Skip to content

Releases

v0.4.0 — WorkBuddy, model catalogs, reminders, agent pool

Choose a tag to compare

@mike-sl-ig mike-sl-ig released this 13 Oct 11:38

0.4.0 — WorkBuddy, model catalogs, advisory reminders, agent pool

New

  • WorkBuddy (Tencent CodeBuddy Code) as a bridged agent: headless CLI, run through the Node that hosts the plugin (interpreter.source: "self"), with its own product identity supplied by run.env (ACC_PRODUCT_CONFIG_PATH).
  • Model catalogs declared per recipe (catalog): read a JSON file (TTL-cached) or ask the CLI itself (kind: "argv"). Filtering is field-level (require: { supportsToolCall: true }) plus excludeIdPattern. A missing or malformed catalog degrades to the recipe's static list, with the reason attached.
  • Free/paid tiers in the model picker, derived from each catalog's own credits multiplier (x0.00 = free).
  • Read-only balance display (ok / unauthenticated / unsupported / error). balance.write is rejected by the schema, so the plugin cannot even describe a call that mutates an account.
  • Advisory per-turn reminders on DSH's agent/pre-step waterfall: enabled role features announce themselves once per turn instead of waiting for the user to name an agent.
  • Agent pool: the panel multi-selects candidates instead of radio-picking one default; defaultAgent is derived from the pool, and the choice of which agent (and how many) is left to the model.
  • Failures explain themselves: a CLI's own error event now reaches the report, last-run.json, the panel and agent_list.

Fixes

  • findOnPath preferred an extensionless POSIX shim over the .cmd/.exe sitting beside it.
  • Catalog discovery no longer lets a PATH entry outrank the install whose product config the CLI needs.
  • JSON array output (what CodeBuddy prints) is parsed as an event stream, so a successful run no longer reports an empty answer.
  • catalog.kind: "argv" was validated but unimplemented — a promise the schema made and the code did not keep.
  • The panel no longer claims the user is signed out when the credential is merely unreadable.

Full suite: 17 test files, 0 failures.

dsh-agent-bridge 0.1.2

Choose a tag to compare

@mike-sl-ig mike-sl-ig released this 30 Sep 16:13

0.1.2

  • Docs only: a full Chinese README (README.zh-CN.md) with a language switcher, plus badges for CI / npm version / downloads / license / node / zero dependencies. Both READMEs ship in the package.
  • Tooling: pack-check now also asserts that package.json carries no byte-order mark — a BOM is legal UTF-8 but breaks JSON.parse everywhere else while npm tolerates it, so nothing else would notice (found the hard way).

dsh-agent-bridge 0.1.1

Choose a tag to compare

@mike-sl-ig mike-sl-ig released this 30 Sep 15:28

0.1.1

  • Fix (live-app breaking): tool results are now returned as content-part blocks. DSH calls render(args, value) and treats the result as blocks, so returning a bare string made every tool call fail in the running app with content.some is not a function. The contract is now asserted per tool by the harness.
  • Fix: the recipe drafter searches nested objects and every JSONL event, so a streaming CLI's answer (part.text) is recovered instead of missed.
  • Fix: the self-test accepts streaming fixtures (*.jsonl), so those recipes really replay their recorded stream.
  • The paid Claude Code adapter is opt-in in the tests (LIVE_PAID=1); the default run spends nothing.