traceverdict 0.2.0__tar.gz

traceverdict-0.2.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 weaqaq
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

traceverdict-0.2.0/PKG-INFO

@@ -0,0 +1,123 @@
+Metadata-Version: 2.4
+Name: traceverdict
+Version: 0.2.0
+Summary: Auditable state-regression evaluation for coding agents; published by the project owner
+License-Expression: MIT
+Project-URL: Homepage, https://github.com/weaqaq/traceverdict
+Project-URL: Repository, https://github.com/weaqaq/traceverdict.git
+Project-URL: Issues, https://github.com/weaqaq/traceverdict/issues
+Requires-Python: >=3.12
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: typer
+Requires-Dist: pydantic
+Requires-Dist: docker
+Requires-Dist: litellm
+Requires-Dist: rich
+Requires-Dist: pytest
+Requires-Dist: GitPython
+Dynamic: license-file
+
+# TraceVerdict
+
+TraceVerdict is an auditable regression-evaluation harness for coding agents. It freezes repository and container state, captures native agent trajectories, and compares the resulting file changes and official test verdicts with paired statistics.
+
+TraceVerdict is not a general-purpose call recorder or agent-rewind system: its unit of evidence is a reproducible coding task with a frozen workspace, environment fingerprint, authoritative patch, and test outcome. It sits below general observability products such as LangSmith and Langfuse as a vertical state-regression layer for coding agents; traces can be exported to those systems. It runs on top of the official SWE-bench harness, whose verdict remains ground truth—TraceVerdict records, reconciles, and compares that verdict rather than replacing the harness.
+
+> This repository is a clean public mirror. Full per-commit audit history and raw run data remain in a private audit repository and can be shown in an interview. Public documents preserve the decision and finding narrative while removing machine paths, credentials, personal information, and infrastructure details.
+
+## What it provides
+
+- adapters for mini-swe-agent 2.4.5 and pinned SWE-agent 1.1.0;
+- immutable config identities and strict provider usage/cost reconciliation;
+- disposable Docker workspaces with patch, artifact, and image-digest evidence;
+- SWE-bench 4.1.0 official-harness verdict integration;
+- paired 10,000-resample bootstrap, exact McNemar, regression alarms, and failure taxonomy;
+- deterministic comparison/report generation with explicit shared-sample disclosure;
+- faithful fault injection that isolates the agent environment from the verifier.
+
+## Evidence, with limits
+
+M3 froze 16 SWE-bench Verified tasks and added 85 paid evaluations. The cross-provider experiment used 16 tasks at `k=2`; its pass delta was `-0.09375` with bootstrap 95% CI `[-0.21875, 0.03125]`. This is engineering evidence for the regression pipeline, not a broad model leaderboard claim. A read-only-workspace regression produced a hard alarm (delta `-0.3125`, CI `[-0.5, -0.125]`) with 32/32 agent-read-only/verifier-read-write isolation records. Total audited API spend was `$1.8627937123` against a `$28` tripwire.
+
+See the [M3 evidence summary](docs/evidence/m3/m3_summary.md) and its [machine-readable companion](docs/evidence/m3/m3_summary.json).
+
+The later [M4-C compatibility arm](docs/evidence/m4c/m4c_codex_arm_20260718.md) ran a pinned Codex CLI locally while keeping subscription credentials off the verifier host. It completed 16/16 official judgments with 6/16 resolved; versus the DeepSeek Flash `k=2` baseline, the authorized asymmetric comparison yielded delta pass `+0.0625`, 95% CI `[-0.15625, 0.3125]`, exact McNemar `p=0.25`, and no regression alarm. This is cross-scaffolding compatibility evidence, not a model leaderboard claim.
+
+## Install
+
+Python 3.12 and Docker are required for real task execution.
+
+```bash
+python -m venv .venv
+python -m pip install -e .
+python -m pip install "mini-swe-agent==2.4.5"
+traceverdict --help
+```
+
+TraceVerdict v0.2 exposes exactly ten top-level commands through either `traceverdict` or the equivalent short entry point `tv`:
+
+```text
+traceverdict run TASK --config CONFIG
+traceverdict suite tasks/self --config configs/dev.yaml --dry-run
+traceverdict compare --baseline BASE --candidate CANDIDATE --task-set tasks/self/task_set.txt
+traceverdict report COMPARISON_ID
+traceverdict inject I3Q --base configs/dev.yaml --output injected.yaml
+traceverdict replay
+traceverdict selftest --config configs/dev.yaml
+tv quick --set model=openai/deepseek-v4-flash
+tv baseline set --config configs/dev.yaml
+tv ingest
+```
+
+`replay` remains an intentionally visible zero-model CI boundary and currently exits with code 2. Paid/provider runs require credentials supplied only through the process environment; no credential file belongs in this repository.
+
+## Daily Mode
+
+Daily Mode keeps its local, ignored state under `.traceverdict/daily/`. First establish an explicit cached baseline (this is the only step that may execute its missing baseline tasks):
+
+```bash
+tv baseline set --config configs/dev.yaml
+tv quick --set model=openai/deepseek-v4-flash
+tv quick --set model=openai/deepseek-v4-flash --set model_params.thinking.type=enabled --full
+```
+
+The default smoke set is frozen to S1/S4/S6; `--full` is frozen to S1-S8 and reuses already completed smoke runs. SWE-bench is deliberately unavailable through `quick`: the 16-task public benchmark is for release evidence, not daily iteration. Derived configs are content-addressed, exact-price-registry checked, and immutable. `baseline update` promotes an already complete candidate without running a model, and refuses correctness or forbidden-path regressions unless `--accept-regression` is explicit.
+
+The one-screen result reports pass delta, token-median delta/ratio, wall-P95 delta/ratio, strict actual cost, cache reuse, and failed tasks. PASS/WARN exit 0, a behavioral FAIL exits 1, and a missing baseline or invalid identity exits 2. The five-minute/$0.005 and full-suite timing figures are experience estimates only; each invocation reports measured time and cost.
+
+`tv ingest [PATH ...]` is passive and starts no model, Docker container, or verifier. It incrementally summarizes stable `codex exec --json` logs and the explicitly versioned July 2026 desktop-rollout compatibility format. Only dates, model, token classes, turns, tool counts, and failure classes are persisted; prompts, answers, commands, output, paths, and credentials are not. Desktop transcript format is not a stable public interface, so required-field drift fails closed instead of being guessed.
+
+This package is published by the project owner. Trusted Publishing through GitHub Actions OIDC is planned as a v0.3 supply-chain improvement.
+
+## Reproduce the frozen report
+
+```bash
+python scripts/finalize_m3.py \
+  --db reports/m3/traceverdict.db \
+  --completion-root reports/m3 \
+  --output reports/m3/final
+```
+
+The command fails closed on an incomplete 16-task/`k=2` matrix, missing metrics or verdicts, environment-fingerprint drift, missing I3Q isolation evidence, or an incomplete SWE-agent trace. Raw databases and trajectories are intentionally absent from the public mirror.
+
+## Findings index
+
+- [F-1](docs/findings/F1.md): deleting tool instructions did not degrade the tested agent at `n=8, k=1`.
+- [F-2](docs/findings/F2.md): truncating observations to 500 characters invited adaptive recovery.
+- [F-3](docs/findings/F3.md): a task-template-loss alarm was unstable when it depended only on cost at `k=1`.
+- [F-4](docs/findings/F4.md): CRLF conversion caused a cross-platform forbidden-path false positive.
+- [F-5](docs/findings/F5.md): strict reconciliation caught a provider reasoning-token pricing partition bug.
+- [F-6](docs/findings/F6.md): a 96-run, four-arm audit separated shared failures into sampling difficulty, runtime constraints, and submitted task-level convergence.
+
+## Build/reuse boundary
+
+Self-built components are the tracer, agent adapters, verifier wiring, snapshot discipline, comparison statistics, taxonomy, injection isolation, and evidence/replay boundaries. Reused components are mini-swe-agent, SWE-agent, the SWE-bench official harness and images, LiteLLM, Docker, and provider APIs. See the [public PRD](docs/PRD.md), [decision log](docs/decisions.md), and [publication provenance](docs/publication_provenance.md).
+
+## Security and provenance
+
+CI runs a fail-closed public-safety scan over text, binary bundles, nested Git authors/content, and repository history. The initial public history is one genuine release commit; it does not imitate the private development history. See [name audit](docs/name_audit.md).
+
+## License
+
+[MIT](LICENSE)

traceverdict-0.2.0/README.md

@@ -0,0 +1,103 @@
+# TraceVerdict
+
+TraceVerdict is an auditable regression-evaluation harness for coding agents. It freezes repository and container state, captures native agent trajectories, and compares the resulting file changes and official test verdicts with paired statistics.
+
+TraceVerdict is not a general-purpose call recorder or agent-rewind system: its unit of evidence is a reproducible coding task with a frozen workspace, environment fingerprint, authoritative patch, and test outcome. It sits below general observability products such as LangSmith and Langfuse as a vertical state-regression layer for coding agents; traces can be exported to those systems. It runs on top of the official SWE-bench harness, whose verdict remains ground truth—TraceVerdict records, reconciles, and compares that verdict rather than replacing the harness.
+
+> This repository is a clean public mirror. Full per-commit audit history and raw run data remain in a private audit repository and can be shown in an interview. Public documents preserve the decision and finding narrative while removing machine paths, credentials, personal information, and infrastructure details.
+
+## What it provides
+
+- adapters for mini-swe-agent 2.4.5 and pinned SWE-agent 1.1.0;
+- immutable config identities and strict provider usage/cost reconciliation;
+- disposable Docker workspaces with patch, artifact, and image-digest evidence;
+- SWE-bench 4.1.0 official-harness verdict integration;
+- paired 10,000-resample bootstrap, exact McNemar, regression alarms, and failure taxonomy;
+- deterministic comparison/report generation with explicit shared-sample disclosure;
+- faithful fault injection that isolates the agent environment from the verifier.
+
+## Evidence, with limits
+
+M3 froze 16 SWE-bench Verified tasks and added 85 paid evaluations. The cross-provider experiment used 16 tasks at `k=2`; its pass delta was `-0.09375` with bootstrap 95% CI `[-0.21875, 0.03125]`. This is engineering evidence for the regression pipeline, not a broad model leaderboard claim. A read-only-workspace regression produced a hard alarm (delta `-0.3125`, CI `[-0.5, -0.125]`) with 32/32 agent-read-only/verifier-read-write isolation records. Total audited API spend was `$1.8627937123` against a `$28` tripwire.
+
+See the [M3 evidence summary](docs/evidence/m3/m3_summary.md) and its [machine-readable companion](docs/evidence/m3/m3_summary.json).
+
+The later [M4-C compatibility arm](docs/evidence/m4c/m4c_codex_arm_20260718.md) ran a pinned Codex CLI locally while keeping subscription credentials off the verifier host. It completed 16/16 official judgments with 6/16 resolved; versus the DeepSeek Flash `k=2` baseline, the authorized asymmetric comparison yielded delta pass `+0.0625`, 95% CI `[-0.15625, 0.3125]`, exact McNemar `p=0.25`, and no regression alarm. This is cross-scaffolding compatibility evidence, not a model leaderboard claim.
+
+## Install
+
+Python 3.12 and Docker are required for real task execution.
+
+```bash
+python -m venv .venv
+python -m pip install -e .
+python -m pip install "mini-swe-agent==2.4.5"
+traceverdict --help
+```
+
+TraceVerdict v0.2 exposes exactly ten top-level commands through either `traceverdict` or the equivalent short entry point `tv`:
+
+```text
+traceverdict run TASK --config CONFIG
+traceverdict suite tasks/self --config configs/dev.yaml --dry-run
+traceverdict compare --baseline BASE --candidate CANDIDATE --task-set tasks/self/task_set.txt
+traceverdict report COMPARISON_ID
+traceverdict inject I3Q --base configs/dev.yaml --output injected.yaml
+traceverdict replay
+traceverdict selftest --config configs/dev.yaml
+tv quick --set model=openai/deepseek-v4-flash
+tv baseline set --config configs/dev.yaml
+tv ingest
+```
+
+`replay` remains an intentionally visible zero-model CI boundary and currently exits with code 2. Paid/provider runs require credentials supplied only through the process environment; no credential file belongs in this repository.
+
+## Daily Mode
+
+Daily Mode keeps its local, ignored state under `.traceverdict/daily/`. First establish an explicit cached baseline (this is the only step that may execute its missing baseline tasks):
+
+```bash
+tv baseline set --config configs/dev.yaml
+tv quick --set model=openai/deepseek-v4-flash
+tv quick --set model=openai/deepseek-v4-flash --set model_params.thinking.type=enabled --full
+```
+
+The default smoke set is frozen to S1/S4/S6; `--full` is frozen to S1-S8 and reuses already completed smoke runs. SWE-bench is deliberately unavailable through `quick`: the 16-task public benchmark is for release evidence, not daily iteration. Derived configs are content-addressed, exact-price-registry checked, and immutable. `baseline update` promotes an already complete candidate without running a model, and refuses correctness or forbidden-path regressions unless `--accept-regression` is explicit.
+
+The one-screen result reports pass delta, token-median delta/ratio, wall-P95 delta/ratio, strict actual cost, cache reuse, and failed tasks. PASS/WARN exit 0, a behavioral FAIL exits 1, and a missing baseline or invalid identity exits 2. The five-minute/$0.005 and full-suite timing figures are experience estimates only; each invocation reports measured time and cost.
+
+`tv ingest [PATH ...]` is passive and starts no model, Docker container, or verifier. It incrementally summarizes stable `codex exec --json` logs and the explicitly versioned July 2026 desktop-rollout compatibility format. Only dates, model, token classes, turns, tool counts, and failure classes are persisted; prompts, answers, commands, output, paths, and credentials are not. Desktop transcript format is not a stable public interface, so required-field drift fails closed instead of being guessed.
+
+This package is published by the project owner. Trusted Publishing through GitHub Actions OIDC is planned as a v0.3 supply-chain improvement.
+
+## Reproduce the frozen report
+
+```bash
+python scripts/finalize_m3.py \
+  --db reports/m3/traceverdict.db \
+  --completion-root reports/m3 \
+  --output reports/m3/final
+```
+
+The command fails closed on an incomplete 16-task/`k=2` matrix, missing metrics or verdicts, environment-fingerprint drift, missing I3Q isolation evidence, or an incomplete SWE-agent trace. Raw databases and trajectories are intentionally absent from the public mirror.
+
+## Findings index
+
+- [F-1](docs/findings/F1.md): deleting tool instructions did not degrade the tested agent at `n=8, k=1`.
+- [F-2](docs/findings/F2.md): truncating observations to 500 characters invited adaptive recovery.
+- [F-3](docs/findings/F3.md): a task-template-loss alarm was unstable when it depended only on cost at `k=1`.
+- [F-4](docs/findings/F4.md): CRLF conversion caused a cross-platform forbidden-path false positive.
+- [F-5](docs/findings/F5.md): strict reconciliation caught a provider reasoning-token pricing partition bug.
+- [F-6](docs/findings/F6.md): a 96-run, four-arm audit separated shared failures into sampling difficulty, runtime constraints, and submitted task-level convergence.
+
+## Build/reuse boundary
+
+Self-built components are the tracer, agent adapters, verifier wiring, snapshot discipline, comparison statistics, taxonomy, injection isolation, and evidence/replay boundaries. Reused components are mini-swe-agent, SWE-agent, the SWE-bench official harness and images, LiteLLM, Docker, and provider APIs. See the [public PRD](docs/PRD.md), [decision log](docs/decisions.md), and [publication provenance](docs/publication_provenance.md).
+
+## Security and provenance
+
+CI runs a fail-closed public-safety scan over text, binary bundles, nested Git authors/content, and repository history. The initial public history is one genuine release commit; it does not imitate the private development history. See [name audit](docs/name_audit.md).
+
+## License
+
+[MIT](LICENSE)

traceverdict-0.2.0/pyproject.toml

@@ -0,0 +1,46 @@
+[build-system]
+requires = ["setuptools>=68", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "traceverdict"
+version = "0.2.0"
+description = "Auditable state-regression evaluation for coding agents; published by the project owner"
+readme = "README.md"
+requires-python = ">=3.12"
+license = "MIT"
+dependencies = [
+    "typer",
+    "pydantic",
+    "docker",
+    "litellm",
+    "rich",
+    "pytest",
+    "GitPython",
+]
+
+[project.scripts]
+traceverdict = "traceverdict.cli:app"
+tv = "traceverdict.cli:app"
+
+[project.urls]
+Homepage = "https://github.com/weaqaq/traceverdict"
+Repository = "https://github.com/weaqaq/traceverdict.git"
+Issues = "https://github.com/weaqaq/traceverdict/issues"
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.setuptools.package-data]
+traceverdict = [
+    "tracer/schema.sql",
+    "resources/daily/configs/*",
+    "resources/daily/tasks/self/*",
+    "resources/daily/tasks/self/_image/*",
+    "resources/daily/tasks/self/S*/*",
+    "resources/daily/tasks/self/S*/verify/*",
+]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+pythonpath = ["src"]

traceverdict-0.2.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

traceverdict-0.2.0/src/traceverdict/__init__.py

@@ -0,0 +1,3 @@
+"""TraceVerdict: stateful regression evaluation system for coding agents."""
+
+__version__ = "0.2.0"

traceverdict-0.2.0/src/traceverdict/adapters/__init__.py

@@ -0,0 +1 @@
+"""adapters: integration layer for agents under test (mini_swe_agent first)."""

traceverdict-0.2.0/src/traceverdict/adapters/codex.py

@@ -0,0 +1,235 @@
+"""Local-only Codex exec adapter for the M4-C compatibility arm (D24)."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+import subprocess
+from pathlib import Path
+from typing import Any
+
+from traceverdict.adapters.mini_swe_agent import AdapterHarnessError, AdapterResult
+from traceverdict.swebench_budget import ADAPTER_WALL_TIMEOUT_GRACE_S
+
+PINNED_CODEX_VERSION = "0.144.4"
+CODEX_TRAJECTORY_FORMAT = "codex-exec-jsonl-1"
+CODEX_BINARY_IN_IMAGE = "/opt/traceverdict/codex"
+
+
+def _config_args(model_params: dict[str, Any]) -> list[str]:
+    """Return the complete behavior-affecting Codex config overrides."""
+    effort = model_params.get("model_reasoning_effort")
+    if effort not in {"high", "medium"}:
+        raise AdapterHarnessError(
+            "model_reasoning_effort must be an explicitly approved M4-C value "
+            f"(high or medium), got {effort!r}"
+        )
+    expected = {
+        "auth_mode": "chatgpt_subscription",
+        "billing_mode": "subscription_unallocatable",
+        "approval_policy": "never",
+        "sandbox_mode": "workspace-write",
+        "web_search": "disabled",
+        "ephemeral": True,
+        "ignore_user_config": True,
+        "ignore_rules": True,
+        "concurrency": 1,
+        "network_mode": "bridge",
+        "container_cap_add": "SYS_ADMIN",
+        "container_seccomp": "unconfined",
+        "inner_sandbox": "workspace-write",
+        "auth_isolation": "per_tool_mount_namespace_hide_codex_home_v2",
+    }
+    mismatches = {
+        key: {"expected": value, "actual": model_params.get(key)}
+        for key, value in expected.items()
+        if model_params.get(key) != value
+    }
+    if mismatches:
+        raise AdapterHarnessError(
+            "immutable Codex M4-C identity mismatch: "
+            + json.dumps(mismatches, sort_keys=True)
+        )
+    nested_expected = {
+        "sandbox_workspace_write": {"network_access": True},
+        "history": {"persistence": "none"},
+        "features": {
+            "shell_tool": True,
+            "shell_snapshot": False,
+            "unified_exec": False,
+        },
+        "hide_agent_reasoning": False,
+    }
+    nested_mismatches = {
+        key: {"expected": value, "actual": model_params.get(key)}
+        for key, value in nested_expected.items()
+        if model_params.get(key) != value
+    }
+    if nested_mismatches:
+        raise AdapterHarnessError(
+            "immutable Codex nested identity mismatch: "
+            + json.dumps(nested_mismatches, sort_keys=True)
+        )
+    if model_params.get("service_tier") != "omitted":
+        raise AdapterHarnessError("service_tier must be explicitly omitted")
+    if model_params.get("temperature") != "omitted" or model_params.get("top_p") != "omitted":
+        raise AdapterHarnessError("temperature and top_p must be explicitly omitted")
+    return [
+        "-c", f"model_reasoning_effort=\"{effort}\"",
+        "-c", "approval_policy=\"never\"",
+        "-c", "web_search=\"disabled\"",
+        "-c", "history.persistence=\"none\"",
+        "-c", "hide_agent_reasoning=false",
+        "-c", "features.shell_tool=true",
+        "-c", "features.shell_snapshot=false",
+        "-c", "features.unified_exec=false",
+        "-c", "sandbox_workspace_write.network_access=true",
+    ]
+
+
+def build_codex_command(
+    *, model_name: str, model_params: dict[str, Any], container_cwd: str
+) -> list[str]:
+    return [
+        CODEX_BINARY_IN_IMAGE,
+        "exec",
+        "--json",
+        "--ephemeral",
+        "--ignore-user-config",
+        "--ignore-rules",
+        "--strict-config",
+        "--model",
+        model_name,
+        "--sandbox",
+        "workspace-write",
+        "-C",
+        container_cwd,
+        *_config_args(model_params),
+        "-",
+    ]
+
+
+def parse_codex_jsonl(data: str) -> list[dict[str, Any]]:
+    records: list[dict[str, Any]] = []
+    for line_number, raw in enumerate(data.splitlines(), start=1):
+        if not raw.strip():
+            continue
+        try:
+            item = json.loads(raw)
+        except json.JSONDecodeError as exc:
+            raise AdapterHarnessError(
+                f"Codex JSONL line {line_number} is invalid: {exc}"
+            ) from exc
+        if not isinstance(item, dict) or not isinstance(item.get("type"), str):
+            raise AdapterHarnessError(
+                f"Codex JSONL line {line_number} lacks an event type"
+            )
+        records.append(item)
+    if not records:
+        raise AdapterHarnessError("Codex produced no JSONL records")
+    return records
+
+
+def run_codex(
+    *,
+    instruction: str,
+    image: str,
+    docker_executable: str,
+    host_work_path: Path,
+    container_cwd: str,
+    model_name: str,
+    model_params: dict[str, Any],
+    litellm_model_registry: Path | None,
+    agent_version: str,
+    cost_limit: float,
+    step_limit: int,
+    wall_time_s: int,
+    work_dir: Path,
+) -> AdapterResult:
+    """Run Codex inside the local agent-layer image; never use a remote host."""
+    del litellm_model_registry, cost_limit, step_limit
+    if agent_version != PINNED_CODEX_VERSION:
+        raise AdapterHarnessError(
+            f"Codex version mismatch: expected {PINNED_CODEX_VERSION}, got {agent_version}"
+        )
+    auth_spec = os.environ.get("TRACEVERDICT_CODEX_AUTH_FILE")
+    if not auth_spec:
+        raise AdapterHarnessError("TRACEVERDICT_CODEX_AUTH_FILE is required locally")
+    auth_file = Path(auth_spec).resolve()
+    if not auth_file.is_file():
+        raise AdapterHarnessError("local Codex auth file does not exist")
+    work_dir.mkdir(parents=True, exist_ok=True)
+    jsonl_path = work_dir / "codex.events.jsonl"
+    stderr_path = work_dir / "codex.stderr.log"
+    command = build_codex_command(
+        model_name=model_name,
+        model_params=model_params,
+        container_cwd=container_cwd,
+    )
+    host_work = str(host_work_path.resolve()).replace("\\", "/")
+    host_auth = str(auth_file).replace("\\", "/")
+    docker_command = [
+        docker_executable,
+        "run",
+        "--rm",
+        "-i",
+        "--network",
+        "bridge",
+        "--cap-add",
+        "SYS_ADMIN",
+        "--security-opt",
+        "seccomp=unconfined",
+        "-e",
+        "CODEX_HOME=/run/traceverdict-codex",
+        "-v",
+        f"{host_work}:{container_cwd}",
+        "-v",
+        f"{host_auth}:/run/traceverdict-codex/auth.json",
+        "-w",
+        container_cwd,
+        image,
+        *command,
+    ]
+    timeout = max(1, int(wall_time_s)) + ADAPTER_WALL_TIMEOUT_GRACE_S
+    try:
+        proc = subprocess.run(
+            docker_command,
+            input=instruction,
+            capture_output=True,
+            text=True,
+            encoding="utf-8",
+            errors="strict",
+            timeout=timeout,
+            check=False,
+        )
+    except subprocess.TimeoutExpired as exc:
+        raise AdapterHarnessError(
+            f"Codex exceeded adapter timeout {timeout}s", exit_reason="TimeExceeded"
+        ) from exc
+    jsonl_path.write_text(proc.stdout or "", encoding="utf-8")
+    stderr_path.write_text(proc.stderr or "", encoding="utf-8")
+    records = parse_codex_jsonl(proc.stdout or "")
+    trajectory = {
+        "trajectory_format": CODEX_TRAJECTORY_FORMAT,
+        "records": records,
+        "jsonl_sha256": hashlib.sha256((proc.stdout or "").encode("utf-8")).hexdigest(),
+        "instruction_sha256": hashlib.sha256(instruction.encode("utf-8")).hexdigest(),
+        "info": {
+            "submission": "",
+            "model_stats": {"instance_cost": None},
+            "returncode": proc.returncode,
+        },
+    }
+    trajectory_path = work_dir / "codex.trajectory.json"
+    trajectory_path.write_text(
+        json.dumps(trajectory, ensure_ascii=False, indent=2), encoding="utf-8"
+    )
+    return AdapterResult(
+        traj=trajectory,
+        traj_path=trajectory_path,
+        returncode=proc.returncode,
+        stdout=proc.stdout or "",
+        stderr=proc.stderr or "",
+        source_artifacts={"codex_jsonl": jsonl_path},
+    )