introspect

Set per-plan caps, quotas, limits, and compute sizes in Companion's billing tiers by deriving the numbers from the existing price/credit ladder, not by inventing round or evenly-doubling numbers. Load when adding or changing a per-tier quantity (automations cap, storage, rate limit, feature quota, or per-plan compute/runtime sizing like vCPU/RAM or sandbox template) or editing PLAN_TIERS / COMPUTER_RUNTIME_TIERS in src/lib/billing/constants.ts. Not for one-off non-tiered constants.

Run Introspect for Codex sessions when the user asks Introspect to inspect recent corrections, update agent instructions, or report what it would remember.

How an LLM agent should emit a structured decision or host-consumed payload — a mode flag, control action, state change, routing choice, or UI data envelope — back to its host system. Use when building or debugging an agentic product where the host asks the model to return only JSON, signal quiet/silent/archive/done/hand-off state, choose an app action, feed a menu, or when markers/tags leak into the transcript or host code regex-scrapes prose.

Drive authenticated web-console and browser forms through semantic controls, Playwright/DOM locators, or form APIs instead of repeated coordinate clicks and blind keystrokes. Use when filling or editing forms in Chrome or the in-app browser, especially admin consoles, OAuth/client settings, secret managers, or pages where focus or typing is failing. Not for read-only page inspection or a single obvious click.

Retrieve an API token, secret, GitHub credential, or remote-session auth path from the machine's own stores before asking the user to paste it or working around auth with file transfer. Use when a CLI errors on auth, an env var like SENTRY_AUTH_TOKEN is unset, a deploy/API call needs a key, or a remote SSH shell cannot access credentials that exist in the GUI/session. Not for a secret that exists nowhere and must be freshly minted through a human-only step (hardware MFA, a one-time email/SMS code) — drive that flow per AGENTS.md instead.

Register a folder as an Obsidian vault before handing off `obsidian://open` links on macOS, so the user actually opens what you generated instead of hitting "Unable to find a vault for the URL".

Polish user-facing product surfaces such as pricing pages, plan cards, feature pages, grant/application packets, navigation rows, empty states, model/settings pickers, provider import/export prompts, provider logos, App Store/TestFlight metadata, icons, categories, onboarding, and marketing UI by reading real reference products and the product's enforced source of truth before writing copy, visuals, budgets, or interaction structure. For research grant packets, keep funder-required research intent distinct from product/prototype framing.

Create, update, prune, and validate agent-loop skills as scoped procedural memory instead of bloating AGENTS.md. Use for reflector decisions involving skill_new, skill_update, skill_prune, reusable workflows from observed failures, or placement decisions between a skill and the global prompt.

Restyle or build an in-app UI component (inputs, chips/tags, tabs, comboboxes, cards, dialogs, chat/message surfaces) by first identifying the right component archetype, reading real reference implementations, matching their concrete details and edge states, then verifying by viewing the rendered component. Use when the user calls a component ugly, off, or wants it to look like a real app or library. Not for marketing/pricing/store copy (use product-surface-polish) or component logic bugs.

Remaster, upscale, or rebuild visual assets and sprite packs with image/upscale/edit APIs. Use when improving pixelated images, choosing GPT/Gemini/Fal/Flux/Topaz/Aura/ESRGAN-style providers, preserving transparent backgrounds, building contact-sheet batches, or generating a production asset pack. Not for ordinary UI component styling or one-off standalone image generation.

Write or revise model-facing prompt text so the desired behavior is stated as the output to produce, then verify it with real response probes. Use when editing AGENTS.md, CLAUDE.md, system prompts, prompt rules, or when prompt wording makes the agent explain, ask permission, or stall instead of acting.

Keep Companion billing money values tied to source of truth: usage markup/credits in billing economics, and subscription plan prices across plan constants, Autumn/Stripe, RevenueCat, and App Store Connect when Apple IAP is in scope. Load when changing CREDIT_MARKUP, usage credit math, plan prices, top-ups, billing reconcile scripts, RevenueCat/App Store product prices, or paywall price displays. Not for per-plan quota or compute sizing; use tiered-plan-limits.

Forecasting and decision support where money, deadlines, or irreversible choices matter; prevents weak-model language from becoming a refusal to build and test a predictive model.

Create, edit, prune, review, and debug AGENTS.md or CLAUDE.md as the always-loaded agent prompt. Use for core_prompt decisions, global agent rules, prompt bloat, instruction failures, or placement decisions between AGENTS.md, hooks, and skills.