Agent OS Knowledge

Start by task

• Learn, apply, and contribute safely: Using Agent OS with Task-Scoped Context - Learn from the KB, apply one to three matching concepts, and suggest sanitized candidates without overriding project authority.
• Run a controller or recover a session: Agent Operating Principles, Persistent Codex Goals, and Named Session Reuse.
• Curate local skill exposure: Curated Skill Exposure Without Source Deletion.
• Delegate work or supervise completion: Controller-Worker Goal Execution Framework, Codex Native Controller-Worker Protocol, Effective Subagent Delegation, Completion Notification without Polling, and Supervisor Wake Principles.
• Load context on change, not every turn (cold-start vs warm continuation): Context Loading Policy.
• Continue an existing PR without redundant context reloads: Context Loading Economy for Agent Handoffs.
• Manage temporary agent evidence: Temporary Agent Evidence Lifecycle.
• Use ChatGPT Projects, files, and Skills safely: ChatGPT Project Context and File Surfaces.
• Choose event-triggered vs scheduled ChatGPT automation: ChatGPT Event-Triggered vs Scheduled Tasks.
• Discover repositories or prove a local web app: Git-Aware Repository Search, Practical Agent Tool Recipes, and Agent-Run Browser E2E and Screenshot Evidence.
• Start or retrofit an agent-managed repository: KISS Repository Starter Adoption.
• Choose development/runtime patterns: Development and Runtime Pattern Selection - Select containers, CI runners, identity, validation, browser E2E, runtime evidence, cost, and cleanup proportionally.
• Handle a connector/platform safety block: Connector Safety Gate.\n- Control AWS durably from GitHub: GitHub OIDC and AWS-Owned Control - Separate runner, OIDC identity, short controller, and AWS-owned long execution.
• Research through X or Grok: Grok Route Selection and Validating X and Grok Research Results.
• Prepare or promote public knowledge: Knowledge Lifecycle and Publication and the contribution guide.

Principles

• Agent Operating Principles - Practical principles for concise, evidence-backed agentic work.

Policies

• Knowledge Lifecycle and Publication - A public-safety gate from candidate capture through retirement, including minimal reusable-learning closeout.
• Connector Safety Gate - Fail closed on repeated connector safety blocks without widening scope, target, permissions, or safeguards.
• Context Loading Policy - Load context on change, not on every turn: event-driven cold-start and warm-continuation reads with explicit reload triggers.

Codex Playbooks

• Persistent Codex Goals - How runtime goals support long tasks without replacing repository truth.
• Named Session Reuse - Conservative recovery of known sessions by durable identity.
• Repository-First Local Agent Control - Discover and attach to local workers by repository identity, with a durable task-runner fallback.
• Curated Skill Exposure Without Source Deletion - Keep reusable source while enabling only a small, reversible task-fit catalog.

Delegation Playbooks

• Controller-Worker Goal Execution Framework - Delegate one complete goal under explicit approval, gates, evidence, and stop conditions.
• Codex Native Controller-Worker Protocol - Use durable goals and results with native queue transport while keeping admission, execution, completion, and acceptance separate.
• Effective Subagent Delegation - Balanced delegation with one writer and explicit evidence.
• Completion Notification without Polling - Durable results followed by one bounded notification.
• Supervisor Wake Principles - Fail-closed, event-driven wake behavior.

Knowledge Use Playbooks

• Using Agent OS with Task-Scoped Context - Learn from Agent OS, apply the smallest relevant knowledge set, and contribute sanitized candidates.
• Temporary Agent Evidence Lifecycle - Bound temporary evidence to one goal, one result, selected proof, and explicit cleanup authority.

Repository Discovery and Tools

• KISS Repository Starter Adoption - Create or retrofit a repository with the current role-specific root truth files while keeping CONTEXT.md current-only.
• Development and Runtime Pattern Selection - Choose reproducibility, containers, runner substrate, OIDC/secrets, validation layers, browser E2E, runtime evidence, cost, and cleanup without forcing one stack onto every repository.
• Git-Aware Repository Search - Separate tracked inventory, working-tree discovery, content search, and direct source verification.
• Practical Agent Tool Recipes - Compact recipes and failure modes for common repository tools.
• Agent-Run Browser E2E and Screenshot Evidence - Automate local startup, live-state assertions, positive and negative browser proof, screenshots, and cleanup.
• Run Playwright Core with Windows Node and Chrome from WSL - Launch installed Windows Chrome for interactive E2E, with CDP attachment as a fallback.

Provider and Integration Playbooks

• Grok Route Selection - Select a bounded research route without exposing authentication material.
• Validating X and Grok Research Results - Source, timestamp, telemetry, and cost gates for research output.
• ChatGPT and Codex Collaboration Protocol - GitHub-only ChatGPT review exchange, native queue transport, milestone sizing, and evidence gates.
• Context Loading Economy for Agent Handoffs - Treat PRs/comments as continuation pointers and reload broader context only on explicit recovery triggers.
• ChatGPT Project Context and File Surfaces - Keep Project instructions, Project files, Library, attachments, runtime copies, and Skills conceptually separate.
• ChatGPT Event-Triggered vs Scheduled Tasks - Use webhook-driven Work tasks for supported external events and schedules for time-driven work.\n- GitHub OIDC and AWS-Owned Control - Keep repository access, runner substrate, OIDC identity, and AWS-owned execution separate; persist execution IDs for restart-safe control.

Model Observations

• Spark versus Luna Evaluation - A dated single-sample comparison of latency and resource use.