Engineering Autonomy program (Phase A)

Phase A proves that one authenticated Matrix task, one fresh forgeadmin approval bound to that exact task, and one confined agent run can edit a disposable repository (src/calc.py → pure add(a, b)), be independently…

Updated

Executive capsule

Phase A proves that one authenticated Matrix task, one fresh forgeadmin approval bound to that exact task, and one confined agent run can edit a disposable repository (src/calc.py → pure add(a, b)), be independently verified on the changed bytes, report the result to Matrix, and leave a checksummed archive. This page is the tracker of record. Until 2026-10-27 that state lived only in workbench/*.zip archives and chat; it is canon here from now on.

Why this matters

Without a single tracker, each session re-derives the program state from archives and chat, and the temptation to call a gate PASS from an exit code grows. This page fixes the vocabulary and the evidence trail.

Who this is for

  • Operators deciding what to approve next in Element.
  • Architects judging whether the execution substrate (laptop vs Granite) matches the policy.
  • Agents that must know which gates are open before claiming progress.

How it works

An operator posts a task in Element; Roadrunner registers a pending request with a canonical digest; forgeadmin posts !approve <request_id> <digest>; the receipt is stored once and cannot be reused; the executor runs only after the receipt matches; a verifier outside the executor checks the changed bytes; the result and a checksummed archive close the task. Each phase below hardens one link of that chain.

Phases

Phase Name State (2026-10-27)
P00 Runtime truth DONE — evidence P00-1-runtime-truth
P01 Matrix admission and approval binding DONE — !approve <request_id> <digest> receipt gate live; receipts immutable; second claim denied; bot cannot self-approve
P02 Minimal task, attempt, verification, workcell-result contracts DONE — P02-1..3
P03 Autonomous execution ACTIVE — see tracker below
P04 Adaptive task decomposition NOT STARTED
P05 Economic model routing NOT STARTED
P06 Element / Matrix UX NOT STARTED (thread approve shape agreed and tested, not wired)
P07 Replay and scaling NOT STARTED

P03 increment tracker

Increment State Note
P03.6f4 PARTIAL Host AF_UNIX isolation stands; stage race closed in f4a
P03.6f4a Accepted offline Descriptor-relative stage; no host worktree mount
P03.6f5 PARTIAL One-file write-back could publish outside the checkout
P03.6f5a BLOCKED_ENVIRONMENT BLOCKED_UNSAFE_PUBLICATION refusal; no kernel conditional replace
P03.6f6 NOT STARTED Confined deterministic verification of a written checkout
P03.6g BLOCKED_POLICY on laptop Real Cursor activation: transport proven to api2.cursor.sh:443; Cursor agent then needs agentn.global.api5.cursor.sh:443, which is not allowlisted. Decision: do not widen the allowlist
P03.7 NOT STARTED Matrix end-to-end canary on the chosen rail

Roadrunner agentic-dev leaves (W1–W4)

Leaf State Shipped as
R01 forge-agents in git + CI DONE autowww/forge-agents main, CI with sibling clones
R02 program canon in repo DONE this page
R03 agent branch lane + rulesets PARTIAL policy + docs/rulesets/*.json; rulesets applied by operator
R04 Fleet agentic_dev_task + source guard DONE forge-fleet v0.3.165
R05 credential plane (courier, git identity) PARTIAL courier service in forge-agents; Granite FLEET_AGENTIC_GIT_TOKEN pending
R06 Roadrunner service + !agent task + broker DONE forge-agents 926edc4; Fleet rollout route v0.3.166
R07 Element repo-task protocol DONE protocol page
R08 code_review_worker + HITL merge DONE Fleet run_review.py v0.3.166; forge-workcells code_review_worker
R09 freshness, branch GC, propagation DONE Fleet stale/hold matrix, /v1/admin/agentic-dev/branch-gc
R10 agentic_dev class across peers DONE node: auto, mesh advise workload: agentic_dev
R11 dual-wiki closure DONE FA-ADV-001, FA-REV-001 pairs; platform catalog/approval model; handbook rebuilt
R12 Element one-task canary NOT STARTED autowww/forge-calc-canary seeded; operator-driven; needs R03/R05/R06 operator steps
R13 web-GPT API/MCP bridge PARTIAL forge-web-bridge + guard + Open WebUI API keys (forge-fleet); lcdl.web_chat MCP tool (forge-lcdl); Granite rollout + key mint pending

How it fits the ecosystem

forge-agents hosts Roadrunner (Matrix adapter, approval receipt, broker); forge-fleet is the only Granite execution rail (cursor_cli, agentic_dev_task); forge-workcells hosts the review workcell; forge-platform owns the approval model, envelopes, and this tracker; Lenses remains the ForgeRun system of record.

Phase A exit gates

Gate Definition State
R1 Approval binding: the executed task digest equals the digest in the forgeadmin !approve receipt; receipt consumed once Mechanism proven; no PASS run yet
R2 Independent verification on changed bytes: a verifier outside the agent process checks the pushed/written file, not the agent's claim No changed bytes yet (src/calc.py still placeholder)
R3 Result and archive: Matrix result event ids, checksums.sha256, archive SHA-256 recorded Archives exist for FAIL runs only

Phase A exit is not claimed.

Exit blockers (current)

  1. No Cursor session has produced a patch on the confined laptop rail (two runs failed: TLS facade defect, then non-allowlisted agent endpoint).
  2. All prepared calc approvals are consumed or expired; the next run needs a fresh receipt.
  3. The Roadrunner code base (forge-agents) only entered git on 2026-10-27 (autowww/forge-agents); Granite could not receive it before.
  4. No repo-aware Granite job exists yet (Fleet cursor_cli host job v0.3.164 is generic).

Decisions

Date Decision
2026-10-27 Execution moves to the Granite Cursor CLI rail (Fleet-managed container, credential supplied from localhost at submit time). The local egress allowlist (api2.cursor.sh:443) is not widened.
2026-10-27 Remediation proceeds through pdca-prompts/roadrunner-agentic-dev/ (R01–R13): git-only code transfer, no SSH, agent branch lane, Element report protocol, separate review workcell with HITL merge.

Evidence log

Archive SHA-256 Verdict
engineering-autonomy-PhaseA-analysis-20261027T101410Z.zip de673539941176370547f97313ac3f388c378bab2c2e876c7a2b485480e7788b BLOCKED (analysis; thread shape tested)
engineering-autonomy-PhaseA-recovery-20261027T113415Z.zip 9b94ca1eb661e950e106f70bb7a667e473d458664fe87a60eeeefe93cb49c023 FAIL (no patch)
engineering-autonomy-PhaseA-run-20261027T121742Z.zip ad85cec0e18620a66a98bbd690c73ebf624180aabcbdf8d6c7167de3184c8a6a FAIL (appr_task_ce8c9548; TLS facade defect, fixed afterwards)

Later run on req_task_3a7d3685 (appr_task_452bf91b): FAIL, exit 1, agent endpoint not allowlisted; staged under workbench/phase-a-run-stage/.

Trust boundary

A human approves every write-capable run with a fresh receipt. Agents cannot approve themselves, cannot reuse a receipt, and cannot claim PASS; only the independent verifier and the archive do. Widening egress allowlists, reusing approvals, or editing archives are outside what any agent may decide.

How to use this page

Before approving anything in Element, read the exit blockers. After any new evidence archive, add a row above with its SHA-256 and the true failing stage. Status words are limited to DONE, ACTIVE, PARTIAL, BLOCKED_*, NOT STARTED, PASS, FAIL.