fix(agent-manager): decode JSON-encoded task arrays

[marius-kilocode]

What Problem This Solves

Agent Manager rejects a start request when a model returns tasks as a JSON-encoded array instead of an array. The outer tool arguments parse successfully, but CLI validation rejects the nested string before permission checks or the extension start event, so no session is created.

Fixes #13596.

Why This Change Was Made

The advertised schema already correctly describes tasks as array-or-null. Provider adapters parse the outer arguments but do not decode nested JSON strings. Changing the advertised schema would encourage the incorrect format and could disturb the existing flat, nullable schema used for provider compatibility.

Decode one JSON layer only for the runtime start operation, then validate the result with the existing task schema and 1–20 item bounds. Keep native arrays, management actions, permission checks, and the extension event contract unchanged. Malformed JSON, decoded non-arrays, double-encoded strings, invalid task objects, empty arrays, and oversized arrays still fail before permission or event publication.

This is a defensive compatibility fix. It does not establish that MiMo generated the reported payload or that an SDK change caused the reported version regression.

User Impact

Local and worktree session requests can proceed when the task list is otherwise valid but JSON-encoded. Models still receive the array-based schema, and genuinely invalid task lists remain rejected. A patch changeset is included.

Evidence

• New regression cases failed on the original decoder and passed after the change. They exercise the actual runtime schema and initialized tool, including normalized task arrays in the published event and zero permission/event side effects for invalid input.
• 49 focused CLI tests and 13 extension boundary tests passed. CLI typecheck, extension compile (including host/webview typechecks and lint), formatting, and the worktree annotation guard passed. Focused lint retained its seven pre-existing warnings with no new warnings.
• vscode-self-test used a credential-free scripted OpenAI-compatible provider and the real extension/backend flow. Before the fix, encoded tasks failed. After the fix, the same payload created a session; native arrays still created sessions; invalid encoded tasks created none. The captured advertised task schema was identical before and after. The isolated test processes were stopped afterward.
• Live MiMo and Windows were not tested. The extension rebuild completed through its installed-Bun fallback after the pinned Bun launcher failed.

Before, the encoded task array is rejected:

After, the same encoded task array succeeds and the extension creates a session:

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (3 files)

• .changeset/agent-manager-json-tasks.md
• packages/opencode/src/kilocode/tool/agent-manager.ts
• packages/opencode/test/kilocode/agent-manager-tool.test.ts

---

_{Reviewed by grok-4.6 · Input: 170K · Output: 16K · Cached: 406.4K}

_{Review guidance: REVIEW.md from base branch main}