Add GPT-6 Astra memory support and reproducible SWE-bench Pro infrastructure

.dockerignore

@@ -9,5 +9,9 @@ benchmark/*
 !benchmark/__init__.py
 !benchmark/codex_adapter/
 !benchmark/codex_adapter/**
+!benchmark/astra/
+benchmark/astra/*
+!benchmark/astra/__init__.py
+!benchmark/astra/pro_lane.py
 !benchmark/container/
 !benchmark/container/**

.superpowers/sdd/2026-09-07-astra-public-benchmark/task-pro-report.md

@@ -0,0 +1,49 @@
+# SWE-bench Pro Container Lane Report
+
+Implemented the minimal in-container Pro solver lane without running model inference or changing
+the benchmark protocol.
+
+## Delivered
+
+- `benchmark/astra/pro_lane.py`: validates the instance ID, exact lowercase base commit, condition,
+  fixed input paths, repository root, and non-existing output before mutation. It archives the base
+  tree, retains ignored installed dependencies, removes future/current tracked content and original
+  Git metadata, creates a deterministic detached single-commit snapshot, verifies tree equality and
+  history size, rejects base trees containing gitlinks before mutation, purges nested Git metadata,
+  forwards the unchanged public prompt to `run_task`, and records original/snapshot commit and tree
+  hashes.
+- `benchmark/astra/pro/Dockerfile.solver`: layers on a controller-supplied task image, uses pinned
+  amd64 Alpine Node and uv builder manifests, installs Codex 0.153.4 with the existing wrapper,
+  adds an independent musl Python 3.12.11 runtime, installs MARGINAL/benchmark packages into that
+  runtime's site-packages, records source/task/tool labels, and invokes the lane CLI from `/app`.
+- `.dockerignore`: admits only the Astra package files required by the solver image.
+- `tests/benchmark/test_astra_pro_lane.py`: focused repository-isolation, validation, overwrite, and
+  forwarding tests.
+
+## Verification
+
+```text
+$ .venv/bin/python -m pytest -q tests/benchmark/test_astra_pro_lane.py
+......                                                                   [100%]
+7 passed in 2.24s
+```
+
+```text
+$ .venv/bin/ruff check benchmark/astra/pro_lane.py tests/benchmark/test_astra_pro_lane.py
+All checks passed!
+```
+
+The image build uses task image
+`jefzda/sweap-images@sha256:d902632d1374cf0282a4ea301b82c296e13a41127308da0204aca87a4ba62c02`.
+Despite earlier evaluator metadata describing Ubuntu 20.04/Python 3.8, this exact image is Alpine
+Linux 3.18.3. The solver therefore uses musl-compatible Node and standalone Python binaries. Its
+final image ID and no-inference executable smoke checks are recorded below once complete.
+
+## Concerns
+
+- Repository sanitization is intentionally destructive and is safe only because the official task
+  container is disposable.
+- Ignored dependency retention depends on the base tree's ignore rules. The verified snapshot tree
+  contains only tracked base-commit files; ignored dependency content remains outside Git.
+- Base commits containing gitlinks are rejected as unsupported infrastructure before mutation.
+- No solver/model request was made during implementation or verification.

README.md

@@ -68,6 +68,27 @@ marginal install codex --autopilot-consent
 
 ## How MARGINAL earns authority
 
+### GPT-6 Astra
+
+The exact model ID `gpt-6-astra` has its own evidence namespace, `openai/gpt-6-astra`.
+When Codex hooks identify that model, decisions and outcomes are recorded locally as work
+happens. Existing unlabelled evidence is not reassigned to Astra. Commons sharing remains opt-in.
+
+A real Astra request was verified with Codex CLI **0.153.4**; **0.147.0** was rejected by the
+provider. This is a tested version, not a claim that every intermediate version is unsupported.
+
+**Astra token savings are not yet demonstrated.** Shadow Mode observes work; it does not
+reduce the model's internal reasoning budget. The complete public OFF/ON evaluation is still
+outstanding. See [Astra evaluation status](benchmark/astra/README.md) for exact preflight facts
+and what remains before a performance claim is justified.
+
+The [SWE-bench Pro protocol and evidence](benchmark/astra/pro/README.md) include all 731
+planned task IDs and the first passing official reference check. That reference check is
+not a model score. A complete OFF/ON run requires 1,462 model executions; no complete
+performance result is claimed until those executions and their grading are available.
+
+### Evidence gates
+
 1. **Observe** — collect derived action, outcome, coverage, state and evidence signals locally.
 2. **Verify** — bind decisions, policy identity, trust state and governance cost into Decision Receipts.
 3. **Earn** — require representative local evidence, clean coverage and explicit promotion.

benchmark/astra/README.md

@@ -0,0 +1,57 @@
+# GPT-6 Astra evaluation status — updated 2026-09-08
+
+**No complete Astra OFF/ON benchmark has been executed. No token-saving claim is supported.**
+
+This directory records preflight facts, not benchmark results. Product integration tests and a
+successful model request do not demonstrate improved task quality or lower token use.
+
+The active experiment is [SWE-bench Pro](pro/README.md). Its complete 731-task manifest,
+protocol and first official reference-validation artifacts are published there. HumanEval+
+preparation was withdrawn before any scored model inference.
+
+## Frozen public benchmark inputs
+
+- Dataset: [ScaleAI/SWE-bench_Pro](https://huggingface.co/datasets/ScaleAI/SWE-bench_Pro),
+  revision `7ab5114912baf22bb098818e604c02fe7ad2c11f`, complete `test` split: **731 tasks**.
+- Evaluator: [scaleapi/SWE-bench_Pro-os](https://github.com/scaleapi/SWE-bench_Pro-os),
+  commit `ca10a60a5fcae51e6948ffe1485d4153d421e6c5`.
+- A single complete OFF/ON comparison requires **1,462 independent task executions**, followed
+  by official grading. It is not equivalent to 1,462 unit tests or grading empty patches.
+- Intended model: `gpt-6-astra`, with identical reasoning effort, prompts, tools, limits and
+  independent state in both conditions. No task has been selected based on its model outcome.
+
+## Observed preflight
+
+| Check | Observation |
+| --- | --- |
+| Codex 0.147.0 + Astra | Provider rejected the request and required a newer Codex version |
+| Codex 0.153.4 + Astra | Minimal no-tool request completed successfully |
+| Successful request usage | 16,474 input; 11,520 cached input; 7 output; 0 reasoning output |
+| Docker | Existing local benchmark VM started; Docker 29.5.2, 6 CPUs, about 12 GiB RAM |
+| Existing task images | Three historical Lite smoke images retained; first Pro image downloaded |
+| First deterministic Pro image | Pinned in [Pro preflight](pro/preflight.json); actual OS is Alpine 3.18.3 |
+| Official Pro grading | First reference patch passed; 1,350.20 seconds under x86 QEMU |
+| Scored Astra task trajectories | **0 baseline / 0 MARGINAL** |
+
+The successful no-tool request is an API/CLI compatibility probe only. Its input token count
+includes the Codex scaffold and cannot be extrapolated into a reliable per-task cost estimate.
+Cached input and reasoning output are subsets, not additional tokens to sum twice.
+
+## Required before results can be published
+
+1. Adapt and validate the runner against current Codex and official Pro task environments.
+   The historical Lite adapter is not a verified Pro execution backend.
+2. Pin images and prompts; validate the evaluator on reference patches in isolated containers.
+3. Publish the executable protocol before scored inference. The dataset pins above alone are
+   **not** a complete preregistration.
+4. Run all 731 tasks in each condition without exposing reference patches or grading tests
+   to the solving agent. Preserve failures, timeouts and every task ID.
+5. Publish sanitized predictions, per-task usage, official grading, intervention traces,
+   environment hashes and analysis. Report correctness and governance overhead together.
+
+The sprint reserves 20% of the short Codex quota window for validation and publication. A
+quota stop must remain visibly incomplete; it cannot become a fabricated full benchmark,
+a smaller benchmark labelled Pro, or a claim that default Shadow Mode saves reasoning tokens.
+
+HumanEval+ was considered as a lower-cost alternative but has not been run or substituted
+for the requested real-repository comparison. No external leaderboard certification is claimed.

benchmark/astra/__init__.py

@@ -0,0 +1 @@
+"""Astra public-benchmark orchestration."""

benchmark/astra/humaneval-plus/Dockerfile.evaluator

@@ -0,0 +1,11 @@
+FROM ganler/evalplus@sha256:26b118098bef281fe8dfe999bf05f1d5b45374b4e6c00161ec0f30592aef4740
+
+COPY evaluator-requirements.txt /tmp/evaluator-requirements.txt
+RUN python -m pip install --no-deps --require-hashes -r /tmp/evaluator-requirements.txt \
+    && python -c 'import importlib.metadata; assert importlib.metadata.version("evalplus") == "0.3.1"'
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    XDG_CACHE_HOME=/tmp/evalplus-cache \
+    HUMANEVAL_OVERRIDE_PATH=/data/HumanEvalPlus.jsonl
+
+ENTRYPOINT ["evalplus.evaluate"]

benchmark/astra/humaneval-plus/README.md

@@ -0,0 +1,51 @@
+# Astra × MARGINAL — full HumanEval+ protocol
+
+**Withdrawn before scored inference.** The user requires SWE-bench Pro. No model samples
+were generated for this suite. The reference-only evaluator check passed 163/164 tasks;
+its remaining failure was not investigated after the scope correction. This directory is
+retained for traceability, not as a performance result or an active preregistration.
+
+This is a **function-level** OFF/ON experiment, not SWE-bench Pro. No result is claimed
+until all 164 tasks have both a real model run and official grading.
+
+## What is being tested
+
+The same Codex 0.153.4 agent and GPT-6 Astra (`low` reasoning) implement every HumanEval+
+v0.1.10 function twice, in separate clean workspaces. Baseline has no MARGINAL benchmark
+adapter; ON uses the existing balanced/diminishing-return experimental enforcement adapter.
+The installed plugin's default Shadow Mode is **not** this intervention.
+
+Each task starts with only its public function prompt in `solution.py`. Reference solutions
+and official base/extra test inputs are not provided to the solver. The solver may perform
+local checks, but may not use the network or files outside its workspace. Execution traces
+must be audited for violations before any results are admitted. No extra repetition is
+requested to make MARGINAL look effective.
+
+## Frozen conditions
+
+The machine-readable [protocol](protocol.json) records task/data hashes, order, model,
+reasoning, timeout, concurrency, analysis and stop rules. Runner source revision, evaluator
+image digest and exact prompt hash are recorded and published before scored inference.
+All original tasks and all extra tests are included; neither `mini` nor `noextreme` is used.
+
+Correctness is primary: official base **and** extra tests must pass. Failed, timed-out,
+missing-usage and infrastructure-invalid lanes remain visible. They are never silently
+replaced with another attempt. The complete experiment requires 328 recorded lanes.
+
+Token accounting uses provider-reported totals and includes all calls, including recovery
+work. Cached input is part of input, and reasoning output is part of output. They are not
+added twice. Subscription quota is not a dollar bill. Unavailable costs stay unavailable.
+
+## Reading the eventual result
+
+Report paired correctness, token use, tool calls, elapsed time, intervention counts and
+governance overhead together. Report overall results and the both-solve subset separately.
+A zero-regression sample is an observation, not a statistical proof of equivalence. One
+sample per task cannot establish production reliability; a public function benchmark may
+also be saturated or present in model training data.
+
+Negative, zero-effect and inconclusive results are valid. This experiment cannot justify
+claiming that installing the default plugin reduces Astra's internal reasoning tokens.
+
+Sources: [EvalPlus v0.3.1](https://github.com/evalplus/evalplus/tree/e5d0ed0bab96280b60b637ec7f15b5e4841b0cb2),
+[HumanEval+ v0.1.10](https://github.com/evalplus/humanevalplus_release/releases/tag/v0.1.10).

benchmark/astra/humaneval-plus/evaluator-requirements.txt

@@ -0,0 +1 @@
+evalplus==0.3.1 --hash=sha256:cd601debb67419113d10ac5c3317689d847f27de5d8cf3837975f3cab571b75d

benchmark/astra/humaneval-plus/protocol.json

@@ -0,0 +1,41 @@
+{
+  "schema_version": 1,
+  "experiment": "astra-humaneval-plus-20260907",
+  "status": "withdrawn_before_scored_inference_user_requires_swe_bench_pro",
+  "dataset": "HumanEval+",
+  "dataset_version": "v0.1.10",
+  "dataset_sha256": "42526ec0e7d5f3ee0b06d6ced98f8c8bae3d76519151bfb3d36f79010645bd7f",
+  "task_inputs_sha256": "8fbac146a932a0e1bdbccf8b0596061d43312dc20143960c48b66029d2dbf425",
+  "task_count": 164,
+  "conditions": ["baseline", "marginal"],
+  "repetitions": 1,
+  "model": "gpt-6-astra",
+  "model_snapshot": "public alias; immutable provider weight snapshot unavailable",
+  "codex_version": "0.153.4",
+  "evaluator_version": "evalplus==0.3.1",
+  "evaluator_image_id": "sha256:a90b5cd691de7334d203b64435aaf1ef10701acf90eeb4455703df3724b5e6aa",
+  "reasoning_effort": "low",
+  "timeout_seconds": 180,
+  "concurrency": 2,
+  "task_order": "ascending sha256(task_id UTF-8)",
+  "condition_order": "baseline first at even task positions, marginal first at odd positions",
+  "baseline": "Codex without MARGINAL benchmark hooks or daemon",
+  "marginal": "existing benchmark adapter: balanced policy, default diminishing-return detector, experimental enforce, unlimited hard budgets",
+  "native_shadow_equivalent": false,
+  "reference_tests_available_to_solver": false,
+  "solutions_per_task_condition": 1,
+  "quality_primary": "official HumanEval+ base AND extra tests pass@1",
+  "failures": "retain every task; no re-solving after inspecting grading; timeouts and invalid infrastructure reported separately",
+  "quality_gate": "no observed baseline-only regressions; descriptive criterion, not a non-inferiority proof",
+  "efficiency": "paired total input+output tokens, cached-input subset, reasoning-output subset, tool calls, wall time, governance overhead and interventions",
+  "cost_usd": "unavailable unless directly reported; no invented zero or API bill for subscription usage",
+  "analysis": "per-task paired results and 95% task-bootstrap interval for token savings; report both-solve separately; do not infer equivalence from non-significance",
+  "stop_rules": "stop new lanes on missing usage, hook failures, invalid isolation or short quota >=80%; retain partial artifacts and label incomplete",
+  "claim_limits": [
+    "Function-level coding tasks, not real-repository SWE-bench Pro",
+    "Experimental benchmark enforcement, not default installed Shadow Mode",
+    "One stochastic sample per task per condition",
+    "Public benchmark contamination and saturation are possible",
+    "Self-published evaluation, not external leaderboard certification"
+  ]
+}