Stage 2 ran against a real daemon, and both live tests pass

.github/workflows/build-deploy.yml

@@ -81,6 +81,7 @@ jobs:
           STACK_NAME: ${{ env.STACK_NAME }}
           STACK_DIR: ${{ env.STACK_DIR }}
           IMAGE_NAME: ${{ env.IMAGE_NAME }}
+          TEST_IMAGE_NAME: ${{ env.TEST_IMAGE_NAME }}
           IMAGE_TAG: ${{ github.sha }}
           GIT_AUTH_TOKEN: ${{ secrets.GIT_AUTH_TOKEN }}
           KOMODO_API_KEY: ${{ secrets.KOMODO_API_KEY }}
@@ -94,4 +95,15 @@ jobs:
             -H "Authorization: token ${{ secrets.FORGEJO_TOKEN }}" \
             "https://repo.indexarr.net/api/v1/repos/indexarr/ops/raw/scripts/komodo-deploy.sh" \
             -o komodo-deploy.sh
+
+          # The acceptance image first, then the runtime image.
+          #
+          # Both are pinned to this commit's sha, and that is not tidiness. The
+          # acceptance service used to pin `:latest`, and `docker compose run`
+          # does not re-pull a tag it already holds -- so the Stage 2 live suite
+          # silently executed a STALE image and reproduced a defect that had
+          # already been fixed and deployed. Evidence from an image nobody can
+          # name is not evidence. The script only rewrites `image: $IMAGE_NAME:*`
+          # lines, so each call touches exactly its own service.
+          IMAGE_NAME="${TEST_IMAGE_NAME}" bash komodo-deploy.sh
           bash komodo-deploy.sh

AGENTS.md

@@ -64,14 +64,16 @@ trade and it is rejected.
 
 | Thing | Location |
 |---|---|
+| **Minimum generic product target and plan template** | `minimal.md` — local-only lifecycle and its acceptance contract |
+| **Exploration-complete implementation backlog** | `BACKLOG.md` — dependencies, decisions, tests, and evidence for every target gap |
 | How to use it | `docs/USAGE.md` — worked example, real output |
 | **Starting a project from nothing** | `docs/USAGE.md` §0 — the four routes in, and which to pick |
 | **Taking on a project already in flight** | `agent-harness adopt`; `src/agent_harness/adoption.py`. **A proposal is never a decision** — nothing is dropped unless a human names it |
 | The first run, needing no credentials | `agent-harness init --demo`; `src/agent_harness/demo.py` |
 | What is configured and what is missing | `agent-harness doctor`; `src/agent_harness/doctor.py` — reports, spends nothing |
 | How to deploy it | `docs/DEPLOYMENT.md` — the two serve modes, and a non-destructive smoke test |
 | Sample plan | `examples/PLAN.md` |
-| The original plan | `docs/HARNESS-PLAN.md` (superseded in part) |
+| The original plan | `docs/HARNESS-PLAN.md` (historical and superseded) |
 | Issue tracker | GitHub, per D1. The only place an issue's *state* lives. |
 | The manifest that seeded those issues | `docs/backlog-seed-2026-08-02.json` — historical, carries no state, not kept in sync |
 | What a route is made of, and how a vendor is added | `src/agent_harness/protocols.py` |

GUI_PLAN.md

@@ -1,7 +1,16 @@
 # Agent Harness GUI Implementation Plan
 
-**Status:** Imported into the current core tree on 2026-08-06; implementation remains
-incomplete. `docs/STATUS.md` is the authority for current and remaining work.
+> **Target alignment — 2026-08-09:** This is no longer a current product
+> sequence. The generic local-only target is [`minimal.md`](minimal.md), and
+> current state is [`docs/STATUS.md`](docs/STATUS.md). Remote Git, hosted CI/CD,
+> publication, organisation-wide RBAC, and session-host expansion below are
+> outside the minimum target. The implemented same-origin GUI remains part of
+> the product and should expose local admission, integration, and delivery
+> evidence as those capabilities land.
+
+**Status:** Historical implementation plan. Imported into the current core tree
+on 2026-08-06; `minimal.md` and `docs/STATUS.md` now define the target and
+remaining work.
 **Plan date:** 2026-08-05
 **Product boundary:** The GUI is built, packaged, served, tested, and documented entirely
 inside `agent-harness`.

README.md

@@ -1,19 +1,22 @@
 # agent-harness
 
-Turns a plan you wrote in markdown into work that coding agents actually do, and tells you
-honestly what happened.
+Turns a plan you wrote in Markdown into locally integrated, locally accepted
+work that coding agents actually do, and tells you honestly what happened.
 
 ```
-PLAN.md ──▶ GitHub issues ──▶ claim ──▶ agent in a terminal ──▶ checks ──▶ review ──▶ PR
-                                          │
-                                          └── you can attach to it, on any device
+PLAN.md ──▶ validate ──▶ local queue ──▶ isolated agents ──▶ gates and review
+                                                               │
+                                                               v
+local Git integration ──▶ local build/deploy/accept/teardown ──▶ evidence
 ```
 
-It is **not tied to any particular project, language or workload.** You supply the plan, the
-provider and the checks; the harness supplies the queue, the claims, the failure model and
-the record of what happened.
+It is **not tied to any particular project, language or workload.** You supply
+the plan, local topology, execution profile, provider and checks; the harness
+supplies the queue, claims, gates, local integration and record of what
+happened. The target contract is [`minimal.md`](minimal.md). Remote Git,
+hosted CI/CD and non-local deployment are optional extensions, not prerequisites.
 
-## Status: pre-alpha — deterministic paths are tested; real use is observed, not proven
+## Status: pre-alpha — foundations exist; the minimal local lifecycle is not complete
 
 It runs as a standalone service; [AIDevEnv](https://github.com/TheDancingDeveloper-org/aidevenv)
 is an optional reference session host,
@@ -26,6 +29,11 @@ follow-up run, so it is an observation, not proof that the harness works against
 fleet. Deterministic fixture success proves wiring, not model quality or unattended
 reliability.
 
+It also lacks deterministic admission of the new minimum plan, a pinned
+per-project execution profile, and final local build/deploy/readiness/acceptance/
+teardown for the product under development. See
+[`docs/STATUS.md`](docs/STATUS.md) for the capability-by-capability comparison.
+
 Three words are used precisely throughout this README, and they are not
 interchangeable:
 
@@ -102,35 +110,43 @@ worker killed mid-item releases it by doing nothing. A heartbeat keeps genuinely
 alive, because "slow" and "dead" look identical from outside and only a live process can
 keep stamping one.
 
-## Definition of done for v1
+## Definition of done for the minimal local product
 
-Expressed as observed behaviour, not internal completeness.
+The full acceptance contract is in [`minimal.md`](minimal.md). In summary:
 
-- [ ] The fleet runs 7 days unattended with no manual restart.
-- [ ] Every failure is diagnosable from the GUI alone, without opening a log file.
-- [ ] Rate-limit errors are classified, and cost caps are never retried.
-- [ ] No single worker's failure pauses another worker.
-- [ ] Reviewer-approved work survives a killed worker.
-- [ ] Delivery rate is no worse than the workload's own pre-harness baseline, at lower cost.
-- [ ] The role→model map can be changed without a redeploy.
-- [ ] Two projects run concurrently without either starving the other.
-- [ ] Deleting `harness.sqlite` changes no audited answer.
+- [ ] An invalid minimum plan gets one complete deterministic rejection report
+  and creates no executable state.
+- [ ] A valid plan and local Git repo are admitted atomically without remote
+  credentials.
+- [ ] A pinned per-project execution profile passes preflight before work begins.
+- [ ] Work survives the existing checks, review, durability, budget and policy
+  gates and is promoted into a local integration branch.
+- [ ] The integrated commit builds, deploys locally, becomes ready, passes
+  acceptance, and tears down with durable evidence.
+- [ ] Two materially different projects complete without changes to core
+  execution-path modules.
+- [ ] The repository gates and genericity enforcement pass.
 
-If all seven hold, v1 is done regardless of what remains unimplemented.
+None of those unchecked outcomes should be inferred from component tests.
 
 ## Documentation
 
-Two documents carry the current state. Everything else is a how-to, an
-operational runbook, or history.
+Four documents separate the target, implementation backlog, implemented
+design, and current state.
+
+- **[`minimal.md`](minimal.md) — what the smallest useful generic product must
+  do.** This is the local-only target and includes the minimum plan template.
+- **[`BACKLOG.md`](BACKLOG.md) — the exploration-complete work to reach that
+  target.** Each item fixes its boundary, dependencies, implementation
+  decision, failure behaviour, tests, evidence, and non-goals.
 
 - **[`docs/DESIGN.md`](docs/DESIGN.md) — how the harness works, and why it is
   shaped that way.** The execution pipeline, the invariants and the failure
   each one came from, model routing and failure classification, the dependency
   graph, durability, and the extension points. Start here to understand it.
 - **[`docs/STATUS.md`](docs/STATUS.md) — where it stands, and everything
-  outstanding.** What is proven, observed and merely tested; the open work in
-  the order it can be done; and how to run the harness against **rdpapp**, the
-  first application it is being tested against.
+  outstanding.** The direct comparison with `minimal.md`, retained foundations,
+  misalignments, milestone exits, and evidence boundary.
 
 ### How to use it
 
@@ -181,6 +197,11 @@ It has no state field and is not kept in sync; GitHub is the tracker (D1).
 A five-minute tour. The full walkthrough, with real output, is in
 [`docs/USAGE.md`](docs/USAGE.md).
 
+The commands below document the current pre-alpha implementation. Some still
+expose optional GitHub publication because that capability predates the local
+target. The end-to-end `minimal.md` path is not available until the gaps in
+`docs/STATUS.md` are closed.
+
 ### 0. See it work, before you configure anything
 
 ```bash

docs/AUDIT-PLAN.md

@@ -5,6 +5,7 @@
 > The durable audit layer, which is built. What survives is in `DESIGN.md`; the plan itself is history.
 >
 > **Current documentation:**
+> [`minimal.md`](../minimal.md) — the current generic local-only target.
 > [`docs/DESIGN.md`](DESIGN.md) — how the harness works and why.
 > [`docs/STATUS.md`](STATUS.md) — where it stands and what is left to do.
 >

docs/COORDINATION-PLANE.md

@@ -4,8 +4,9 @@
 >
 > **Half superseded, half still the only record.** The typed work graph (§8) is built and is described in `DESIGN.md`. The message ledger, rooms and oversight actor remain **proposed and not built**, and this document is still the only place they are specified — that part is not replaced by anything.
 >
-> **Current documentation:** [`docs/DESIGN.md`](DESIGN.md) — how the harness
-> works and why. [`docs/STATUS.md`](STATUS.md) — where it stands.
+> **Current documentation:** [`minimal.md`](../minimal.md) — the generic
+> local-only target. [`docs/DESIGN.md`](DESIGN.md) — how the harness works and
+> why. [`docs/STATUS.md`](STATUS.md) — where it stands.
 >
 > Where this document and the code disagree, the code is right.
 

docs/DESIGN.md

@@ -8,6 +8,16 @@ here. **Where a claim here disagrees with the code, the code is right and this
 document is wrong**; that is rule 7 of [`AGENTS.md`](../AGENTS.md) and it
 applies to this file first.
 
+The target product boundary is now [`minimal.md`](../minimal.md): validated
+plans, local agent execution, local Git integration, and local product
+build/deploy/acceptance. This file still describes remote issue, branch, pull
+request, and reconciliation capabilities because they exist in the pre-alpha
+implementation. They are optional legacy/extension paths, not prerequisites or
+the target lifecycle. Missing target capabilities are recorded only in
+[`STATUS.md`](STATUS.md); their exploration-complete implementation work and
+dependency spine are in [`BACKLOG.md`](../BACKLOG.md). This design document
+does not duplicate that roadmap.
+
 What is built, what is half-built, what has never run against a real fleet, and
 what is blocked on what: [`docs/STATUS.md`](STATUS.md) owns all of it. Nothing
 here is a status claim, and where a design element exists but is not yet
@@ -17,14 +27,20 @@ reachable, this says so in one clause and moves on.
 
 ## 1. What this is, and the one idea
 
-A queue and a delivery pipeline for a fleet of coding agents.
+A queue and gated execution foundation for a fleet of coding agents. The
+currently implemented remote path is:
 
 ```
 PLAN.md ──▶ issues ──▶ claim ──▶ implement ──▶ checks ──▶ review ──▶ PR
                                                      │
                                every stage recorded, append-only
 ```
 
+The target minimum path replaces the remote endpoints with deterministic plan
+admission, a local queue and integration branch, then an integrated local
+build/deploy/readiness/acceptance/teardown lifecycle. Those gaps are not
+silently treated as implemented here.
+
 You supply the plan, the model provider and the checks. The harness supplies
 the queue, the leases, the failure model, the gates and the record of what
 happened. It writes no code itself and judges no code itself; it decides *what