dsh-iris

Progressive capability selection for DeepSeek Harness. The current model reads short descriptions, selects a capability, and uses DSH's normal loading and execution path.

简体中文

Iris is named after the eye's iris: it adjusts capability exposure within the current Agent's mode and permissions. It preserves native persona and does not require another routing model or any GPT-6 API dependency.

How selection works

Adaptive modes expose core tools and an Agent-scoped iris:capabilities section. A small catalog is described directly; larger catalogs retain iris_search for capabilities outside the card budget. Already available native tools, Skills and MCP tools follow their native routes without redundant activation.

short description → current model selects capability ID
  visible native Tool / MCP → normal DSH call
  hidden authorized inherited Tool → iris_activate → normal DSH call
  configured local Provider → iris_activate → import/apply/verify → normal DSH call
  Native Skill → native skill route (reveal its entry only if needed)
  insufficient description → iris_search → select the result's route

Metadata reads never import Providers, load Skill bodies, connect MCP servers or execute business tools. Description text grants no permissions. Iris does not install community plugins, create a second Skill/MCP runtime, or proxy calls through a generic JSON execution tool.

Modes and host boundaries

Mode	Behavior
Minimal / Preserve	Native persona, tools and order; no Iris controls, cards or automatic activation. Unknown custom modes default to Preserve.
Standard	Core tools and budgeted cards; explicit activation for hidden or unmounted capabilities; search for the long tail.
Code / PTC	Native run_code and DSH SDK. A running program keeps its bindings; newly activated capabilities wait for the subsequent host step.
Cordis / creation	Native persona, creation restrictions and creation entries remain available, including cordis_mount / cordis_unmount. A failed search does not authorize plugin creation.

DSH rc.6 restrictions filter inherited tools, but cannot safely hide tools registered in the Agent's own scope. Such MCP tools remain native-visible. A Provider registering several Agent-owned tools exposes its actual Provider loading unit. Iris reports that boundary instead of promising one-tool isolation. Optional inherited tools use precise disclosure or a small explicit native dependency group; unknown extensions do not share an all-extensions switch.

MCP runtime identity uses the full registered name, such as mcp__github__create_issue; a short display name is not an execution identity. DSH retains authorization, approval, guards, cancellation and ownership. Configured but disconnected MCP activation is unsupported because the installed host lacks a sufficient public enumeration/connection lifecycle contract.

Install and configure

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

Use the Iris control beside the mode selector and Settings → Iris for configuration and telemetry. These installation commands are documentation; local repository verification does not install or publish a plugin.

Default card budgets are 8 cards, 200 Unicode characters per description, and 1800 Unicode characters for the section. They are initial product parameters, not token estimates. Whole over-budget description fields are omitted with a search fallback; IDs and routes are never cut. Native Skill descriptions are left with their native owner to avoid duplicate cards.

iris:
  enabled: true
  policy: auto
  legacyInterface: false
  cards:
    maxCards: 8
    descriptionChars: 200
    totalChars: 1800

New sessions register iris_search and iris_activate. Set legacyInterface: true to retain iris_recommend; it is not registered by default. Existing sessions retain their captured policy, Provider configuration, budgets and control set to avoid changing replay contracts. Changes apply to new sessions; disabling Iris immediately tears down its active runtimes. Older configurations receive the new defaults.

Configured local capabilities support description, whenToUse and bilingual keywords. Configured kind: skill is not a local Skill loader; use DSH's native Skill configuration. Local Providers require explicit activation and policy checks.

For the local text-tools example, a card can lead directly to:

iris_activate({ capabilityId: "tool:text_word_count" })
→ import / apply / verify / reveal
text_word_count({ text: "Iris opens only when needed." })
→ native ToolRuntime

The intended direct-card path has one activation control call and one Provider apply, with no search/recommend prerequisite. See integration tests for measured fixture behavior; this trace is not a model performance measurement.

State and evidence

Settings separates catalogued, mounted, visible and staged capabilities. Catalog counts and exposed tool-schema counts are different units; one run_code schema can expose many SDK capabilities. Host registration does not prove current-program readiness. rc.6 has no public structured current-step binding inventory, so SDK readiness requires the pinned host execution contract rather than a substring match.

The architecture and optimization audit record implementation boundaries and verification. Historical single-pair A/B results remain dated sanity checks, including slower Code execution; persona rewrites in older experiments are a confound. The new offline comparison checks transport and verifiers only. It establishes no token, latency or task-success improvement. No paid model evaluation is included in this round.

Development

Node >=22 and pnpm 11.7.0; the local integration baseline uses DSH 0.1.0-rc.6.

pnpm install --frozen-lockfile
pnpm typecheck
pnpm lint
pnpm test
pnpm build
pnpm smoke:release
pnpm smoke:package
pnpm benchmark:optimization

smoke:package inspects the real tarball and installs it offline into a temporary DSH host fixture. It imports server exports and resolves the browser client; browser rendering still requires DSH Web. Benchmarks and Python caches are excluded from npm packaging. CI runs local checks without credentials, paid requests or publishing. See benchmark instructions for the remaining live-evaluation requirements.

License

MIT for Iris. See LICENSE. Retained BFCL research fixtures have separate upstream provenance requirements documented in their README; they are not included in the npm package.