Skip to content

Yuimi-chaya/codex-development-guidelines / skills/adopt-agent-policy

adopt-agent-policy

Verified

Configure Codex user-level AGENTS.md when the user asks to install, adopt, or update these guidelines. Inspect the actual environment, relevant Codex configuration, and exposed models first; research public model facts, recommend a configuration, collect unresolved preferences together, and apply the authorized changes. Honor an explicit project-level target. Reviewing the repository or ordinary coding does not request installation.

references/model-selection.md

View on GitHub
# Model Research and User-Selected Roles

On an installation request, inspect the exposed inventory and check relevant public evidence before the consolidated recommendation. Also use for requested model/configuration advice or an intentional refresh, not a market survey in every coding thread. A consultation can remain read-only; the bundled policy assets suffice if a policy draft is also requested. Recommendation, selection, policy writing, benchmarking, and installation have distinct scopes.

## Start with the exposed inventory

Read the actual provider/harness inventory through permitted tools and relevant non-sensitive settings as described in [discovery.md](discovery.md). Record exact selectable IDs, provider, reasoning controls, per-agent overrides, visual input, tool access, and session-sharing constraints. A configured or cached name is not proof of current selectability. Do not scan credentials to discover hidden models. Public availability does not prove account or role access.

If enumeration is unavailable, use explicitly supplied facts with provenance and label unverified controls. Do not invent a model, select a similar-looking name, or claim that an override occurred. Disclose any permitted harness-default fallback.

## Verify public information

Search the public web for the exposed model identities using official model/release documentation and the relevant provider's pricing and capability documentation. Fetch the supporting pages, not only snippets. Resolve provider aliases only with evidence; a label resembling a public model is not proof of that backend. Do not send keys, private endpoints, nonpublic deployment IDs, or sensitive configuration to search engines. For a private alias, use a verified public mapping or leave identity unknown.

Record source and check date for material claims. Distinguish announcement, availability, snapshot, and alias dates. Do not infer a release date from a version number or claim the newest exposed name is the newest public model. When sources disagree, report the disagreement; when browsing is unavailable or declined, provide a clearly limited comparison without fabricated current prices or rankings.

Separate current-provider charges from upstream list prices. Include currency, units, input/cached-input/output, applicable image/tool/reasoning charges, and batch or other conditions where documented. An upstream price is background evidence, not the user's verified bill. Do not infer an effective price from a subscription label or assume a markup.

Speed evidence should distinguish time to first token, output throughput, tool/network waits, and end-to-end task time. Label provider claims, published benchmarks, local observations, and unknowns separately. Note reasoning settings, output length, concurrency, and task/tool conditions where available; results under different conditions are not a controlled comparison.

## Recommend by role, not a universal ladder

| Role | Capability gate | Preferences after the gate |
|---|---|---|
| Throughput and discovery | Adequate extraction, locating, summarization, and instruction reliability | Prefer recent releases, low applicable cost, and fast useful throughput; an older better fit remains eligible |
| Ordinary implementation and verification | Sufficient coding, tool use, and test interpretation for the scope | Balance completion quality, cost, and time |
| Advanced interactive execution | Relevant visual interpretation, parameter/action understanding, and actual tool/session access | Compare complete-task success, latency, retries, media load, and handoff cost |
| High-risk architecture and review | Strong reasoning and evidence evaluation for the risk | Prioritize quality and appropriate independent checks within the user's budget |

These are suggested responsibilities, not mandatory agents or one fixed ranking. One model may cover several roles. An executor may use the same or a stronger model than the primary Agent. A strong coding score alone does not establish strong UI execution; do not require weak-model failure before a difficult interactive task receives an appropriate model.

For browser/Computer Use, separate interaction method, model, execution context, and resource limits. Prefer authorized structured access for equivalent work, and isolate long/high-output interaction when the executor has the required capabilities. Keep a bounded observe-decide-act stage with its owner; do not recommend permanent per-click tutoring by the primary. Short direct work and unavailable isolation remain subject to the user's browser policy.

## Present a shortlist

Use a compact comparison for candidates relevant to the user's roles, not an exhaustive provider dump. Include:

| Field | Required distinction |
|---|---|
| Identity and access | Exact exposed ID/provider, verified public identity or unresolved alias, usable roles/tools |
| Release | Relevant release/snapshot dates and which kind of date each is |
| Cost | Applicable provider rates/units/conditions; upstream-only or unknown explicitly labeled |
| Capability | Evidence relevant to the proposed role, limitations, reasoning/tool support |
| Speed | First-token, throughput, whole-task evidence and its conditions; unknown where unmeasured |
| Recommendation | Role fit, reason, alternative, tradeoffs, and confidence |
| Provenance | Sources, check date, observed versus published versus inferred |

Explain whether the user is trading money, waiting time, reliability, or retained context, rather than collapsing them into an unexplained score. Hypothetical round-count arithmetic is allowed only when labeled illustrative, not as a measured benchmark.

Do not run paid probes or benchmarks merely to fill missing cells. Such tests require applicable authorization, bounded cost/time, appropriate fixtures, and no unapproved external side effects. If approved, measure complete tasks and failures as well as successful individual calls. Do not install a benchmark dependency without permission.

## Let the user decide

Include unresolved role/model/reasoning choices and fallback authority in the consolidated configuration proposal, using [interview.md](interview.md) internally. Recommend a usable bundle with alternatives rather than asking the user to fill out a role matrix. Reuse explicit choices; a rejected recommendation is not a reason to keep persuading or silently substitute a model. Offer a permitted disclosed default or deferred choice when exact selection is unsupported.

For interactive work, resolve only relevant priorities, direct-use exceptions, and progress/stop limits. A model upgrade does not grant more permissions or reset the stage's retry/resource budget. Retention and media controls remain separate; do not promise that closing an executor deletes its history.

If adoption proceeds, put approved operational mappings and limits into self-contained effective instructions. Keep research tables, volatile prices, and question text out of runtime policy; an optional [adoption record](profile.md) may preserve requested provenance. Ordinary tasks reuse approved mappings. Refresh only affected choices when availability, meaning, or the user's priorities change.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall