Update Model Pricing

Keep the daemon's offline cost fallback aligned with current public OpenAI Standard API pricing and the codex-acp version selected by AgentConnect's runtime registry or an explicit runtime override.

Workflow

1. Read packages/daemon/src/usage/openai-public-pricing.ts, its focused test, and packages/daemon/src/runtimes/registry.ts. Resolve the effective codex-acp package/version from the current ACP registry entry or the deployment's explicit runtime command/args; it is not pinned in pnpm-lock.yaml.
2. Read references/pricing-sources.md and fetch the current official sources. Do not rely on remembered prices or third-party tables.
3. Re-check the resolved codex-acp tag's TokenCount mapping and compare it with upstream main before changing the formula. Confirm whether ACP input, cache-read, cache-write, output, and thought fields are disjoint and whether PromptResponse usage is per-turn.
4. Update the manifest with exact published model IDs, prices, dated snapshots, thresholds, and the verification date. Cover text/reasoning models that codex-acp advertises or AgentConnect explicitly supports; do not broaden the manifest to unrelated OpenAI products. Keep unknown models unsupported; never infer a price by prefix or strip an unlisted date suffix.
5. Update tests in the same patch. Cover every alias, the exact long-context boundary, cache behavior, unknown models, malformed usage, and reasoning non-duplication.
6. Run the focused daemon tests, daemon typecheck, and formatting check. Review the final diff for unrelated product or telemetry changes.

Repository Rules

• Price Standard API tokens in USD per one million tokens. Do not branch on ChatGPT subscription versus API-key authentication.
• Prefer an ACP-provided cost for a turn. Run fallback calculation only when that turn reports no cost.
• Keep cost display unchanged. Do not add estimated labels, approximation glyphs, or CP/Web schema fields unless explicitly requested.
• Calculate each Codex turn with the model selected for that session at turn start; do not reprice a whole session using its final model.
• Treat output tokens as already including reasoning tokens. Never add thought tokens a second time.
• Apply a long-context tier only when the public table publishes one and request input is strictly above its documented threshold.
• Keep runtime pricing offline and deterministic. Never fetch prices from the daemon at runtime.
• Preserve historical stored amounts; a price refresh affects only future fallback turns.

Validation

Run:

pnpm --filter @agentconnect.md/daemon exec vitest run test/openai-public-pricing.test.ts test/local-store.test.ts test/daemon-webchat.test.ts test/acp-host.test.ts
pnpm --filter @agentconnect.md/daemon typecheck
pnpm format:check

If codex-acp's usage mapping changed, add a regression that demonstrates the new upstream shape before modifying the fold or cost formula.