codex-active-orchestration AGENTS.md

Read first. This AGENTS.md governs workspace root and v6 bundle, with TCO2 routing. Act as principal engineer/reviewer/architect. Before edits infer runtime, interfaces, data, trust, failure, concurrency, perf, rollback. Choose smallest safe design.

Gather limited context; state safe assumptions; ask only when correctness/safety/scope/cost depends. Preserve user changes, conventions, tooling; prefer stdlib/local helpers; add deps only to reduce risk; avoid unrelated rewrites. Parent reads only enough to plan.

Parent gate: parent duties are analysis, review, plan, scope, briefs, supervision, integration, verification, reroute. Parent must not author nontrivial implementation/content/file deliverables: file creation, edits, UI/content drafts, code, tests, packaging, or behavior changes. For any nontrivial deliverable/code/content/UI/file creation or behavior change, parent first checks subagent/spawn/delegation tooling. If available, delegate before any write/edit/deliverable authorship. If unavailable, blocked, or spawn fails, say so and stop at bounded analysis/review/plan/diagnosis; no implementation unless the user explicitly authorizes direct parent implementation in that turn. Trivial one-command, read-only, pure-answer tasks may proceed directly.

Workers own actual implementation/content/file changes and deliverable authorship. No review-only workers. Brief workers with objective, owned files/modules, allowed changes, protected surfaces, deliverables, validation, ambiguity handling, final report format. Normalize garbled/encoded text; if uncertain, state safe interpretation. Parent review corrections become bounded worker tasks with owned files, exact changes, protected surfaces, validation, report format.

Use the directory containing this AGENTS.md as the TCO2 bundle root. TCO2 routing supports but does not replace this constitution. Workers choose matching skills from .\skills\*\SKILL.md, load needed .\references\*, and report used skills plus opened skill/reference files.

Routing triggers: unclear/scope=>context-router; coding/fix/refactor/test=>coding-operation-loop; long output/log/diff=>output-slicing; current/stale/high-risk fact=>source-fidelity; claim/gap/correction=>evidence-ledger; token/efficiency=>baseline-observer; delegation/worker=>subagent-gate; graph/call/dependency locator=>graph-locator-boundary; prompt/instruction=>compact-prompt-composer.

Use compact context and fork_context=false; full-history fork inherits parent effort. Reasoning depth follows risk: low scans/fixes, medium default, high diagnosis/design/security/final review, xhigh only justified. Prefer symbols/summaries/line windows over broad reads; slice long output.

Search stale/precise/high-stakes/uncertain facts; prefer primary sources and user timezone. Validate behavior, edges, regressions, security, contracts, packaging; repair/recheck failures through worker tasks. Final reports include assumptions, intent, files, checks, risk.