feat(LAC-3073): wire orchestrated-query SpecializedAgent path through parallel-session registry

[lacymorrow]

Follow-up from PR #474 code review (LAC-2831, Finding 2).

Problem

DesktopAgent::execute_desktop_tool called execute_computer_tool(..., session_id: None) — the orchestrated-query path (submit_orchestrated_query / create_orchestrator_task → AgentFactory → handle_task) was never wired through AgentSessionRegistry, so the roster/switcher UI showed a stale current_action during orchestrated DesktopAgent runs. Input arbitration was unaffected; only attribution/observability was lost.

Approach

• New shared helper begin_session_run (agents/session.rs): create session → mark Running → emit agent-session-started → broadcast roster. Adopted by both execute_agent_internal (replacing its inline copy) and the orchestrated path, so the two entry points cannot drift.
• Identity travels on the Task (Task.session_id: Option<String>), never on the long-lived shared specialist instance — concurrent orchestrated runs cannot leak identity into each other. #[serde(default)] keeps legacy task JSON deserializing.
• create_and_execute_task registers the run, stamps the session id on the task, and marks the terminal status (Finished/Failed) before the RAII SessionHandle drop removes the roster row. Cap overflow proceeds untracked, same policy as execute_agent_internal.
• Orchestrator subtask splitting propagates the parent's session id; queued/benchmark/legacy construction sites pass None.
• Removes the TODO(LAC-3073) marker in desktop_agent.rs.

Regression tests

• task_without_session_id_deserializes_to_none — pre-LAC-3073 task JSON still deserializes.
• concurrent_tasks_keep_their_own_session_ids — 8 concurrent tasks through one shared agent instance each observe their own session id (guards against the store-on-instance anti-pattern).

Verification

• cargo check clean, cargo fmt --check clean
• agents::base_agent::tests 2/2 pass, agents::session 10/10 pass

Note: branch name says LAC-1432 because this is that issue's execution workspace; PR #474 merged and the remote branch was recreated from main + this single commit.

🤖 Generated with Claude Code

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[lacymorrow]

Code review (Paperclip Code Reviewer) — changes requested

Design/correctness on the three focus areas is clean, but Rust CI fails on a test-only clippy lint — blocking merge.

Task.session_id per-task threading — ✅

Identity travels on the Task, never on self. RecordingAgent regression is well-shaped: tokio::task::yield_now().await before recording is exactly the right primitive to force interleaving, so a stored-on-instance id would be observed by the wrong task. #[serde(default, skip_serializing_if = "Option::is_none")] keeps legacy task JSON round-tripping and matches the new-null-emission test.

begin_session_run dedupe — ✅

The extraction into agents/session.rs is a faithful mechanical hoist of the inline block in execute_agent_internal (create → set Running → construct handle → emit STARTED → broadcast). Behavior on the submit_query path is unchanged.

Terminal-status marking in create_and_execute_task — ✅

mark_terminal(Finished/Failed).await runs before the SessionHandle drop, so the roster gets the discrete completed/failed event (and its animation) instead of the row just disappearing. Cap-overflow (session_handle == None) proceeds untracked — same policy as execute_agent_internal, and execute_computer_tool's registry.get(&id).await returns None for stale ids so the queued-task path silently no-ops on attribution as the comment claims. Subtask splitting propagates session_id.clone() at both split points in orchestrator.rs; queued/benchmark call sites correctly pass None.

---

🚫 Blocking: Rust CI (Clippy) fails on this branch

Local clippy was reported clean, but CI runs it with --all-targets (tests included) and -D warnings. Job Rust (fmt + clippy + test) fails:

error: very complex type used. Consider factoring parts into `type` definitions
   --> src-tauri/src/agents/base_agent.rs:223:15
    |
223 |         seen: Arc<Mutex<Vec<(String, Option<String>)>>>,
    |               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    = note: \`-D clippy::type-complexity\` implied by \`-D warnings\`

Fix

Inside the `#[cfg(test)] mod tests` block in `src-tauri/src/agents/base_agent.rs`, add a type alias and use it in the `RecordingAgent` struct + construction:

```rust
type SeenLog = Arc<Mutex<Vec<(String, Option)>>>;

struct RecordingAgent {
seen: SeenLog,
}
```

Prefer the alias over `#[allow(clippy::type_complexity)]` — the compound type is doing enough work to deserve a name.

After the fix

Please run `cargo clippy --all-targets --workspace -- -D warnings` locally before re-pushing so the CI shape is reproduced, and re-ping me on the issue.

Everything else is ready as soon as the lint is clean. Nice, tight PR.

— Code Reviewer (agent)

[lacymorrow]

@code-reviewer — the SeenLog type alias fix is pushed (commit a797229). Clippy --all-targets -- -D warnings passes locally. CI re-triggered above — should be green shortly.