Openplan auto

OpenPlan Auto

Use this as the orchestration layer for unattended or low-interaction workflows.

Core Rule

Auto mode can gather evidence, classify, review, summarize, and prepare proposals. It cannot make human-owned decisions, accept risk, approve changes, or silently mutate durable goal/intent.

Skill Order

Default chain:

core -> operate -> (align/research if needed) -> action/test/review/garden -> record proposal -> handoff -> html brief

Use only the stages needed.

Stage	Skill	Auto Can Do	Must Stop For
Principles	$openplan-core	Apply invariants	Conflicting human/platform instruction
Startup	$openplan-operate	Load minimal context, choose mode	Destructive or broad action
Alignment	$openplan-align	Infer safe route, ask limited questions only as the human-facing coordinator	Scope/architecture/risk decision
Research	$openplan-research	Gather bounded evidence	Turning findings into decisions
Test	$openplan-test	Choose proportional checks, prove regression coverage, report skipped checks	Approval, merge/release, or risk acceptance
Review	$openplan-review	Read-only findings	Approval or fixes
Garden	$openplan-garden	Semantic doc report	Durable doc rewrite
Record	$openplan-record	Propose memory/docs updates	Ambiguous durable goal/intent update
Docs Init	$openplan-docs-init	Initialize governed document generation	Durable governance/design decision
Handoff	$openplan-handoff	Snapshot state	None, unless it would assert unverified facts
HTML	$openplan-html-brief	Temporary human frontend	Persisting or committing HTML

State Machine

1. INIT: read request and local instructions.
2. CLASSIFY: choose primary mode and surfaces.
3. EVIDENCE: collect minimal durable evidence.
4. DECIDE_ROUTE: bounded action, align, research, test, docs-init, review, garden, record proposal, handoff, HTML brief, or block.
5. EXECUTE_SAFE: perform read-only work or explicitly authorized writes inside scope.
6. VERIFY: run the proportional test, review, garden, or evidence check for the artifact touched.
7. RECORD: propose durable memory/docs updates only when warranted.
8. HANDOFF: write or emit a minimal continuation snapshot when work is long-running, cross-agent, blocked, or risky.
9. REPORT: summarize scope, evidence, actions, decisions needed, next safe action, and residual risk.
10. STOP: stop at the human decision boundary.

Closed Loop

An auto run is closed only when every routed branch ends in one of these states:

• completed in scope: evidence and checks are reported, with no approval claim;
• blocked: the missing artifact, decision, or authority is named;
• human decision required: options and evidence are prepared, but no decision is made;
• handoff emitted: another agent/session can continue from a minimal snapshot;
• disposable brief emitted: HTML is clearly temporary and points back to durable artifacts.

Closure rules:

• If files are changed, run the relevant proportional check before reporting.
• If docs, memory, handoff, or generated documentation are touched, run $openplan-garden when in scope; otherwise report the skipped check, impact, and next safe action.
• If durable memory/docs should change, use $openplan-record as a proposal layer.
• If uncertainty affects route choice, use $openplan-research before design or execution.
• If broad docs generation starts, use $openplan-docs-init before writing durable docs.
• If bounded action changes implementation artifacts, verify with proportional checks and report no approval claim.
• If testing strategy, regression proof, or skipped-check meaning matters, route through $openplan-test.
• If review or garden reports blocking findings, do not auto-fix unless the original task explicitly authorized fixes in that scope.
• If a handoff or report contains facts another agent will rely on, include evidence paths and residual risk.

Human Decision Boundaries

Stop and report instead of continuing when the next step requires:

• accepting or deferring risk;
• changing durable memory or source-of-truth docs from ambiguous input;
• approving, merging, releasing, or declaring success;
• choosing architecture or scope;
• bypassing a safety rule;
• destructive file or branch operation;
• conflicting instructions or unclear authority;
• source-of-truth conflict between docs, memory, specs, or code;
• a subagent needs human input;
• a check is unsafe, unavailable, or outside the allowed scope but materially affects confidence;
• review/garden returns a blocking finding and fixes were not explicitly authorized;
• expanding beyond the user's requested scope.

Report BLOCKED when progress requires a missing artifact, unavailable evidence source, unsafe/out-of-scope check, or inaccessible authority source.

Report HUMAN_DECISION_REQUIRED when progress requires a human choice, risk acceptance, scope/architecture decision, approval, bypass, or durable intent change.

Auto-Allowed Work

Allowed by default when safe:

• read files;
• inspect diffs or provided artifacts;
• run non-destructive checks;
• perform explicitly authorized scoped writes;
• produce alignment summaries;
• produce review/garden reports;
• propose record updates;
• create handoff snapshots;
• create temporary HTML briefs under /tmp.

Output Contract

At the end of an auto or YOLO-style run, report:

Mode:
Scope:
Evidence collected:
Actions performed:
Findings/proposals:
Test evidence:
Checks run:
Checks skipped:
Human decisions required:
Blocked items:
Next safe action:
Handoff snapshot:
Temporary HTML brief:
Residual risk:

References

• references/orchestration.md: detailed flow.
• references/stop-rules.md: stop conditions.
• assets/claude-code-command.md: Claude Code command/prompt seed.