#21556: fix(agents): graceful fallback when spawned model is not in allowlist

## Summary When `sessions_spawn` is called with a `model` that is not in the agent's model allowlist, the current behavior hard-fails with `status: "error"`, making callers uncertain whether the task started. This leads to accidental duplicate spawns. ## Changes - **Graceful fallback**: when `sessions.patch` rejects the model with "model not allowed" or "invalid model", the spawn continues with the default model instead of returning an error - **Clear warning**: the response includes `warning: "model not allowed: ... — task will run with the default model instead"` so callers know what happened without re-spawning - **Type update**: added `warning?: string` to `SpawnSubagentResult` to make the field part of the public contract - **Tests**: added test coverage for the allowlist fallback path (status="accepted", modelApplied=false, warning present) and the hard-fail path (non-allowlist errors still return status="error") ## Behavior | Scenario | Before | After | |---|---|---| | Model not in allowlist | `status: "error"` (task not started) | `status: "accepted"`, `modelApplied: false`, `warning: "..."` | | Invalid model name | `status: "error"` | `status: "error"` (unchanged) | | Model in allowlist | `status: "accepted"`, `modelApplied: true` | unchanged | ## Notes - This aligns the git source with the behavior already shipped in the 2026.2.17 release build, adding the missing `warning` type field and improving the warning message clarity - Callers should check `modelApplied` to confirm the requested model was applied; `warning` is present when it wasn't <!-- greptile_comment --> <h3>Greptile Summary</h3> Implements graceful fallback when `sessions_spawn` is called with a model that's not in the allowlist. Previously, such calls would hard-fail with `status: "error"`, causing uncertainty about whether the task started and leading to duplicate spawns. Now, when the gateway rejects a model with "model not allowed" or "invalid model" errors, the spawn continues with the default model and returns `status: "accepted"` with `modelApplied: false` and a clear warning message. - Added `warning?: string` field to `SpawnSubagentResult` type (src/agents/subagent-spawn.ts:52) - Modified error handling in `spawnSubagentDirect` to detect allowlist-related errors and fall back gracefully (src/agents/subagent-spawn.ts:205-214) - Updated tests to distinguish between allowlist errors (graceful fallback) and other errors (hard fail) - Added comprehensive test coverage for both "model not allowed" and "invalid model" fallback scenarios <h3>Confidence Score: 5/5</h3> - This PR is safe to merge with no identified risks - The implementation is straightforward, well-tested, and follows existing patterns in the codebase. The string matching for error detection (`includes("model not allowed")` and `includes("invalid model")`) correctly aligns with the canonical error messages from `model-selection.ts`. The type change is additive and backwards-compatible. Test coverage is comprehensive, covering both fallback paths and the unchanged hard-fail path for non-allowlist errors. - No files require special attention _{Last reviewed commit: 29cca80} <!-- greptile_other_comments_section --> <!-- /greptile_comment -->