Praxist Takeover Codex

Praxist Takeover: Codex-native mode

Run the complete Praxist takeover workflow while making Codex-native mode the explicit operator choice. The mode uses the operator's saved ChatGPT login. The user invokes this skill once; the current agent performs the checks and shell operations. Do not ask the user to type Praxist commands, provider names, model names, or API keys.

Composition Contract

Load the currently installed praxist-takeover SKILL.md and follow its complete current workflow. In particular:

• run complete onboarding first;
• use the complete current task-initialization or repair workflow when needed;
• preserve evaluator, protocol, lane-routing, resource, and launch-readiness checks;
• preserve the user's explicit protocol-intent permissions, including an intentionally reduced or incomplete authoritative mode; Codex-native mode changes runtime/authentication only and must not strengthen task semantics;
• preserve the task-initialization closing-policy regression and positive mature normal-close gate whenever the task defines close-grade evidence;
• validate with praxist resolve;
• launch detached, confirm status, and report the independent monitor command.
• preserve the common agent.reasoning_effort policy; use max unless the user explicitly selected another supported value.

Do not copy or reconstruct an abbreviated takeover from memory. This skill overrides only authentication/runtime selection and the unambiguous-path confirmation behavior described below. If a base takeover instruction conflicts with this file on those points, this file wins.

Treat invocation of praxist-takeover-codex as the user's explicit choice of:

agent runtime: agent_runtime:codex_sdk
agent system: codex_sdk
model provider: model_provider:openai_compatible
model: gpt-5.6-luna unless the user explicitly selects another model
mode: Codex-native mode
authentication: saved ChatGPT login
API-key use: disabled for the Praxist model connection

Carry that choice through task initialization, generated launch guidance, the actual start command, and the final summary. Do not replace it with the normal DeepSeek/Claude default merely because task initialization recommends that profile for users who have a DeepSeek key.

Codex-native mode automatically enables Praxist's lossless context efficiency. Do not add a task-local compressor, memory database, persistent Codex conversation, or relay. Do not ask the user to configure the optimization. After launch, state that finding-only wakeups are batched while lifecycle and resource events remain immediate, and include the diagnostic skill for measuring session count and cached versus uncached input.

Preconditions

The research project must already run on this machine with its required code, data or simulator, and task environment. Apply the base takeover refusal rules when those task prerequisites are missing.

Praxist must provide the Codex runtime extra. In the selected Praxist Python environment, verify that openai_codex and mcp import successfully and that the runtime distributions are openai-codex==0.147.0 and claude-agent-sdk==0.2.136. If they are missing or mismatched, use praxist-runtime-install to install the tested runtime dependencies only; do not enter provider-key configuration and do not ask the user for a key.

Praxist and the research project may intentionally use different Python environments. Run Praxist CLI/runtime checks with the Praxist interpreter. Run evaluator and harness tests with the task interpreter; when that interpreter does not install Praxist, expose the selected Praxist source/package root only for that test subprocess through PYTHONPATH. Do not install Praxist into the task environment or create another environment merely to make a harness test import it.

Codex-native Preflight

Use the same Python environment and Praxist installation that will launch the run. First run the route-aware host gate:

praxist doctor --codex-native --task-path /absolute/path/to/task --json

Add --model <selected-model> when the user explicitly chose a model other than the profile default.

It must report agent_runtime:codex_sdk, model_provider:openai_compatible, auth_mode: codex-native, ready SDK/MCP dependencies, a valid saved login, and codex_model_catalog: ok for the selected model. This single gate uses the SDK-bundled Codex binary, performs a bounded login probe, and reads the account model catalog without making a research request. Do not duplicate it with a second model-catalog probe. Never read auth.json directly or print tokens, account identifiers, or credential contents. A successful login probe means the SDK-bundled binary reported exactly Logged in using ChatGPT. Ignore PRAXIST_CODEX_BIN only for this probe and launch so a stale custom binary override cannot make preflight and execution disagree. Do not alter the operator's persistent setting.

The default model for this skill is gpt-5.6-luna. Preserve a model explicitly selected by the user and pass it to the doctor gate. Never infer a different model from stale task documentation, shell variables, a previous resolve snapshot, or memory. When login succeeds but the selected model is absent from the account catalog, stop before resolve/start and report the available model identifiers; do not launch a cohort to discover the mismatch.

When the login probe succeeds, continue without asking any authentication question. When it fails, stop before task launch and ask the operator to run praxist setup --profile codex-native --install-skills codex in a local interactive terminal. That explicit setup route opens and verifies the SDK-pinned Codex login flow. Do not offer an API key as an automatic fallback, do not switch providers, and do not invoke a different Codex executable.

Task-Initialization Overrides

Keep the complete current task-initialization workflow. For a task newly created through this Codex-native entry only:

• do not launch until the current task-init evaluator fan-out preflight has passed, including its task-defined one-unit canary and any explicitly declared external-evidence trust check;

• generated resolve/start instructions must use the selected Codex runtime, provider, and model rather than the standard DeepSeek/Claude recommendation;

• use dig_lite.contract.min_rejected_alternatives: 2 unless the user already chose a stricter value, while preserving candidate count, mechanism and intervention diversity, selected-contract, ablation, forbidden-change, metric-signature, fail-fast, validator, and retry requirements;

• do not create a task-local agent_runtime:codex_sdk shadow, relay, transport, or model-specific schema adapter. Praxist owns runtime compatibility.

Do not silently rewrite the scientific contract of an existing task merely because it uses a stricter rejection count. Repair it only when the current task-init workflow is already repairing that harness or a bounded preflight shows the stricter task-owned value is not stable for the selected model.

Low-Interaction Task Selection

Use the user-provided project/task path when present; otherwise start from the Codex invocation directory.

Invocation of this skill is sufficient authorization to use a task path when one of the following resolves unambiguously:

• the current directory is itself one valid task project; or
• complete task initialization just created one task project for the current research project; or
• a bounded local scan finds exactly one launch-ready task belonging to the current research project.

Print the absolute path before launch, but do not ask for a redundant confirmation in those unambiguous cases. Ask one concise path-selection question only when multiple plausible task projects remain, the candidate is outside the current research project, or ownership cannot be established. Never guess among multiple active research projects.

All scientific and safety confirmation gates from task initialization remain in force. This low-interaction rule removes only redundant task-path confirmation; it does not authorize inventing missing objectives, data, simulators, evaluators, or baseline evidence.

Codex-native Resolve

Run every resolve with the same sanitized environment and explicit runtime, provider, and model that will be used for start:

env -u OPENAI_API_KEY \
    -u CODEX_API_KEY \
    -u CODEX_ACCESS_TOKEN \
    -u OPENAI_BASE_URL \
    -u PRAXIST_CODEX_BIN \
    -u MODEL \
    -u PRAXIST_MODEL \
  praxist resolve /absolute/path/to/task \
    --codex-native \
    --runtime agent_runtime:codex_sdk \
    --model-provider model_provider:openai_compatible \
    --model gpt-5.6-luna

Use the user-selected model in place of Luna when applicable. A successful resolve creates a run-like artifact snapshot whose exit condition is resolve_only; report it as configuration validation, not as a research run or failed launch. Track actual starts only from praxist start output.

The inherited task-initialization workflow must also run its real evaluator summary contract probe. Add --result-summary <actual_summary_path> to the same sanitized resolve command. When the task requires maturity ratios, do not start if this check cannot extract finite effort_ratio and coverage_ratio; repair the task evaluator or follow the explicit user-approved legacy fallback from task initialization. A stage label is not a substitute for ratio telemetry. The inherited workflow must also preserve its optional effective-configuration provenance check for configuration-sensitive treatments; Codex-native mode does not weaken that check or turn it into a gate for ordinary experiments.

Codex-native Launch

Do not persist provider settings or alter the user's shell configuration. Build one launch subprocess whose inherited environment omits these variables:

OPENAI_API_KEY
CODEX_API_KEY
CODEX_ACCESS_TOKEN
OPENAI_BASE_URL
PRAXIST_CODEX_BIN

Removing them only from the launch subprocess is required even when onboarding reports that they are present. This prevents API-key billing or a custom endpoint from taking precedence while preserving the user's environment for unrelated work. Do not unset unrelated task credentials.

Start through the selected Praxist executable with explicit matching runtime and provider flags:

env -u OPENAI_API_KEY \
    -u CODEX_API_KEY \
    -u CODEX_ACCESS_TOKEN \
    -u OPENAI_BASE_URL \
    -u PRAXIST_CODEX_BIN \
    -u MODEL \
    -u PRAXIST_MODEL \
  praxist start \
    --codex-native \
    --task-path /absolute/path/to/task \
    --agent-system codex_sdk \
    --runtime agent_runtime:codex_sdk \
    --model-provider model_provider:openai_compatible \
    --model gpt-5.6-luna \
    --daemonize \
    --json

The command is a behavioral template, not a request for user input. The current agent must execute it. On platforms without env -u, construct an equivalent child environment with only the listed entries removed.

Pass the selected model explicitly. The default is gpt-5.6-luna; preserve an explicit user selection after the account-catalog probe. Remove MODEL and PRAXIST_MODEL from the launch subprocess so stale defaults cannot override the verified choice. Preserve explicit cohort, generation, strategy, and run-directory choices when they do not conflict with Codex-native authentication.

Do not use praxist configure-llm, a relay provider, OPENAI_API_KEY, DeepSeek, OpenRouter, or Anthropic for this launch. Saved ChatGPT authentication is valid only with native model_provider:openai_compatible through agent_runtime:codex_sdk.

Post-Launch Verification

Treat a successful launcher response as startup-confirmed. Parse the JSON response, then use praxist status --run-id <run_id> --json and bounded log inspection to confirm all of the following:

• the selected task path matches the intended task;
• runtime is agent_runtime:codex_sdk;
• provider is model_provider:openai_compatible;
• model is the account-catalog-verified selection;
• the daemon remains alive and did not report saved-login or account-type failure;
• no fallback provider or relay was selected.

Praxist independently verifies that the app-server account type is chatgpt. If startup fails that check, report the failure and stop. Never retry with an API key or another provider.

Final Handoff

End with a compact table containing:

Do not open the foreground monitor automatically. Clearly state that the run is detached and continues after the current operator-agent session closes. Use praxist status --json and filter its returned records by run_id; do not scan unrelated rows when the direct selector is available. Prefer praxist status --run-id <run_id> --json. The foreground TUI selector remains praxist --monitor --run-id <run_id>.