fix(agent): scope supervisor review spend per conversation

[Temikus]

TL;DR: supervisor reviews billed every call to one cost-tracker session that nothing ever resets, so the per-session cost_limit_hard became a lifetime budget. Once spent, every later review failed closed to human approval until restart.

Found on my own instance: 60 reviews decided normally, then 17 straight errors over 27 hours, all session "supervisor:default" exceeded hard cost limit, leaving 7 routine run_javascript calls queued on me.

• Session key is now supervisor:{agent}:{convID}, so the budget rolls over with the conversation like every other session.
• Attribution: the session registers under the supervisor's own agent name. The key's first segment otherwise parsed as a phantom agent called supervisor, which hid the spend from cost_summary and skipped the supervisor's own configured limits.
• Diagnosability: a failed review records detail.cause (cost_limit/timeout/provider_error) and the notice names the budget, since "Supervisor unavailable" is wrong when nothing is down.

APPROVE/DENY/ESCALATE are unchanged, and the error fall-through stays fail-closed.

Out of scope: a dedicated supervisor budget knob, and the zero cache hits on review calls (~32k uncached input tokens each).

Evidence from the live instance

audit_events over the last week, category supervisor, every one of them a run_javascript review (everything else is covered by permanent auto-approve scopes, so JS is the only tool that reaches the supervisor):

Window	Outcome
2026-09-14 to 2026-09-20T04:06Z	60 x APPROVE
2026-09-20T04:33Z to 2026-09-21T07:35Z	17 x ERROR ... exceeded hard cost limit

cost_summary at the time: session supervisor:default held $5.01 over 136 messages and 4.36M input tokens, with sessions: 1, under an agent named supervisor that does not exist in config.

Verification

Both fixes are pinned by tests that fail without them. Reverting the session key alone:

--- FAIL: TestSupervisorReview_SpentBudgetDoesNotWedgeOtherConversations
    conversation B statuses = [supervisor_error], want supervisor_approved:
    conversation A's spend must not wedge B

Reverting the RegisterSessionAgent call alone:

--- FAIL: TestSupervisorReview_SpendAttributedToSupervisorAgent
    cost attributed to phantom agent "supervisor"; want the supervisor's own name
    got [{Agent:supervisor ...} {Agent:default ...}]

Tests drive a real Engine, Router and CostTracker with a mock provider, so the limit check runs through llm.Router rather than a stub.

just hook is green: fmt-check, vet, lint, lint-ui, test, test-ui, openapi-check.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GPHZAN5uWR4fMd2QtcLd4D

[github-actions]

🚀 Released in v0.47.3