English | 中文

dsh-llm-kiro

A Kiro provider for the DeepSeek Harness LLM seam.

The package is a standalone Cordis plugin: it mounts in any harness composition through cordis.yml, registers the kiro provider route on ctx.llm, and needs no changes to the harness itself.

What it handles

Every runtime request carries the harness attribution User-Agent (from @deepseek-ai/dsh-llm attributionHeaders()) plus the AWS SDK-shaped x-amz-user-agent telemetry header the backend expects.

Install

The package is ESM and declares the harness packages it talks to as peer dependencies — the host DSH installation provides them. It also declares the dsh bundle manifest (cordis.patch.yml ships with the package), so it mounts as a profile bundle as well as a plain plugin dependency.

dsh plugin --profile <profile> add @joshryandavis/dsh-llm-kiro

Mount it in cordis.yml:

- id: llm-kiro
  name: '@joshryandavis/dsh-llm-kiro'
  config:
    region: us-east-1        # optional; the credential's region wins when omitted
    reasoningEffort: high    # optional; off | low | high | max (default high)

That is the whole configuration for an existing kiro-cli or Kiro IDE login: credentials are picked up ambiently, refreshed silently, and written back. A static bearer token works too:

- id: llm-kiro
  name: '@joshryandavis/dsh-llm-kiro'
  config:
    bearerTokenEnv: KIRO_BEARER_TOKEN   # stored via the harness credentials service

Config reference

All fields are optional; the llm-kiro user-settings section uses the same schema and can be edited at runtime — a change reaches the next request without a restart.

Models

Bootstrap catalog (harness id → wire id):

Both dashed and dotted spellings resolve to the wire id. The bootstrap list is the pre-discovery fallback and the curated menu order; the authenticated management catalog (List-Available-Models, per credential, per region) refreshes matched entries' capacities, effort schemas, and reasoning metadata and appends models the bootstrap list has not learned yet. Discovery reaches the harness model picker: reading the model list serves the cache immediately and refreshes it in the background for the next read, so a slow or unreachable management endpoint never delays a picker or a request. An explicit non-empty models list replaces discovery entirely and is served verbatim.

The catalog cache at ~/.kiro-management-models-cache.json is shared with pi-provider-kiro: that tool writes format version 1, this provider reads that layout and writes version 2.

Development

pnpm install
pnpm run build       # tsc -b: compiles this package; deps resolve from the published @deepseek-ai packages
pnpm test            # vitest suite (105 tests): endpoints, store, refresh, models, effort, serialize,
                     # translate, adapter transport, real Loader composition

Tests resolve the harness packages from the published @deepseek-ai/* packages in node_modules (peer dependencies plus the dev-dependency set at the versions the code is developed against), so they exercise the exact published APIs. Credential tests are fully isolated from the host machine's real kiro-cli / Kiro IDE stores.

Known limitations and deferred work

• Interactive login is not bundled: with no ambient credentials the provider fails with MISSING_CREDENTIAL and points at kiro-cli login or the Kiro IDE. A device-code flow can be added later on the management seam.
• Legacy inline thinking dialect: reasoning arrives through the native thinkingText events; models that emit <thinking> tags inside content keep them in the visible text block.
• Replay state is not emitted; history is resent as provider-neutral content on follow-ups.
• The management catalog refresh is pull-based, not scheduled: reading the model list serves the cached or bootstrap catalog and refreshes the cache for the next read, so the first read after the TTL can lag one launch behind the authenticated catalog.

License

MIT