fix(agent-editor): stop dropping behavior_policy.supports_team

[arnaud-moona]

Description

The advanced panel of the custom agent editor is a free-form JSON editor, but handleJsonChange
whitelists the keys it keeps. For behavior_policy it preserved only supports_side_question, so
typing

{ "behavior_policy": { "supports_team": true } }

validated, showed no error, and was silently discarded on save — the value never reached
/api/agents/custom, which does accept it.

That flag gates team-mode eligibility and is seeded false for custom agents, so a custom ACP agent
could not be made team-selectable through the UI at all, even when it advertises
agentCapabilities.mcpCapabilities.stdio: true at initialize. The only remaining routes were the
HTTP API (business endpoints require a session, so unreachable from outside the app) or editing
aioncore's SQLite directly.

The change parses supports_team alongside supports_side_question, building the policy object
incrementally so neither key clobbers the other; surfaces both keys in the skeleton JSON so the panel
is self-documenting; and declares the field on CustomAgentAdvancedOverrides and BehaviorPolicy,
where it had been dropped while the backend still reads it.

Serialization moved to advancedOverrides.ts so the round-trip is unit-testable — the parsing lived
in an inline useCallback and was unreachable from a test. Behaviour is preserved, including two
pre-existing quirks now pinned by tests rather than left implicit: valid JSON that is not an object
("s", null) clears the error and keeps the previous bag, hence a three-way outcome
(ok / ignored / invalid); and an array parses as an empty bag, because typeof [] === 'object'
has always let it through.

Related Issues

• Companion issue on deriving team_capable from the handshake (backend-side): to be linked

Type of Change

• fix — Bug fix (non-breaking change which fixes an issue)
• feat — New feature (non-breaking change which adds functionality)
• perf — Performance improvement
• refactor — Code restructuring (no behavior change)
• Breaking change (fix or feature that would break existing functionality)
• docs — Documentation update

Atomic PR Checklist (Rule 1)

• This PR contains exactly one feature or bug fix that cannot be further decomposed
• The PR title follows Conventional Commit format, for example fix(scope): subject (English)

The three commits are the fix, its regression test, and the formatting pass — one change, not three.

Local Checks (Rule 3)

• bun run format — changed files formatted; full format:check passes (2084 files)
• bun run lint — 0 errors (912 pre-existing repository warnings only)
• bunx tsc --noEmit — no type errors
• bunx vitest run — 518 files passed, 1 skipped; 4,957 tests passed, 5 skipped, 0 failed
• i18n validated (node scripts/generate-i18n-types.js + node scripts/check-i18n.js)
• New/changed user-facing text uses i18n keys (no new user-facing text)

Runtime Verification

• Verified on macOS
• Verified on Windows
• Verified on Linux
• I have performed a self-review of my own code

Screenshots

Not applicable; the panel already existed and its rendering is unchanged. The only visible difference
is that the skeleton JSON now lists supports_team alongside supports_side_question.

Additional Context

Tests. tests/unit/settings/advancedOverrides.test.ts covers the round-trip: supports_team
survives alone and alongside supports_side_question, non-boolean flags are ignored rather than
coerced, blanks are still dropped, and buildAdvancedJson → parseAdvancedOverrides preserves what
the panel displays. Writing them caught a mistake in the first draft, which asserted arrays were
ignored — that would have been a silent behaviour change smuggled into a bug fix.

Coverage. Codecov flags the patch at 16%. The uncovered lines are the thin component-side binding
in InlineAgentEditor.tsx, not the logic, which now lives in a separate module and is covered.
Covering the remaining branches would require a DOM test mounting the whole editor (CodeMirror, Arco,
ipcBridge, the emoji picker) to exercise three ifs. Happy to add one if you would rather have it.

Flaky test in CI, green locally. An earlier CI run on this branch failed on
tests/unit/renderer/mermaidBlockPanZoom.dom.test.tsx. The full suite passes locally on the same
commit (4,957 passed, 0 failed), so the failure is load-sensitive rather than a regression — and it
is unrelated either way: the test exists on main, this branch neither adds nor modifies it, and
MermaidBlock shares no code with the files changed here.

The trace points at a race: the test captures inner via diagram.firstElementChild immediately
after findByTestId, so if the async mermaid render replaces that node the reference is detached and
its transform never updates — which matches the observed scale(1) after a zoom-in click. Flagging
it in case it is worth hardening separately.

Scope. This makes the existing flag reachable. It does not address why a custom agent that
declares MCP stdio needs a manual flag at all, when team_capable is documented as "MCP stdio
capable, computed by backend" — the stored mcp_capabilities currently keeps only http and sse,
so the advertised stdio never reaches the catalog. That belongs in the backend and is filed
separately.

[codecov]

Codecov Report

❌ Patch coverage is 86.66667% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...pages/settings/AgentSettings/InlineAgentEditor.tsx	20.00%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!