agentic-cli-config

name	agentic-cli-config
description	Keep the project's Claude Code / Codex / Gemini CLI configuration in sync with upstream — model catalogue, effort levels, context windows, beta headers, default flags. Use when the user says "mets à jour les modèles", "refresh CLI config", "y a-t-il un nouveau modèle Claude/Codex", "check effort levels", "context window à jour", "1M beta header", or before any release notes / model launch (Anthropic, OpenAI). Runs WebSearch + fetches official docs, diffs against the current catalogue, and patches the source files.
disable-model-invocation	false
allowed-tools	Read, Edit, Write, Glob, Bash(ig *), Bash(bun *), WebSearch, WebFetch

agentic-cli-config

Single-source-of-truth refresher for everything that depends on upstream CLI behaviour (Claude Code, Codex, Gemini). When Anthropic or OpenAI ship a model, change a flag, deprecate a beta header, or move defaults, this project drifts silently. This skill detects the drift and patches it.

Files this skill owns (the canonical surface)

File	Owns
src/lib/chat/context-windows.ts	DEFAULTS map (cli:model → tokens), CLI_FALLBACK, supportsOneM()
src/lib/chat/models.ts	CODEX_MODELS, DEFAULT_CODEX_MODEL, claudeExtraArgs(), codexExtraArgs()
src/components/ide/model-pill.tsx	CLAUDE_MODELS, OPENCODE_MODELS, GEMINI_MODELS, DEFAULT_BY_CLI
src/components/ide/reasoning-pill.tsx	EFFORTS_CLAUDE, EFFORTS_CODEX, ReasoningEffort union, DEFAULT_EFFORT
src/lib/chat/claude-adapter.ts	ANTHROPIC_BETAS env (e.g. context-1m-2025-08-07), --effort mapping
src/lib/chat/codex-adapter.ts	--effort mapping for codex
src/components/assistant-ui/status-button.tsx	Hint copy under the 200K/1M chips

If a fact lives elsewhere it should be migrated here first.

Authoritative sources (always check these, in order)

Claude / Anthropic

1. Claude Code CLI docs — https://code.claude.com/docs/en/model-config (effort levels, defaults per model, env vars)
2. Claude Code release notes — https://code.claude.com/docs/en/release-notes (look for "v2.x", "default effort", "model rollout")
3. Anthropic platform release notes — https://platform.claude.com/docs/en/release-notes/overview (model launches, beta headers, deprecations)
4. Anthropic models reference — https://docs.anthropic.com/en/docs/about-claude/models/overview (context windows, ids)
5. Beta headers index — https://docs.anthropic.com/en/api/beta-headers (current header names, sunsets — e.g. context-1m-2025-08-07 retiring on Sonnet 4.5/4 on 2026-04-30)

Codex / OpenAI

1. OpenAI Codex CLI repo / README — https://github.com/openai/codex (flags, model ids, --effort)
2. OpenAI release notes / model index — https://platform.openai.com/docs/models (gpt-5.x lineup, context windows)

Gemini

1. Gemini CLI docs — https://github.com/google-gemini/gemini-cli
2. Gemini API model list — https://ai.google.dev/gemini-api/docs/models

Prefer WebFetch on these exact URLs over WebSearch — fewer hallucinations, dated content. Use WebSearch only to discover new URLs (release titles, blog posts) and then WebFetch the canonical doc.

Workflow

1. Snapshot the current state

Run, in parallel:

ig "DEFAULTS|CLI_FALLBACK|supportsOneM" src/lib/chat/context-windows.ts
ig "CLAUDE_MODELS|OPENCODE_MODELS|GEMINI_MODELS|CODEX_MODELS" src/components/ide/model-pill.tsx src/lib/chat/models.ts
ig "EFFORTS_CLAUDE|EFFORTS_CODEX|ReasoningEffort" src/components/ide/reasoning-pill.tsx
ig "ANTHROPIC_BETAS|context-1m" src/lib/chat/

Read each match and tabulate (model id, label, tokens, default effort, beta headers).

2. Pull upstream truth

Foreach CLI in {claude, codex, gemini}:

• WebFetch the model-config / release-notes URL
• Extract: model ids · context window (input tokens) · default effort · supported --effort values · required beta headers · deprecation dates

Cross-check at least two of the listed sources before changing a fact. A single blog post is not enough — the canonical doc must agree.

3. Diff, then patch

Build a tiny diff table:

model              tokens    default-effort   1M?    beta-header
claude-opus-4-7    1_000_000 xhigh            yes    (none, native)
claude-opus-4-6    200_000   high             yes    context-1m-2025-08-07
claude-sonnet-4-6  200_000   high             yes    context-1m-2025-08-07
claude-haiku-4-5   200_000   medium           no     —
gpt-5.5            400_000   high             —      —

For each row that differs from the codebase, edit the appropriate file. Never invent values — if the doc is silent, leave the existing value and note it in the report.

4. Verify

bunx tsc --noEmit 2>&1 | grep -E "(context-windows|reasoning-pill|model-pill|claude-adapter|codex-adapter|status-button|models)\.ts" || echo OK
bun run lint 2>&1 | tail -20

Then run the relevant Playwright smoke if effort/context UI changed:

bunx playwright test e2e/smoke.spec.ts -g "model pill\|status"

5. Report

Output a short summary: what changed, what stayed, what was uncertain (and the URL to confirm later). Cite source URLs.

Hard rules

• Never add a model id that isn't documented on an official Anthropic/OpenAI/Google page. No "claude-opus-5-0" guesses.
• Never flip supportsOneM on by default for a model unless the doc explicitly lists 1M. Beta-header models stay opt-in.
• Never remove an existing model id without checking it's actually deprecated upstream (the user may still pin it). Mark legacy models in the description ("Legacy · 200K") instead.
• Always preserve the claude-adapter.ts ANTHROPIC_BETAS injection when 1M is selected. If Anthropic ships a new beta header (e.g. context-2m-…), add it under a new override value — don't overwrite the existing one.
• Always keep EFFORTS_CODEX capped at the levels Codex actually supports. The defensive xhigh|max → high map in codex-adapter.ts is the safety net, not a substitute for the right list.
• Always update status-button.tsx hint copy when the rules change (currently mentions context-1m-2025-08-07).

Quick triggers

User says	Do
"y a-t-il un nouveau Opus / Sonnet / Haiku ?"	Steps 1→3 for Claude only
"Codex a sorti gpt-5.6 ?"	Steps 1→3 for Codex only
"refresh CLI config" / "mets à jour les modèles"	Full sweep, all CLIs
"1M context broken on Sonnet 4.6"	Confirm beta header in Anthropic docs, verify supportsOneM + ANTHROPIC_BETAS injection
"default effort changed in Claude Code v2.X"	WebFetch release notes, update DEFAULT_EFFORT and per-model defaults

Out of scope

• Don't touch UI styling, copy unrelated to model facts, or unrelated adapters.
• Don't change MAX_HISTORY_TURNS / MAX_PROMPT_CHARS — those are project-tuned, not upstream-driven.
• Don't auto-bump to a new major Claude Code CLI version unless the user asks (breaking flags possible).