Historical (2026-08-31): accurate when written; see CHANGELOG for what changed since.

Autonomous Workflow V2 Component Probes

Candidate revision: 0030e12 plus this evidence note
Date: 2026-08-31
Scope: sequence 15 of docs/dynamic-workflow-v2-execution-plan.md

These probes exercised the model-sensitive boundaries separately before any end-to-end autonomous canary. Every paid call used a Relay GPT-5.6 Luna pool. No Minimax or Fable model was eligible. A failed or rejected model response would have stopped the sequence; none occurred.

Results

Boundary	Pool and model	Result	Wall	Known tokens	Durable evidence
Initial planner program	`opencode2` / `relay/gpt-5.6-luna`	Accepted by the real V2 planner-response parser without correction	58.0 s	1,230	`~/.bullswarm/runs/out-1788187364874-j3wtn.md`
Bounded mutating worker	`opencode2` / `relay/gpt-5.6-luna`	Changed only `src/counter.js` in a disposable repository; external tests passed 2/2	32.9 s	241	`~/.bullswarm/runs/out-1788187490363-5dkb2.md`
Evidence envelope and worker self-validation	`opencode2` / `relay/gpt-5.6-luna`	Worker ran `check-v2-evidence.js`; runtime parser accepted one scoped passing requirement with no concerns	50.6 s	487	`~/.bullswarm/runs/out-1788187584619-o9wjr.md`
Consolidated semantic-gap planning	`opencode2` / `relay/gpt-5.6-luna`	Accepted exactly one bounded fix plus one evidence action; no reviewer, repair, or reverify loop	36.7 s	756	`~/.bullswarm/runs/out-1788187686903-z2ubs.md`
Authentication replacement	simulated unavailable Luna, then `opencode2:relay-2` / `relay-2/gpt-5.6-luna`	First attempt classified `auth` and quarantined; second pool completed immediately	0.1 s + 28.6 s	61 + 168	isolated route state and output under the temporary `bullswarm-v2-route-y45TRb` fixture

The routing probe used the production dispatchV2Action and watchOnce seam. Only the first connector was synthetic: it emitted an authentication failure. The replacement was the real configured Relay-2 OpenCode connector. The probe passed only after asserting the exact attempt order, quarantine record, two durable decision-log entries, successful output content, and the replacement model ID. Its isolated Bullswarm home prevented the simulated failure from quarantining a real user pool.

Presentation boundaries

The result, quiet-watch, and human-TUI boundaries are deterministic consumers of durable V2 state, so they were probed without another paid model call:

native V2 runs list, show, and stable result envelope: 1 focused test passed;
compact V2 watch heartbeat, freshness, terminal result handoff: 1 focused test passed; and
durable timeline, live-agent filtering, narrow/mobile layout, and non-clearing spinner repaint: 3 focused tests passed.

The broader result/watch/dashboard component batch passed 54/54 tests. These checks prove each sequence-15 boundary in isolation. They do not count toward the sequence-16 five-run end-to-end streak.

Sequence-15 conclusion

All five model-in-the-loop component probes completed without content rejection, schema correction, semantic repair loops, ownership violations, or use of a disallowed model. Planner validity, worker execution, evidence validation, gap handling, provider replacement, stable result consumption, quiet watch, and wide/mobile TUI rendering are ready for end-to-end canaries.