Status: Accepted (shipped in v2.x — evolved from the original v2.4.0 estimate) Date: 2026-05-20 Amended: 2026-07-03 (update status and implementation details post-ship)

When multiple skills run in the same session, the context window fills progressively. After skill 3, the window is ~60% full and recency bias degrades reasoning quality. By skill 5, quality degrades ~50% relative to a fresh start. Re-reading the same files (CONVENTIONS.md, state.yaml) in every skill wastes tokens redundantly.

Each skill spawn receives a fresh, isolated context window (target: 200K tokens). The orchestrator passes only the files explicitly required by that skill via a <files_to_read> declaration. Session history is not passed. Prior decisions are surfaced via specs/state.yaml, not conversation replay.

The original design envisioned a single orchestrate-project coordinator. The shipped implementation uses two complementary mechanisms:

1. delegate-task / dispatch-agents: Sub-agents get fresh context windows with only the skill’s SKILL.md, the current epic capsule, and state.yaml — no session history or prior conversation. dispatch-agents runs multiple sub-agents in parallel on disjoint tasks.

2. session-state: Handoff data (handoff.next_skill, epic_cycle.*, metrics.*) is written to specs/state.yaml at the end of each critical-path skill. The next agent reads state.yaml to resume — no conversation replay needed. This replaced the original STATE.md concept (state.yaml is the YAML cockpit format adopted in v2.0.0).

• Quality is consistent across all skills regardless of session length.
• Token cost per skill drops ~20% (no re-reading).
• Orchestrator complexity increased as forecast, but the dual-mechanism design (isolated agents + state.yaml handoff) proved simpler than the original single-coordinator model.
• Skills can chain themselves via handoff.next_skill in state.yaml — the isolation is at the agent level, not the skill level. A skill’s SKILL.md declares its next step, and the host harness launches a fresh agent to execute it.