Plan: TAC RSI Loop Engineering with Plan F3

Purpose

Create a Plan F3 implementation plan for turning the current TAC Product Factory into a truthful loop-engineering system: HTML-first plans, signed deploy boundaries, verification receipts, event-driven operation, and carefully scoped RSI claims.

Problem

The repo already has pieces of RSI and Product Factory behavior, but the product narrative can drift into overclaiming. Factory-level deploy loops are operational, while broader agent-level auto-patch remains unproven except for tracked canary recovery evidence. The system needs a boring, machine-checkable plan that prevents future agents from confusing evidence, aspiration, and production truth.

Solution

Use Plan F3 as the canonical planning fabric. Encode the four loops directly in product metadata and receipts, validate plan/deploy/RSI truth in tests, and keep the deploy surface signed-action-only through git-proxy:8099/deploy. The lazy win: constants, receipts, and tests; no new framework.

Implementation Phases

IMPORTANT: Execute every phase and task step by step, in order, top to bottom.

Status markers: [] idle · [wip] in progress · [x] complete · [f] failed.

[x] Phase 1: Lock Current Truth

Normalize Product Factory truth into constants and tests.

1.1 RSI claim constants

• [x] Keep RSI_DECISION = healthy_but_autopatch_unproven.
• [x] Keep auto_patch_proven = false until degraded-skill recovery receipt exists.
• [x] Track rsi_canary_recovery_evidence = true separately from broad autonomy.

1.2 Testing Strategy

Use pytest to lock the claim model.

• [x] uv run --with pytest pytest -q — proves current factory tests pass.

🔁 Do not exit this phase until all product factory tests pass.

[x] Phase 2: Encode Four Loops

Expose the loop-engineering model as metadata, not prose-only marketing.

2.1 Product metadata

• [x] Add loop_engineering.agent_loop: ADW phases run tools until completion.
• [x] Add loop_engineering.verification_loop: plan validation, pytest, verifier_result, receipts.
• [x] Add loop_engineering.event_driven_loop: webhooks, cron, JSONL events, dashboard refresh.
• [x] Add loop_engineering.hill_climbing_loop: signed patch requests, rollback, receipts.

2.2 Testing Strategy

Assert loop metadata exists in generated spec and receipt.

• [x] uv run --with pytest pytest pipeline/test_product_factory.py -q — proves metadata is emitted.

🔁 If metadata appears only in README/dashboard and not receipts, fail the phase.

[x] Phase 3: Preserve Signed Deploy Boundary

Prevent future agents from reintroducing raw shell deploy through product flows.

3.1 Deploy route assertions

• [x] Assert all product outputs use git-proxy:8099/deploy.
• [x] Keep deploy-webhook:8098 only as legacy/internal note.
• [x] Confirm deploy payloads remain signed actions such as patch_skill_from_pr.

3.2 Testing Strategy

Run product and deploy-webhook tests without live VPS calls.

• [x] uv run --with pytest pytest pipeline/test_product_factory.py pipeline/test_deploy_webhook.py -q — proves local signed deploy contract.

🔁 Do not run live deploy checks unless explicitly authorized.

[x] Phase 4: Add Plan F3 Meta-Prompt

Make future plans inherit the loop model automatically.

4.1 Meta-prompt file

• [x] Create plans/meta-prompts/loop_engineering.md.
• [x] Require all Product Factory plans to include deploy route, receipts, RSI claim scope, and four-loop mapping.
• [x] Reference this meta-prompt from generated plan templates or docs.

4.2 Testing Strategy

Use existing plan validator checks; keep it file-based.

• [x] uv run --with pytest pytest pipeline/test_plan_validator.py -q — proves strict plan validation still works.

🔁 If this needs a new framework, stop. A markdown meta-prompt is enough.

Validation Commands

Execute these commands to validate the entire plan is complete:

• [x] uv run --with pytest pytest -q — all repo-scoped tests pass.
• [x] rg -n "git-proxy:8099/deploy|auto_patch_proven|loop_engineering" products pipeline plans — expected truth markers exist.
• [x] rg -n "command\"\s*:" deploy_webhook.py dark_factory.py trigger_webhook.py adw_modules products — no raw command deploy payload in product path.

Notes

Tradeoffs

• Skipped a new orchestration framework; constants and tests are enough.
• Skipped live deploy probing; local contract tests are safer unless explicitly requested.
• Separated canary evidence from broad auto-patch proof to avoid RSI overclaim.

References

• disler/planf3
• Anthropic: When AI builds itself
• products/receipts/rsi-proof-20260619T093724Z.md
• products/AGENT_PRODUCTS.md

Amendments