Autonomous Workflow

Iron Law

NO SKIPPING PHASES. Explore, then Plan, then Code, then Verify. Every phase produces an artifact.

Explore > Plan > Code > Verify

1. EXPLORE: Agent(Explore) for unfamiliar code; independent Bash queries in one message; read referenced files. LSP: goToDefinition from issue keywords, findReferences for impact.
2. PLAN: TaskCreate per deliverable, dependencies via addBlockedBy, display the plan.
3. CODE: TaskUpdate(in_progress); implement; commit incrementally (Tier 1); TaskUpdate(completed) only after the per-task gate. LSP: hover before modifying.
4. VERIFY, four mandatory layers: a. Static: lint/test/typecheck in parallel. With lsp.diagnosticsAsQuality, LSP diagnostics add a signal (errors P1, warnings P2), never replace CLI checks. b. Runtime: build, start, verify; on failure enter the debug-fix-retest loop (bounded by closedLoop.maxDebugIterations). c. Review: self-review with fix-forward; fix P1/P2 immediately. d. Verdict: when verdict.enabled, dispatch verdict-judge with acceptance criteria + evidence bundle only (no diff, rationale, journal). Per criterion PASS/FAIL/NEEDS-HUMAN-REVIEW; FAIL enters a fix loop, NEEDS-HUMAN-REVIEW escalates.

Task-Tool Contract

TaskCreate in PLAN (one per deliverable, imperative subject, acceptance criteria in description); TaskGet before working a task; TaskUpdate(in_progress) before starting; TaskUpdate(completed) only after the per-task gate; TaskList at checkpoints.

Per-Task Verification Gate

TaskUpdate(taskId, status: "completed") is blocked until ALL hold:

1. All tests pass (existing and new); otherwise the task stays in_progress in the debug-fix-retest loop.
2. The task's verification command was run and its output recorded as evidence now, not at VERIFY.
3. Every modified file is classified; out-of-context files resolved (separate commit, removed, or user-approved).
4. When settings.json testing.tddMode is enforce (default), RED-GREEN-REFACTOR was observed: failing test before implementation.

Three-Tier Safety

Tier	Actions	Behavior
1 Autonomous	Commits, branches, file edits, staging	Execute
2 Journal	Push, PR creation, issue assignment	Execute, log to journal
3 Confirm	Merge, release, force operations	Human confirmation, always

settings.json tiers; promotable (journal→confirm), never demotable. Full tables: references/three-tier-safety.md.

AskUserQuestion Enforcement

"Use the AskUserQuestion tool" means invoke the tool with contextual options, never plain text.

Decision Journal

• Init: {journal-dir}/issue-{N}.md at branch creation (default .decisions/).
• Log: hooks auto-log edits and commits to a gitignored trail; skills add timestamped entries (category, decision, rationale, risk).
• Summarize for the PR body: public entries only; internal (security rationale, credentials, vulnerabilities) never reaches PR bodies, comments, or logs.
• Anti-estimation guard: no calendar-time estimates in entries; t-shirt sizing (S/M/L) only when asked.

Bounded Verification

Ceilings (qualityCheckMaxIterations, closedLoop.maxDebugIterations default 5, fixForwardMaxIterations) are safety nets, not budgets. Run, fix, re-run until convergence. Halt only for genuine non-convergence (same failure across the last 3 iterations with no progress AND ceiling reached), then file a six-field escalation per skills/llm-operator-principles/SKILL.md § Genuine non-convergence. Never silently loop past the ceiling.

Stop Conditions

Plan >10 tasks for one issue: decompose first. Contradictory EXPLORE signals: ask before planning. >5 files modified without a commit: stop; tasks too large.

Closed-Loop Mandate

Fix failures yourself; never report and move on. Minimum runtime verification per project type: runtime-verification. Only markdown-only, config-only, dependency-bump-only may skip runtime checks (with the evidence that skill requires); any other skip requires a six-field escalation.

Graceful Degradation

No agent teams: sequential. No quality commands: Skill(capability-discovery), then runtime-only. No LSP: grep. No gh: git-only.

Proactive Autonomy

Try first; present 2-3 options with a recommendation, never open questions; use the six fields in references/escalation-format.md. Escalate only for: irreversible actions (Tier 3); genuinely ambiguous product/architecture decisions; out-of-whitelist runtime skips; genuine non-convergence. Never for: Tier 1/2 actions; decisions with clear policy; any P1/P2/P3 disposition (fix it — no follow-up issues in default mode); approaching a ceiling.