Context Loading Economy for Agent Handoffs

The canonical cold-start versus warm-continuation rule lives in the Context Loading Policy (load on change, not on every turn). This playbook is the handoff-focused procedure for applying that rule.

Use the smallest durable pointer sufficient to continue safely. The owning Issue/PR records current work; the latest relevant authorized comment is the delta. Optional rereading is not optional authority. Task history does not replace enduring requirements, host constraints or observed runtime evidence.

Normal continuation

When repository identity, objective and governing context remain usable:

owning Issue/PR -> relevant authorized delta -> current HEAD/diff
-> changed governing context and required runtime evidence -> continue

Give the existing pointer and exact comment URL only when useful. Do not copy the full plan, completed history or file-reading checklist. go stays within the selected scope and mode; a newer or agent-authored comment is not approval merely because it is the latest comment.

Full context reload triggers

Reload affected context when the session lacks it, a governing file materially changes, repo/goal/host identity is ambiguous, evidence contradicts cached state, or a new authority/security/production/destructive boundary is entered. A PR number or milestone change alone does not require a full reload.

Use existing fingerprints/references when available, not a new caching system. If freshness cannot be established, reload the relevant contract before action.

Bootstrap / recovery order

Follow repository-specific cold-start requirements. Generally establish:

  1. repository/worktree identity, HEAD and local AGENTS.md;
  2. owning Issue/PR and relevant authorized decision/handoff;
  3. applicable operating contract, desired-state configuration and host/project environment context;
  4. unique restart state and live facts required for the action.

A locally required SPEC.md cannot be skipped because a generic starter makes that filename optional. Missing authority blocks the controlled action.

Handoff economy

The existing record must preserve outcome, allowed scope, acceptance, stop gates and relevant research/evidence. Same-PR corrections need head, delta and proof, not another packet. The collaboration playbook owns transport and human-copy handoff rules.

Project and local context

The core-file model separates reusable Agent OS policy, repository contracts/configuration, GitHub current-work records and host-specific knowledge.

Preserve the existing dotfiles/local host profiles and active-host selection. Project ENV.md, when needed, records workload dependencies and approved execution targets, not another copy of the entire machine inventory. Resolve the execution host, not just the browser host. If that host forbids a required tool installation, select an authorized compatible target or report the blocker. Remote compute does not authorize moving corporate data elsewhere.

Credentials/authentication state stay out of these documents. Private facts stay in appropriately private/local sources, never public Agent OS. Manage ChatGPT Project source membership through its owning product surface and preserve unique information before removing duplicates.

Safety invariants

Stop or reload on uncertain identity, conflicting evidence/authority, material HEAD changes or entry into a new trust boundary. Preserve explicit no-go gates and enforced controls. A worker observation, recommendation or receipt is not approval. Reconstruct missing authority; never guess.

Adoption pattern

Before removal or renaming, map unique content to its surviving home and audit scripts, prompts, links and runtime consumers. CONTEXT.md and CHATGPT.md remain where they contain unique recovery state or adapter rules.

Prove normal warm continuation and an authority/host-sensitive case on one non-critical pilot; a paper policy review is not a live pilot result. Keep rollback, then simplify downstream repositories gradually when next touched. Do not mass-delete files or disturb live delivery for smaller file counts.

project learning -> sanitize/deduplicate -> Agent OS
-> starter where universal -> existing repos when next touched