing them all at once adds irrelevant choices, lengthens the prompt, and can make capability selection less predictable.

The community often summarizes this as “GA DeepSeek V4 is overfit to Minimal.” A more careful statement is that V4 appears unusually sensitive to its first-turn Tool and prompt environment, while its published Code Agent evaluation used DeepSeek Harness Minimal.

Without Iris, each DSH mode exposes capabilities in its native way. With Iris, the mode still defines what the Agent may use. Iris only changes what the model sees right now: it starts with the essentials and adds more when the task calls for them.

Harness mode → full capability ceiling
                       │
                    dsh-iris
                       │
          only the essentials now
                       │
       search / explicit demand / recovery
                       │
                 reveal on demand

✨ Why this is a lighter approach

Iris does not replace DeepSeek Harness. It gives less software more ownership.

Approach	What it usually owns	Iris instead
Forked preset or global prompt patch	The mode persona and prompt	Keeps native presets and yields to an explicit prompt owner
Always-on capability bundle	Imports and exposes every Tool up front	Catalogues metadata first; imports local Providers only on activation
Second router Agent / LLM loop	Another decision and execution loop	Reuses the current Agent loop and deterministic DSH lifecycle hooks
Parallel Skill or MCP runtime	Loading, execution, and teardown	Routes back to DSH's native Skill, MCP, ToolRuntime, approval, and guard owners
Global Tool mutation	Every Agent sees the same change	Mounts and reveals capabilities in the requesting Agent's scope

That is the practical difference: Iris controls capability exposure. It does not build another Harness beside DSH.

🧭 What changes when you enable Iris

Choose Enable Iris — All four modes start minimal and expand capabilities when needed beside the DSH mode selector. Choose Disable Iris — Use DeepSeek's native behavior to return to the unmodified DSH experience.

Mode	Initial experience	On demand
Minimal	Preserves native DSH Minimal exactly at Iris's verified boundaries	Iris adds no controls or automatic activation
Standard	Core capabilities and lightweight Iris controls	Adds native capability groups or lazily activates configured Tools
Code / PTC	Native run_code with a small, step-stable SDK	Stages compatible bindings for the next normal step
Creator / Cordis	Core capabilities first	Adds the privileged Creator control plane only for creator intent

During a session Iris only adds capabilities; it does not repeatedly hide and re-add them between turns. Settings → Iris shows what the model can see now, what remains available, prompt/schema/SDK sizes, and the capabilities recently added.

🔀 Discover, then use the right owner

Iris gives the Agent one place to search without forcing every capability through one lifecycle:

tool:text_word_count     → iris_activate → lazy Agent-scoped Provider mount
skill:repo-review        → native skill route → DSH loads the Skill body
mcp:github/create_issue  → connected MCP Tool → direct DSH execution

• iris_search and iris_recommend(query) read metadata only.
• iris_activate(tool:...) loads and verifies a configured local Provider once.
• iris_activate(skill:...) delegates to the native DSH Skill route without loading the body itself.
• An already-connected MCP Tool is returned as directly available; Iris does not reconnect it.
• UNKNOWN_TOOL enters the same activation pipeline as a deterministic fallback. Iris never replays the failed call.

DSH remains responsible for the Agent loop, ToolRuntime, Skills, MCP transport, Code SDK, approval, guards, execution, cancellation, sessions, and Cordis lifecycle.

🚀 Quick start — no config-file editing required

Install Iris into the Web profile:

dsh plugin --profile web add dsh-iris
dsh --profile web

Then use the UI:

1. Select Enable Iris beside the DSH mode selector.
2. Open Settings → Iris to choose Auto, Preserve, Adaptive, Adaptive Code, or Adaptive Creator behavior.
3. Configure metadata discovery, logging, and local Providers with controls in the page.
4. Start a session. Search or activate a capability only when it is useful.

A configured Provider is only a catalogue declaration. Bundle startup, search, and recommendation do not import or apply it.

iris_search({ query: "count words" })
→ tool:text_word_count metadata
→ Provider apply count: 0

iris_activate({ capabilityId: "tool:text_word_count" })
→ resolve → policy → import → mount → verify → reveal
→ Provider apply count: 1

text_word_count({ text: "Iris opens only when needed." })
→ normal DSH execution pipeline

See examples/local-text-tools for a dependency-free example.

✅ What works today

• Minimal-first disclosure for Standard, Code / PTC, and Creator / Cordis.
• Canonical Minimal preservation as the control experience.
• Agent-scoped Tool activation, authoritative verification, single-flight, rollback, and teardown.
• Unified metadata discovery and routing across configured Tools, model-invocable DSH Skills, and connected MCP Tools.
• Step-stable Code SDK updates between normal model steps.
• additionalContexts handoff after recovery; no transparent Tool replay.
• English and Simplified Chinese DSH Web UI with writable settings and live capability telemetry.
• Reproducible Vanilla-vs-Iris benchmark harness. Until a credentialed run is saved, the report says results pending rather than inventing numbers.

🛣️ What Iris can grow into

The next features stay behind the same ownership rule:

• benchmark-guided capability-pack tuning rather than hand-wavy “smaller is better” claims;
• better task-aware recommendations when measured retrieval misses justify them;
• a community registry contract followed by explicit approval and installation;
• broader Skill and MCP lifecycle support only where DSH exposes safe public, Agent-scoped seams;
• evidence-guided adaptation across sessions without creating a second Agent loop.

Configured-but-disconnected MCP lazy activation remains intentionally deferred: current DSH does not expose the clean configured-server enumeration and Agent-owned lifecycle seam Iris would need.

🛠️ Development

pnpm install
pnpm typecheck
pnpm lint
pnpm test
pnpm build
pnpm pack --dry-run

📄 License

MIT. See LICENSE.