dsh-short-tool-ids
适用于 DeepSeek Harness 的作用域限定、可逆工具调用 ID 兼容插件
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:mrbeandev/dsh-short-tool-ids#e88d7d951beced07e2e97728a5961d44ff5abb0dShort Tool Ids 以 dsh-short-tool-ids 发布，当前版本为 0.2.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
An experimental DeepSeek Harness plugin with a per-provider Use short tool-call IDs checkbox in Settings → Models, on configured pi-ai provider cards.
Fixes ordinary oversized tool-call IDs in OpenAI-compatible Chat Completions
history, such as maximum length 64, got length 81, without editing saved sessions.
This was built to recover a real DSH chat session that could no longer accept messages. The session history contained tool-call IDs with 81 characters, but the OpenAI-compatible API receiving the request accepts at most 64. Every new turn resent the invalid history, so the session failed before the model could respond with:
Invalid 'input[5].call_id': string too long. Expected a string with maximum length 64, but got a string with length 81 instead.
The original route used a custom cc provider. DSH's built-in normalization
did not cover that provider during same-model replay, and renaming it to
openai: cc would not activate the exact upstream openai branch. Renaming it
to openai would still leave the same-model replay gap.
Use this plugin when an OpenAI-compatible Chat Completions provider rejects a resumed or tool-using chat because a tool-call ID is too long. Enable the toggle only for the affected provider in Settings → Models. The plugin shortens IDs at request time, rewrites matching tool results consistently, and leaves the saved session history unchanged. It is a targeted compatibility workaround, not a general fix for Responses or Anthropic message IDs.
Install it into the Web profile with one command:
dsh plugin --profile web add dsh-short-tool-ids
The package declares the bundled host patch and browser registration in its
package.json metadata, so a global npm install is not required. Stop the
existing DSH Web process when active work is finished, then restart it normally
(without the previous short-ID --patch flag):
dsh web --no-open --port 3080
Refresh the existing Web page. In Settings → Models, enable Use short tool-call IDs for the provider that returns the ID-length error. Retry that chat.
To install from a local checkout instead, from a terminal:
dsh plugin --profile web add "/absolute/path/to/dsh-short-tool-ids"
Stop the existing DSH Web process when active work is finished, then restart it
normally (without the previous short-ID --patch flag):
dsh web --no-open --port 3080
Refresh the existing Web page. In Settings → Models, enable Use short
tool-call IDs for the provider that returns the ID-length error. Retry that chat.
All provider toggles default OFF, including cc when upgrading from v0.1.0.
The profile installation persists across restarts. Each enabled provider covers all its eligible models and sessions in that process. It does not enable other profiles or separately launched worker processes. Keep the local package folder in place if the package manager installed it as a link.
Do not load this package and the old startup overlay together. Two installs of the same adapter wrapper are rejected. The host client-package manifest and browser bundle are cached, so upgrading from the host-only plugin requires a server restart plus page refresh. There is no standalone UI server to start and no core Web rebuild is necessary; the plugin uses an existing settings slot.
Rollback:
dsh plugin --profile web remove dsh-short-tool-ids
# Stop/restart the existing server without any short-ID --patch flag.
Or switch a provider's toggle off; subsequent dispatches stop rewriting its IDs.
short-tool-ids.providers settings
map, not in API keys or the provider's own configuration schema.openai-completions models only. Other protocols and native
adapters are untouched even when that provider's switch is on.call_ plus 48 SHA-256 hex
characters (53 total). Calls and matching results receive the same mapping.call_id|item_id IDs untouched; this is NOT a general
Responses/Anthropic compatibility fix. Unsupported signed/native history that
would require an unsafe rewrite is refused rather than stripped.The checkbox disables itself for loading, unavailable, read-only or non-host settings and during saves. Path-scoped writes retain other provider settings and use revision fencing. Save conflicts/failures show an error rather than a false success; different provider cards subscribe to the shared settings scope.
No guarantee of zero regressions. This is a version-gated prototype wrapper,
not an officially supported adapter-decorator API. DSH's ordinary llm/stream
middleware sees frozen durable requests and cannot replace their histories.
The plugin resolves the adapter from the running CLI's dependency tree, wraps
PiAiAdapter.stream() and prepared-call dispatch reversibly, and retains original
prepared handles/model snapshots. Runtime versions are deliberately pinned:
| Package | Tested version |
|---|---|
@deepseek-ai/dsh-llm-pi-ai | 0.1.5-rc.2 |
@earendil-works/pi-ai | 0.85.1 |
| Inspected CLI | 0.1.5-rc.1 |
Untested adapter versions refuse plugin startup and require review. Disable or
remove the plugin if an upgrade prevents startup. It must not be combined with
another wrapper of the same methods. The UI occupies the keyed
settings.models.provider-card slot for llm-pi-ai; that slot supports only one
occupant per key, so another plugin using the same key needs coordination.
Renaming a provider to literal openai: cc does not trigger the upstream exact
provider === 'openai' branch. Even renaming to openai leaves the upstream
same-model normalization gap.
Zero npm runtime/build dependencies: server code is ESM; the dependency-free client builder emits DSH's lazy ModuleLoader format using the shell's existing React.
npm run build
npm test
DSH_TEST_HARNESS_ENTRY="/absolute/path/to/dsh/lib/bin.js" npm run test:integration
npm run pack:check
Use a workspace-local npm cache if desired (npm_config_cache="$PWD/.cache/npm").
Optional DSH_TEST_SESSION=/path/to/decoded-session.jsonl enables private-history
replay checks. Never package or commit real session history. Generic test runs
skip integration tests without the explicit installed CLI path; the integration
script refuses a missing path rather than silently skipping adapter verification.
Tests include frozen-history preservation, matching results, boundary lengths, collisions, live provider toggles, protocol isolation, prepared snapshots, unload/duplicate guards, iterator cancellation/errors, settings schema, actual adapter HTTP serialization, provider checkbox writes/errors/read-only state and lazy browser-bundle registration. Mock HTTP tests use a temporary loopback SSE endpoint and fake credentials, close it afterward, and make no paid API calls. UI tests use a fake React hook harness, not a full browser DOM test.
v0.1.0's live repair was confirmed with a successful reply on a previously stalled session. The v0.2.0 provider-card UI was checked in a running DSH web session: toggle rendering, saving, and persistence across a page reload.
MIT. Maintainer release steps are in RELEASE.md.