resolveModelForDelegateTask returned input.userModel immediately without
availability checking when it was set, silently ignoring the user's
configured fallback_models. Effect: a team-mode category member
(e.g. hyperplan's 'artistry' category) configured with
{ model: "opencode/gemini-3.1-pro", fallback_models: [...] }
spawned with the unreachable primary even though a listed fallback was
reachable, leading to a dropped/broken team member instead of graceful
degradation.
Now, when the provider-models cache is warm AND userFallbackModels is
non-empty, the function verifies userModel is reachable via
fuzzyMatchModel. If not, it iterates userFallbackModels and promotes the
first reachable entry. Cold-cache (first-run) behavior is preserved -
the userModel is returned as-is when availability data is unavailable,
matching the existing 'trust the user' contract covered by the
pre-cache test fixtures.
Adds three regression tests covering: (1) unreachable primary + reachable
fallback -> fallback promoted, (2) reachable primary + reachable fallback
-> primary kept (fast path), (3) unreachable primary + unreachable
fallback -> legacy trust-user behavior preserved.
Sync the PR branch with the newest dev branch and resolve the new import-level conflicts in background-agent manager and runtime-fallback tests. Preserve both the delegated bootstrap coverage from this branch and the newer upstream test utilities and runtime wiring changes, then re-verify the affected delegated fallback suites and typecheck.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
The circuit breaker in manager.ts uses recordToolCall to detect when a subagent gets stuck repeating identical tool_use blocks. It passed partInfo.state?.input as the tool-input signature. When a model (Kimi K2.6 in the reporter's case) emits duplicate tool_use parts faster than the tool actually starts running, state.input is still null, so loop-detector falls back to the bare 'tool::__unknown-input__' signature. As soon as one part has state.input populated (next event), the signature flips to 'tool::{actual-args}' and the consecutive counter resets to 1, repeatedly. The breaker never reaches its 20-call threshold.
Add a top-level input?: Record<string, unknown> field to the local MessagePartInfo interface and prefer state.input when present, falling back to the part's own input when state is still pre-running. The OpenCode part payload carries the tool input as soon as the tool_use block is generated, so this fallback restores signature stability across the model's repeated emissions.
Verification: added 2 regression tests in manager-circuit-breaker.test.ts. Test 1 (reproduce) emits 20 part.updated events with only top-level input and asserts the task is cancelled by the breaker — fails before the fix, passes after. Test 2 confirms that when state.input IS present, it still wins over the top-level input (precedence preserved). All 10 manager-circuit-breaker tests pass, all 20 loop-detector tests pass, typecheck clean.
The task() tool schema declares run_in_background as a REQUIRED parameter, but three prompt-template files contained category-based task() examples that omitted it. When agents copied these patterns into actual tool calls, the runtime validator threw 'run_in_background parameter is REQUIRED', breaking delegation.
Add run_in_background=false to the four offending category-based examples: 2 in src/agents/dynamic-agent-category-skills-guide.ts (Delegation Pattern + Category Domain Matching CORRECT/WRONG pair), 1 in src/agents/hephaestus/gpt-5-3-codex.ts (frontend example), and 1 in src/agents/sisyphus/gemini.ts (Example 4 DELEGATE). Subagent-type explore/librarian examples already correctly use run_in_background=true. atlas/shared-prompt.ts examples were already correct.
Verification: bun test src/agents/ — 373 pass, 0 fail. bun run typecheck — clean. Grep confirms no remaining category task() examples lack the parameter.
Cubic AI reviewer flagged the use of the untrusted Host header as the
URL base in startCallbackServer. The server only binds to 127.0.0.1,
so hardcoding "http://127.0.0.1" as the URL base is robust against
malformed or manipulated Host values and matches upstream behavior
prior to the node:http refactor.
The dist-bundle regression tests in
src/shared/dist-bundle-bun-globals.test.ts are guarded by
`test.skipIf(!existsSync("dist/index.js"))` and dist/ is
gitignored, so they silently skipped in CI which ran tests before
the build step. Adding the build step earlier ensures the
regression guard runs and a future raw `Bun.*` leak in the bundle
fails CI.
Existing test only checked `globalThis.Bun` top-level destructures
and `__require` calls. Add two new test cases:
1. Raw Bun runtime API scanner: scans dist/index.js for any
`Bun.<anyMethod>(` or `Bun.<anyMethod>.` call outside shim-safe
patterns (runtime.Bun, globalThis.Bun, typeof Bun, and the
"Bun is not defined" error-message string). Uses a negative
lookbehind so shim indirection (`runtime.Bun.foo`) passes.
2. Node smoke test: imports dist/index.js under
`node --input-type=module` and asserts stderr contains no
`ReferenceError` and no `Bun is not defined`. The existing
case 3 only checked exit code, which masked lazy-evaluation
crashes that fire after import. Reading exports forces lazy
module-level evaluation paths to run.
bun-install.ts streamToText() was reachable from the plugin bundle via
the cli/config-manager barrel re-export. Replace with the WHATWG
standard `new Response(stream).text()` pattern which works
identically in Bun and Node and avoids the last raw Bun.* runtime
call in dist/index.js.
The OAuth callback server was using Bun.serve, which would crash if
mcp-oauth code paths ever entered the plugin bundle. Switch to
node:http.createServer with the same WHATWG behavior:
- Binds to 127.0.0.1, preserves 200/400/404 status codes
- Translates fetch(Request)→Response to (req, res) callback style
- Clears OAuth timeout on success/error/missing-param paths
- Replaces server.stop(true) with server.close() for shutdown
Functional behavior and response bodies unchanged.
YeonGyu-Kim