goal-branch-orchestrator

name	goal-branch-orchestrator
version	0.2.117
description	Runtime-only branch orchestrator for an audited branch prompt and existing branch worktree. Use when goal-main-orchestrator has passed prompt audit, created a branch integration worktree, and launched a branch session that must run skill/CLI bootstrap, optionally use CLI-only Lite advisors for packet planning/context packing/completed-worker summaries/blocked triage, create path-safe worker/research-worker/reviewer packets with telemetry, choose allowed per-worker routes from the ds-pro-max -> ds-flash-max -> codex-spark -> codex-mini ladder (bridge deepseek leading, native Codex fallback) for normal workers, keep worker launcher slots saturated with ready workers, integrate results, dispatch a read-only heavy-model reviewer, and return only when the branch prompt's falsifiable Definition of Done is satisfied or blocked.

Goal Branch Orchestrator

Runtime wrapper for one audited branch worktree. Do not create the branch integration worktree, rewrite manifest/prompts, launch plan-amender, or do worker implementation yourself.

Start

Resolve the installed skills root:

GOAL_SKILLS_ROOT="${CODEX_HOME:-$HOME/.codex}/skills"
if [ ! -d "$GOAL_SKILLS_ROOT/goal-branch-orchestrator" ] && [ -d "$HOME/.agents/skills/goal-branch-orchestrator" ]; then
  GOAL_SKILLS_ROOT="$HOME/.agents/skills"
fi

Then print the compact deterministic phase table and follow it:

python3 "$GOAL_SKILLS_ROOT/goal-branch-orchestrator/scripts/runtime_phase_manifest.py" --markdown

Runtime Rules

• Use generated packets, route JSON, scheduler JSON, status/review JSON, telemetry JSON, diffs, tests, and validator defects as the working surface.
• Verify prompt audit passed and run the model catalog phase before choosing worker routes.
• Launch independent worker packets as a rolling saturated pool up to max_active_worker_packets.
• Preserve the manifest-configured worker route-class ladder by default. Do not pass --allow-route-pruning unless a validator, model catalog, route-health artifact, explicit operator request, timeout, budget cap, or provider failure justifies the shorter ladder with a concrete reason.
• Create reviewer packets only when pre-review validation succeeds and the branch status has no unresolved blocker evidence.
• Reviewer promotion is validator-owned: use validate_branch_status.py/create_pre_review_gate.py to ensure the branch review artifact is promoted to the canonical branch review path and launcher config route policy is consistent.
• If deterministic branch repair was command-verified after worker route failures, promote it through promote_worker_repair_evidence.py before pre-review; do not commit repair code while leaving the canonical worker status blocked.
• Use separate child worktrees for write-capable workers when owned paths do not conflict.
• Do not poll active worker/research/reviewer packet logs. Inspect artifacts only after launchers exit or return terminal status.
• Use context_pack.py/packet context excerpts instead of broad reads.
• In CLI branch-control mode, do not read, tail, grep, cat, or inspect your own redirected branches/Bxx.codex.log or branches/Bxx.codex.final.md. Do not read memory files or produce memory citations. If worker scheduler evidence has ready unlaunched work, create/run the worker packet; do not self-monitor. Finish only after writing and validating branches/Bxx.status.json, or after writing validator-visible blocked/partial branch evidence.
• Do not read or search skills/*/scripts/*.py during normal orchestration, including with rg, grep, cat, sed, or head. Inspect Python source only when a script failed and debugging that script is the assigned task.

Details On Demand

Open detailed references only after a phase script or validator points at an ambiguity:

• references/branch-runtime-contract.md for validate_branch_status.py defects.
• references/lite-advisor-contract.md before creating Lite packets.