fix: close control-plane auth gaps and hold daily budgets under concurrency

[SairajMN]

Summary

Four control-plane and budget bugs, each fixed in its own commit and covered by a
failing test that now passes:

1. An unauthenticated write path could start a run, and unauthenticated read/write
   paths could leak or inject authority. The control plane now fails closed on
   every gate.
2. The daily run ceiling / budget check recorded the run only after it finished,
   so overlapping events could all pass the check and all start runs. Ceilings now
   hold under concurrency.

Commits (merge-request scope)

• 7f1d692 fix: gate POST /v1/action behind the control token

  • POST /v1/action resumed a waiting node and re-ran the runtime (spending money)
    with no auth. Same resume as the protected /v1/agent/completions and
    /v1/agent/runs/{id}/resume routes. Now gated with require_control.
  • Proof: tests/test_control_plane_auth.py::test_the_action_route_is_a_write_path_and_fails_closed

• 8e0f9fe fix: gate /facts, /documents, /memory/search behind the control token

  • These memory write/read routes inject evidence the agent later treats as
    authorised, with no token. Now gated with require_control.
  • Proof: tests/test_control_plane_auth.py WRITE_PATHS (facts/documents)

• a2a0a49 fix: gate GET /subscriptions behind the control token

  • GET /v1/agent/subscriptions returned every subscription's allowed side
    effects, budgets and instructions with no auth while the write path was gated.
    Now gated, matching the write path.
  • Proof: tests/test_control_plane_auth.py::test_the_subscription_read_path_is_gated_like_its_write_path

• 2193fa8 fix: hold daily run ceiling and budget under concurrent events

  • admit_run checked max_runs_per_day / daily_budget against the window ledger,
    but the run was only recorded after runtime.run() finished. Concurrent
    process() calls could all observe count/spend = 0 and all start runs.
  • Fix: atomically reserve a run slot (and daily-budget remainder when set) under
    the event-store lock at admit time; settle actual spend after the run without
    double-counting. Failed runs refund the reservation but keep the consumed slot.
  • Proof: tests/test_autonomy_governor.py::test_the_daily_run_ceiling_holds_under_concurrent_events

What breaks (before)

• Anonymous caller can POST /v1/action to approve/reject a pending decision and
  re-run the graph.
• Anonymous caller can read every subscription's authority object
  (side effects, budgets, instructions) via GET /subscriptions.
• Anonymous caller can inject facts/documents (evidence the agent trusts) and run
  memory searches.
• With max_runs_per_day=1, firing several overlapping relevant events
  (asyncio.gather on engine.process) starts multiple runs and records zero (or too
  few) max_runs_per_day refusals — the ceiling is bypassed.

Fix

• s16code/ui/routes.py — gate POST /v1/action with require_control.
• s16code/routes.py — gate /facts, /documents, /memory/search with require_control.
• s16code/events/routes.py — gate GET /subscriptions with require_control (match write path).
• s16code/events/store.py — EventStore.window_try_admit_run (check + reserve in one
  locked section) and EventStore.window_settle_run (reservation -> metered spend).
• s16code/events/governor.py — admit_run routes through the atomic reserve; settle_run added.
• s16code/events/engine.py — settles actual spend after the run; failed runs refund
  reservation but keep the slot.

Test plan

• pytest tests/test_control_plane_auth.py — auth gates fail closed (503 when token
  unset, 401 on wrong token) for every gated path; new read-path test red before, green after
• pytest tests/test_autonomy_governor.py — new concurrent ceiling test red before
  fix on that case, green after
• Sequential max_runs_per_day still bounds runs and records refusals
• Sequential daily_budget still caps the window and shrinks effective per-run budget
• Concurrent events with max_runs_per_day=1 -> exactly one run and three max_runs_per_day refusals
• Self-trigger, source rate limit, triage ceiling, liveness, and morning-report tests still pass
• Full suite: 362 passed (361 prior + new concurrent ceiling test)
• Reviewer: confirm diff has no .env, tokens, credentials, or local EA wiring
• Reviewer: optional — concurrent daily_budget stress (same pattern as max_runs) if desired beyond the unit test

No secrets, real messages, or API keys in this diff.

[theschoolofai]

Session 16 — graded

Score: 0.

Duplicate on both halves. The daily-budget concurrency work is #3 (2026-08-09) and the control-plane auth gaps are #16 (2026-08-13 16:44), both earlier. Solid work, but points go to the earliest filer of each bug.

[theschoolofai]

Session 16 — regraded ✅

Score: +100. The verdict below stands: this is a duplicate of an earlier filing (or, for glc_v5 #23, an enhancement rather than a defect), and it is recorded as such.

What has changed is the credit. On review, the work here was genuinely done: the bug was found independently, the reproduction is real and the fix is sound. Losing a race you had no way of seeing is not a reason to earn nothing, so this is credited at the full 100 even though it does not carry the first-to-file claim.

---

Original assessment, unchanged:

Duplicate on both halves. The daily-budget concurrency work is #3 (2026-08-09) and the control-plane auth gaps are #16 (2026-08-13 16:44), both earlier. Solid work, but points go to the earliest filer of each bug.