Unattended Driving

Unattended Driving (the Driver) is the flow engine introduced in RunLogos 1.0.0. It drives one or more AI agents to take a change proposal all the way from planning to deployment and archival — you only step in at a few key confirmation points, and can walk away from the screen the rest of the time.

In one sentence: you decide what to build and whether to release; the Driver makes sure the agents finish every step in between — correctly, and to convergence.

The ten-stage pipeline

1. Write proposal — the agent drafts proposal.md and tasks.md
2. Plan tasks — break down [delta] / [deploy] task lists
3. Write deltas — produce spec change files (requirements / design / API / tests)
4. Merge deltas — merge into the main specs (spec-complete)
5. Plan slices — the slice planner cuts self-contained slices from finalized specs and real test IDs
6. Implement — deliver business code + tests + the OpenLogos reporter, slice by slice
7. Verify — full regression via openlogos verify (failures enter a repair loop automatically)
8. Deploy — release per the merged deployment plan (automatic in full-auto; confirmed in semi-auto)
9. Smoke — production smoke via openlogos smoke
10. Archive — close out the proposal (automatic in full-auto; confirmed in semi-auto)

Every stage lights up in the CLI panel in real time, and the proposal lifecycle log page can replay the whole run.

Two authorization modes

Automatic never means untraceable — auditing lands in two places: each auto-passed flow gate appends a line to GATE_AUTO_PASSED, while command actions (verify / deploy / smoke / archive) leave their own existing markers (VERIFY_PASS, DEPLOY_DONE, SMOKE_PASS, and the archive record) rather than synthetic gate lines.

The one hard red line: loop-exhausted

Standing authorization has exactly one boundary that no mode may cross — the loop-exhausted red line: when the repair loop hits its round limit and tests are still failing, unconverged code is never auto-released in any mode, fully automatic included. The dialog offers only "take over and fix" or "abort" — there is no "confirm release" button.

Remember the boundary in one sentence: auto-pass only ever happens after the code is green; red code always stops for a human. In semi-automatic mode, the deploy, smoke, and archive gates additionally stop for your confirmation each time.

The multi-agent review gate

With multi-agent review enabled, a second agent adversarially reviews the primary agent's output at key nodes (proposal / deltas / code):

• The Driver recomputes the gate verdict from the finding snapshot — it never trusts the reviewer's self-reported verdict;
• A BLOCK enters a triage sub-loop: the primary agent addresses or rebuts every finding, then re-review;
• The loop is bounded: after the round limit, your configured conflict policy (stop for me / reviewer wins / producer wins / strict-then-lenient) decides;
• Progress is visualized as a six-state sub-row in the panel, with each round's review and triage text available for inspection.

Self-healing

• Verify repair: on failure, the failure summary is dispatched back to the agent to fix, then verify reruns until convergence or a clear stop;
• Smoke repair: fix, then rerun verify first, then redeploy if needed, then smoke again;
• Blocked triage: structural gaps in proposal artifacts enter a bounded triage-repair inner loop, with the gate as the final judge;
• Watchdogs: the completion handshake, heartbeats, and dispatch barriers all carry watchdogs with lease renewal — eliminating both "working but judged stuck" and silent hangs.

Next steps

• Follow the Quick Start to complete your first unattended run
• UI reference, configuration, and troubleshooting guides are currently available in Chinese