koma-miko-dsh (alpha)

Native DeepSeek Harness adapter for Koma Miko Agent Specs. The first alpha targets the exact DSH Developer Preview version verified below; later RC compatibility is not implied.

What it proves

The adapter uses three native DSH extension points:

DSH event	Miko checkpoint	Behavior
tools/pre-execute	PREPARE / PRE_ACTION	DENY blocks; REVIEW uses DSH's native approval path
tools/result	evidence	Only a final isError: false result becomes observed evidence
agent/turn-stopping	COMPLETE	Missing obligations steer one more model step, with a bounded loop guard

The normal CLI, Web UI, and any other DSH surface render the host's own approval, tool-error, and steering interaction. Miko does not add a separate GUI.

DSH Code Mode is covered at the native sub-call boundary. The outer run_code transport is delegated because each tools.* sub-call re-enters tools/pre-execute and tools/result; Miko does not inspect or retain program text.

Install the alpha

dsh plugin --profile miko-lab add koma-miko-dsh@alpha
dsh --profile miko-lab --dump-config

For local development, build from the Koma root and install or link the package into a disposable DSH profile:

Build from the Koma root, then install or link the package into a disposable DSH profile:

npm run build -w koma-miko -w koma-miko-dsh
dsh plugin --profile miko-lab add ./packages/koma-miko-dsh
dsh --profile miko-lab --dump-config

Place a DSH-named miko.json in the session workspace. The narrow example is in examples/ui-change. It expects DSH tool names such as skill, read, write, edit, and bash, rather than Claude Code's capitalized names.

Repeat the bounded live evaluation

After building the package, set ANTHROPIC_API_KEY in the parent process and, if dsh is not on PATH, point MIKO_DSH_BIN at the executable. Then run:

npm run eval:dsh-live -w koma-miko-dsh

The runner creates a disposable DSH home and fixture, installs the local bundle, uses claude-haiku-4-5, disables title generation and telemetry, disables retries, caps the run at eight agent requests and 768 output tokens per request, and restricts the available tools. It then verifies the compressed DSH session artifact rather than trusting the model's final claim. The API key is read only from the process environment; the runner never opens an env file. Temporary files are deleted unless MIKO_DSH_KEEP_TEMP=1 is set.

Set MIKO_DSH_PLUGIN_SOURCE to an absolute .tgz path to test the exact packed artifact rather than the workspace directory. A passing report includes the observed tool sequence, request count, token usage, and model-run latency.

To turn an exact successful command into completion evidence, override the bundle row in the profile's cordis.patch.yml:

- id: koma-miko-dsh
  name: koma-miko-dsh
  config:
    specPath: miko.json
    checks:
      - name: targeted-tests
        tool: bash
        argument: command
        equals: npm test -- Hero

An isError: false result with that exact argument records only check_passed:targeted-tests. The command, prompt, source code, model response, and tool output are not copied into Miko evidence. Background commands never count as a passing check because their successful result means only that a job started.

Configuration

Field	Default	Meaning
specPath	miko.json	Agent Spec path relative to each session workspace
taskTags	[]	Deployment-owned tags; prompt text is never classified
missingSpec	warn	Warn once per unguarded workspace, or use silent
reviewPolicy	ask	Map REVIEW to DSH ask, or make it deny
failureMode	open	Adapter failures delegate; closed throws
unknownRisk	high	Conservative risk for unknown/custom tools
riskOverrides	[]	Exact { tool, risk } deployment overrides
checks	[]	Exact successful-tool matches for named check evidence
evidenceArgumentNames	path keys only	Primitive arguments allowed into tool_succeeded evidence
maxCompletionSteers	2	Prevent a missing obligation from causing an infinite stop loop

Host dependency rule

@deepseek-ai/cordis and every @deepseek-ai/dsh-* package are peers, not runtime dependencies. A DSH profile must resolve the host's one shared copy. Shipping a second dsh-tools instance can split singleton symbols and break the tool pipeline. The exact compile target for this experiment is DSH 0.1.1-rc.2. Peer versions are pinned exactly because compatibility with a later Developer Preview is not assumed.

Honest boundaries

• Miko records that DSH successfully loaded a Skill; it cannot prove the model understood or followed it.
• This experiment cannot force a model to choose a Skill before the first relevant action; it can block that action and explain what is missing.
• The first alpha deliberately starts a fresh evidence epoch on every DSH resume/restart. Required Skills, references, artifact changes, and checks must be observed again.
• Live adapter state is not replayed from the DSH session log after process restart or plugin hot reload. This is an explicit alpha policy, not silent recovery.
• Completion steering is corrective, not an unbounded hard lock; the configured loop guard eventually lets the turn close.
• Exact command matching is deliberately narrow. Miko does not infer from arbitrary terminal text that tests passed.
• Miko complements DSH approval and sandbox policy; it does not replace either.

The live host gate passes: three bounded Haiku sessions from a packed adapter showed blocked write → observed skill/reference → allowed write → observed exact check → allowed completion. The measured 3/3 result is a narrow integration signal, not a general model-reliability claim.

Current measured results are in docs/evals/miko-dsh-alpha.md.