full-auto

full-auto

Closed-loop autonomous transcript → running-app pipeline. Transcript in → running local docker-compose stack out. Zero prompts to the operator. Bounded fix loop.

When to use

• A transcript exists (file path, pasted text, or sitting in the prior message)
• Operator wants end-to-end build without being interrupted
• Output should be a running, locally-deployed app — not just a spec

When NOT to use

• Operator can answer questions interactively, wants reviews mid-build → use cto-loop (semi-auto with 2 gates) or sonzai-sdk wizard (spec + plan, no build)
• One-line ask, no transcript → use sonzai-sdk wizard
• Transcript describes something Sonzai SDK doesn't do (trading, payments, video editing) → halt; do not improvise

Triggers

• Operator pastes a transcript and says "build this" / "go full auto" / "ship it autonomously"
• Operator runs /full-auto <path> with a transcript file path
• Operator's prompt mentions "transcript" + "build" / "implement" / "ship" + no requested review

Hard rules

1. Never ask the operator a question. All ambiguity resolves to a documented default. The masterplan + final report surface what got guessed.
2. Always-search-current-state. No package version, install command, docker image tag, or library API from training memory. Verify at runtime. See version-search.md.
3. Bounded fix loop. Max 5 fixer cycles. If still failing, write .full-auto/BLOCKED.md and exit.
4. Local commits only. No git push, no gh pr create, no remote deploys (Fly / Cloud Run / etc.). Operator owns those decisions.
5. No mock SDK calls. Use the real SDK against a real SONZAI_API_KEY. If absent, abort with a clear message (autonomous pipeline cannot prompt for a key).
6. Drift check first. Fetch live OpenAPI before anything else (https://api.sonz.ai/docs/openapi.json).
7. No invented symbols. Every Sonzai call verifies against live OpenAPI or public SDK source.
8. No tenant names. Anywhere — code, docs, commits.
9. postgres default ≠ postgres mandatory. Stateless apps skip the database.
10. Brownfield: integrate, never recreate. Edit the existing repo; don't scaffold alongside.

Pipeline (Phases 0-5)

→ Full detail: pipeline.md

#	Phase	Reference	Output
0-pre	Notify-setup	notify-setup.md	Detect Gmail+Slack MCPs, load recipient config
0a	Transcript analysis	transcript-analysis.md	in-memory synthesis
0b	Project type detection	project-type-detection.md	greenfield or brownfield
0c-G	Tech-stack derivation (greenfield)	tech-stack-derivation.md	tech_stack
0c-B	Brownfield audit	brownfield-audit.md + subagent-prompts/auditor.md	brownfield-context.md
0d	Wizard answer derivation	answer-derivation.md	8 sonzai wizard answers + rationales
1	Masterplan assembly	masterplan-assembly.md + templates/masterplan.md.template	docs/cto-review/<date>-masterplan.md
2	Build	builder-dispatch.md + subagent-prompts/{builder,version-checker}.md	local commits, files in CWD
3	Local deploy + QA	local-deploy.md + qa-loop.md + templates/	running app + QA report
4	Feedback iteration (auto-fixer)	feedback-iteration.md + subagent-prompts/fixer.md	bounded retry loop
5	Final report	final-report.md.template	docs/cto-review/<date>-final-report.md

Required cross-skill dependencies (within this plugin)

• sonzai-sdk (sibling) — archetype playbooks, feature references, spec templates, drift-detection guide
• cto-loop (sibling) — only if operator opted into the semi-auto path; otherwise unused

Optional cross-skill (degrades gracefully if absent)

• sonzai-internal-staff (sibling, install-time gated) — adds workspace awareness when present; full-auto runs fine without it on public sources alone

Optional MCPs (degrades gracefully)

• Browser MCP for UI testing: mcp__chrome-devtools__* or mcp__playwright__*
  • If neither is loaded → UI tests fall back to "API contract exercised only; frontend not driven by browser"
  • If present → drives the UI in a headless browser, captures screenshots + console errors

Red flags (halt and write BLOCKED.md)

• Transcript names a non-Sonzai product as the target stack
• Transcript asks for game state / billing / payments / video editing / trading
• Operator's prompt contains "don't use Sonzai" or "build without the SDK"
• Builder subagent returns status: blocked — do NOT re-dispatch; exit with the blocker
• Auto-fixer cycle 5 reached with QA still failing — write BLOCKED.md, leave compose running, exit
• Drift check failed twice — write BLOCKED.md, exit
• SONZAI_API_KEY missing from environment AND from .env — halt with "set the key, then re-run"

What success looks like

Operator gets:

• A new or modified repo with working code, local commits, no remote push
• docs/cto-review/<date>-masterplan.md (the autonomous masterplan; approval section emitted but not marked)
• docs/cto-review/<date>-final-report.md summarizing what was built, what was tested, what defaults were used, and what's left for the operator
• A running docker compose stack on localhost — operator can docker compose down when ready
• Optionally: .full-auto/BLOCKED.md if something couldn't be resolved within 5 cycles

Operator does NOT get prompted at all during the run.

Versioning

Ships with the sonzai-sdk plugin in sonzai-claude-skill repo. See repo CHANGELOG for release history.