z: closed-loop execution engine

Execute the five-step flow as a continuous state machine. Do not pause to ask permission. Advance autonomously across all phases until VERIFIED_COMPLETE.

1. Physical state sensing

Run the physical state validator before taking action:

ws-delivery-status --json

Physical state	Condition	Mandatory action	Next transition
(Not in worktree)	Primary checkout	Run init to claim issue and create isolated worktree	Enter worktree, sense state
UNKNOWN	ws-delivery-status exited 4: git or GitHub could not be read	Fix the cause in reasons[]; re-run. Never read it as a pass	Sense state
DRAFT	Working tree dirty	Implement minimal logic; verify focused tests pass	Run git commit, sense state
LOCAL_VERIFIED	Committed, unmerged	Invoke auto to push branch and open pull request	PR opened, sense state
IN_REVIEW	PR open	Wait for CI events, resolve review feedback, merge PR	PR merged, sense state
CODE_LANDED_STAGE1	Merged on main	Run ws-delivery-status --assert-complete --reality-probe "<CMD>"	Probe passes (exit 0)
VERIFIED_COMPLETE	Merged + reality verified	Run close cleanup; advance to next phase or conclude	Next phase or done

A clean checkout of main reports CODE_LANDED_STAGE1 for prior tasks. The tool has no state for a task that is not begun. Do not probe unbuilt work; run init to claim an issue and begin the task.

2. State branches

Branch A: DRAFT (Working tree dirty)

1. Apply changes per ssot. Do not add failing items to allowlists ("exclusion is not resolution").
2. Run focused tests per smoke (pytest <file>::<test> -x).
3. Commit with an ASD-STE100 conventional commit message and re-sense state.

Branch B: LOCAL_VERIFIED (Committed, unmerged)

1. Verify HEAD commit is pushable. Open a pull request using auto and re-sense state.

Branch C: IN_REVIEW (PR open)

1. Trigger the review and merge gate tool (ws-pr-gate --merge or platform equivalent).
2. Bind the PR with ccd_pr and wait for ci-monitor-event. Hosts without ccd_pr run ws-pr-gate in the background. Never run a wait loop or a watch command in the foreground.
3. Run long test suites with run_in_background. Resolve actionable feedback and re-sense state.

Branch D: CODE_LANDED_STAGE1 (Merged on main, reality unverified)

1. Formulate a read-only, non-mutating Business Reality Probe <CMD> verifying live behavior. Prohibit bare echo/printf, trivial prints, mutating commands, and active exemptions.
2. Execute gate: ws-delivery-status --assert-complete --reality-probe "<CMD>".
3. If ws-delivery-status exits 0 (VERIFIED_COMPLETE), proceed to Branch E. If it exits 3, fix the probe and re-run it.

Branch E: VERIFIED_COMPLETE (Physical proof achieved)

1. Invoke close to record evidence and clean up.
2. If subsequent phases exist, do not pause or ask permission; begin next phase immediately.

3. Loop invariants

1. Continuous Autonomous Drive (Zero Decision Offloading): Never ask permission between phases. The agent owns technical execution end-to-end. Stop only for Stage 3 production deploy.
2. Multi-Phase Goal Progression: When goals require multiple phases, chain them sequentially through the state machine until all goals reach VERIFIED_COMPLETE.
3. Zero self-grading: Completion requires ws-delivery-status --assert-complete --reality-probe '<cmd>' to exit 0 (VERIFIED_COMPLETE). Without --assert-complete, exit 0 means only Stage 1 Landed.
4. Anti-Exclusion Invariant: Exclusion is not resolution. Probes with active exemptions fail.
5. Compaction resilience: When resuming after disruption, run ws-delivery-status --json to recover state without asking the user what to do.