Audit/reconcile 04a capability recovery 38ca6156

.github/workflows/foundation.yml

@@ -17,4 +17,38 @@ jobs:
           python-version: "3.x"
       - name: Check foundation contract
         run: python tools/check_foundation.py
+      - name: Run baseline fixture
+        run: python -m arkx.baseline
+        env:
+          PYTHONPATH: src
+      - name: Run unit tests
+        run: python -m unittest discover -s tests -t . -v
+        env:
+          PYTHONPATH: src
+      - name: Compile Python sources
+        run: python -m compileall -q src tests tools
 
+  packaged-cli:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.13"
+      - name: Build and install wheel
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install build twine
+          python -m build
+          twine check dist/*
+          python -m venv .venv-release-check
+          . .venv-release-check/bin/activate
+          python -m pip install --no-index --find-links dist codepro
+      - name: Verify packaged public boundary
+        run: |
+          . .venv-release-check/bin/activate
+          test "$(codepro --version)" = "codepro 0.3.0.dev0"
+          codepro doctor --json
+          codepro baseline
+          python -m pip check
+          python -m unittest discover -s tests -t .

.gitignore

@@ -7,6 +7,7 @@ __pycache__/
 .mypy_cache/
 .coverage
 htmlcov/
+*.egg-info/
 
 # IDE and operating-system files
 .idea/
@@ -18,3 +19,6 @@ Thumbs.db
 experiments/**/runs/
 experiments/**/artifacts/
 
+# Local preflight workspaces and temporary virtual environment
+.p82-local-preflight-*-workspace/
+.venv-mini/

AGENTS.md

@@ -0,0 +1,120 @@
+# Codepro agent operating rules
+
+The product identity is Codepro. `arkx` is the current Python namespace and
+must not be treated as a separate product identity.
+
+## Scope
+
+Treat this repository as an evidence-oriented software-engineering project.
+The presence of a module, adapter, test, or document does not prove that the
+corresponding capability is qualified for adoption.
+
+## Required implementation standard
+
+Every material capability change must have:
+
+1. an explicit contract or decision record;
+2. a falsifiable hypothesis and empirical acceptance criteria;
+3. executable tests for deterministic behavior and failure boundaries;
+4. recorded raw evidence, including environment and identity where relevant;
+5. an independent acceptance decision before promotion.
+
+Local unit-test success is local evidence only. It must not be reported as
+upstream, scientific, benchmark, or provider evidence.
+
+## Harness rules
+
+- Preserve provider, model, version, configuration, budget, runtime, and
+  treatment identity in every execution record.
+- Distinguish identity levels: `experiment_id` identifies the protocol,
+  `trial_id` identifies a frozen comparison cell, and `attempt_id` identifies
+  one execution attempt. A verifier `run_id` must be unique for every
+  control, patch, and retry that can produce a distinct result; it must never
+  be reused merely because the task is the same.
+- Use stable identity for equivalent frozen protocol inputs, but include the
+  control role, patch/diff identity, and attempt semantics wherever a provider
+  or verifier can cache results.
+- Execute serially when the experiment requires controlled comparison.
+- Never silently overwrite an attempt directory or select a fallback executor.
+- Persist execution artifacts atomically and keep execution, verification,
+  acceptance, and promotion as separate states.
+- A qualification record must identify the dataset revision or fingerprint,
+  task/base revision, environment image and digest when applicable,
+  OS/architecture, executor/provider/model/configuration, verifier identity,
+  budget, treatment, control role, and whether the provider was called.
+- A provider-free verifier qualification requires both controls: an empty or
+  no-op patch must fail as expected and an official gold/oracle patch must
+  resolve as expected. A missing control is `NOT_EXECUTED` or `BLOCKED`, never
+  a successful qualification.
+- Represent missing identity, missing evidence, blocked infrastructure, and
+  indeterminate outcomes explicitly; none may become success by inference.
+- A failure must retain enough raw output and environment facts to support
+  diagnosis and reproducibility.
+
+## Routing and promotion
+
+Route the request before selecting an executor. Simple, localized, and
+repository-wide paths are treatment regimes, not implicit executor cascades.
+Fallbacks, executor switches, scope expansion, acceptance, and promotion must
+be explicit and evidence-backed.
+
+Do not promote a provider, model, executor, or treatment until the declared
+authority has accepted the required evidence. A Gold or infrastructure
+qualification validates its stated authority and does not by itself prove
+mechanism effectiveness or general model quality.
+
+## Agent execution in local repositories
+
+The execution agent is a task worker, not an implicit release manager.
+
+- Start each task by recording the workspace, branch/worktree, base revision,
+  and pre-existing working-tree changes.
+- Treat one task, one workspace, and one attempt as the default unit. Retries
+  receive new attempt identity and preserve the previous artifacts.
+- Preserve unrelated local changes. Never use broad `git add -A`, reset, clean,
+  checkout, stash, or file deletion to make a workspace look clean unless the
+  user explicitly authorizes that exact operation.
+- The default deliverable is a verified patch plus evidence. Do not create a
+  commit, branch, tag, merge, or pull request unless the request explicitly
+  authorizes it.
+- When a commit is authorized, make at most the task-scoped commit after the
+  declared checks pass. Do not create checkpoint or speculative commits.
+- When a pull request is authorized, first locate the task's existing branch
+  or open PR and update it when appropriate. Do not create duplicate PRs for
+  retries, blocked attempts, or the same task. A blocked, failed, or
+  indeterminate attempt is not PR-ready.
+- Never infer completion from a clean diff, a created commit, or an open PR.
+  Completion requires the declared verification and acceptance evidence.
+
+The complete workspace/patch/commit/PR protocol is in
+[`docs/agent-execution-workflow.md`](docs/agent-execution-workflow.md).
+
+## Change discipline
+
+- Preserve unrelated working-tree changes.
+- Prefer the smallest change that satisfies the contract.
+- Add or update tests with behavioral changes.
+- Update the relevant architecture, protocol, or decision documentation when
+  the boundary changes.
+- Run the focused tests first, then the complete test suite when practical.
+- Do not claim merge readiness or a green local baseline while required
+  evidence or acceptance records are missing, or while the declared local
+  suite is not green. Historical passing runs must retain their revision,
+  command, environment, and limitations.
+
+## Documentation status
+
+- Normative rules live in this file and in the current experimental protocol.
+- Decision records explain why a boundary exists; they do not silently replace
+  the current protocol.
+- Logs and manifests are evidence, not policy. A historical log must not be
+  presented as the current state without its revision and collection date.
+- When a protocol boundary changes, update the protocol, the relevant
+  architecture page, and one decision record together. Avoid adding a new
+  document when an existing normative contract can be updated.
+
+## Current baseline
+
+The recorded external authority result is `resolved: true` for the official
+SWE-bench Docker Gold path on the frozen tasks under `logs/evaluation/`. This
+does not qualify a real provider/model or the A/B/C/D mechanism trial.

README.md

@@ -1,33 +1,75 @@
-# Arkx
+# Codepro
 
-Arkx is a greenfield project for empirical, incremental development of reliable capability.
+Codepro is a repository for incremental implementation and empirical
+evaluation of software capability.
+
+The current Python import namespace is `arkx` for compatibility; the product,
+distribution, CLI, and release identity are `codepro`.
 
 > **MINIMUM SUFFICIENT ARCHITECTURE**  
 > **FOR MAXIMUM RELIABLE CAPABILITY**
 
-This repository is the canonical root for the project. It intentionally contains only the foundation in this bootstrap: the project contract, experimental protocol, structural placeholders, and a minimal CI check.
+This repository is the canonical root for the project. It contains runtime
+contracts, an execution harness, the project contract, the experimental
+protocol, and structural checks.
 
-## Scope of this bootstrap
+## Current scope
 
 Included:
 
 - explicit project invariants;
-- a small, evidence-oriented experimental protocol;
-- stable top-level boundaries for future work;
-- a deterministic structural check and minimal CI.
+- an evidence-oriented experimental protocol;
+- stable top-level boundaries;
+- a deterministic structural check, executable contracts, and a Treatment A
+  baseline harness.
 
-Deferred:
+Still deferred or unqualified:
 
-- routing and planning;
 - multi-agent orchestration;
-- OpenHands, ReX, mini-SWE-agent, and SWE-agent integrations;
+- OpenHands, ReX, and SWE-agent integrations;
+- real provider/model adoption and the A/B/C/D mechanism trial;
 - product dependencies or automatic migration from `smag-rex`.
 
 ## Development posture
 
 Every capability must be introduced through a falsifiable hypothesis, an explicit implementation, an executed experiment, and recorded evidence. Local success is not scientific or upstream evidence. No fallback, executor switch, or scope expansion may be silent.
 
 See [the project contract](docs/project-contract.md) and [the experimental protocol](docs/experimental-protocol.md).
+Agent workspace, patch, commit, retry and pull-request behavior is governed by
+[the agent execution workflow](docs/agent-execution-workflow.md).
+
+## P0 baseline and telemetry
+
+P0 provides executor-agnostic execution records, explicit event metrics, nullable resource measurements, and fail-closed status resolution. It does not implement performance optimization, routing, planning, or any executor. See [the P0 metrics contract](docs/baseline-metrics.md).
+
+Install the development distribution in a clean environment and run the
+deterministic fixture and tests:
+
+```text
+python -m pip install .
+python -m arkx.baseline
+python -m unittest discover -s tests -t . -v
+```
+
+The distribution is currently `codepro==0.3.0.dev0`; it is a packaged
+pre-release chassis. The public commands are limited to health and the
+deterministic baseline:
+
+```text
+codepro --version
+codepro doctor --json
+codepro baseline
+```
+
+There is intentionally no task-submission or executor command yet.
+
+P1 adds deterministic task characterization from explicit signals only. See [the P1 contract](docs/task-characterization.md). It does not select executors or route work.
+
+P2 adds deterministic progress/stagnation assessment from explicit snapshots only. See [the P2 contract](docs/progress-stagnation.md). It does not control execution or authorize retry/replan.
+
+P3 adds bounded rule-based routing and evidence-based escalation decisions. See [the P3 contract](docs/routing-escalation.md). It does not invoke concrete executors.
+
+Block B adds independently removable P4 repository planning/state, P5 patch verification, and P6 handoff/context accounting. P8.2a adds a Treatment A baseline runner and official SWE-bench acceptance adapter; neither owns general orchestration or promotion.
 
 ## Checks
 
@@ -36,4 +78,3 @@ The bootstrap check is dependency-free:
 ```text
 python tools/check_foundation.py
 ```
-