v1.3 - Output Quality Gate and Promotion State Machine

New in v1.3

Three new rules (23-25) that formalize automated output verification and workflow tier management:

• Rule 23: Output Quality Gate -- Every iteration must pass automated QA checks before being counted as successful. Artifact existence alone is insufficient.
• Rule 24: Promotion State Machine -- Projects with test/staging/production workflow tiers get an automated promotion ladder. Consecutive QA passes promote; consecutive failures demote or halt.
• Rule 25: QA Report Archival -- QA results logged to append-only JSONL for trend analysis and regression detection.

Origin: Derived from OTR v2.0, where Iteration 5 falsely declared success and Iteration 6 was the first real end-to-end visual pipeline run.

Total rules: 25 (was 22 in v1.2)

v1.2 - Operational Discipline

v1.2 - Operational Discipline

14 new protocol rules derived from real production failures.

New Rule Categories

• Operational Discipline (9-14): Completion verification, heartbeat, mirror freshness, env health, artifact manifests, iteration tagging
• Session/Task Lifecycle (15-18): Session start/end, task ownership with TTL, file intent, atomic fixes
• Conflict Resolution (19-22): Env change notification, canonical timestamps, dispute resolution, priority interrupt

Every rule traces to a specific failure during the OldTimeRadio v2.0 project (Claude + Antigravity on RTX 5080).

Service Restart Policy (Rule 20)

• Projects define a RESTART_POLICY in HANDSHAKE.md with triggers, no-restart conditions, and exact restart procedure
• Operator agents execute restarts autonomously -- human never restarts the service manually
• Integrates with ENV_CHECK: failed health check triggers restart, re-check, then proceed or escalate
• Rule Zero: source code changed = restart, config/data changed = no restart
• Hot reload attempted first; full kill-wait-relaunch as fallback

Git Push Protocol (Rule 21)

• 5-method escalation: GitHub API (most reliable) > git+token > git+SSH > Windows MCP > human command block (last resort)
• Full GitHub API recipe with code: blob > tree > commit > update ref
• Token management: never commit tokens, never log to shared files
• Lockstep verification mandatory after every push
• Common failure mode fixes (index.lock, credential prompts, BOM, timeouts)

v1.1 — Blocked Agent Protocol + Sync Check

What's new

• Blocked Agent Protocol — when an agent can't complete a task (wrong role, no terminal, resource limit), it logs [BLOCKED], tails the other agent's mirror log, and hands off. No silent freezing. No human middleman.
• 3-Strike Rule — same failure 3 times? Full stop. Tail both logs. Write [BLOCKED-3X]. Act from context, not blind retry.
• sync_check.py — run python scripts/sync_check.py from any platform to get a health report: which agents are active, when they last spoke, any open BLOCKED/STALLED entries, unresolved push requests.
• Protocol template updated with rules 6-7 in the agent prompt section.

Install

Same as v1.0 — copy the agent-handshake/ folder into your project root and run python scripts/discover_agents.py --init .

v1.0

Initial release — multi-agent collaboration protocol