test(app-server): cover external-agent cancellation and approval paths

[andrei-hasna]

Summary

Adds regression coverage for live external-agent runs before /external-agent execution is ungated, per todos task 5ddd56e4. Tests only — no product behavior changes; the current deny/reject gating is asserted as a regression guard.

Unit tests (thread_external_agent_processor)

• Terminal events (Completed/Failed/Cancelled) mark the run terminal (so execute() cannot double-emit Failed); progress events (Status/ProposedAction) do not.
• Terminal events with no subscribers are dropped yet still recorded as terminal (late-event filtering).
• Events fan out to every subscribed connection.
• request_permission denies by default (RejectOnce) — approval-denial guard.
• perform_action rejects managed actions with the exact gating reason and emits a ProposedAction notification — action-routing-failure guard.
• is_cancelled tracks the CancellationToken.

Integration tests (suite/v2/thread_external_agent)

• Cancel rejects empty and unknown run ids with the expected messages.
• An active live run reaches a terminal event and is removed from the active-run registry, so a follow-up cancel reports it inactive. Terminal event is accepted as Cancelled or Failed so the registry-cleanup guarantee holds regardless of sandbox availability.

Validation

• cargo test -p codex-app-server --lib external_agent — 63 passed.
• cargo test -p codex-app-server --test all thread_external_agent_cancel — 2 new integration tests pass.
• cargo fmt --all, cargo clippy -p codex-app-server --all-targets — clean.

Note: the pre-existing thread_external_agent_start_emits_run_event_and_validates_runtime test asserts the claude runtime is gated on "missing runtime"; it fails in environments where a real claude binary is on PATH. This is environment-dependent and untouched by this change.

🤖 Generated with Claude Code