goal-main-orchestrator

name	goal-main-orchestrator
version	0.2.117
description	Runtime-only main orchestrator for prepared /goal job bundles. Use when a /goal session has been launched from a goal-preflight bootloader and must consume existing job.manifest.json, main.prompt.md, and branch prompts; first run skill availability bootstrap and fail-closed prompt audit with telemetry, optionally use CLI-only Lite advisors after audit or completed branch artifacts for advisory summaries, then create path-validated branch worktrees, dispatch goal-branch-orchestrator sessions within the hard agent limit, summarize packet telemetry, and finish only when the main prompt's falsifiable Definition of Done is satisfied.

Goal Main Orchestrator

Runtime wrapper only. Do not implement branch work, rewrite preflight artifacts, or absorb goal-branch-orchestrator instructions into main context.

Start

Resolve the installed skills root:

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

Then print the compact deterministic phase table and follow it:

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

Runtime Rules

• Use job.manifest.json, main.prompt.md, prompt-audit.json, scheduler JSON, branch status/review JSON, telemetry JSON, and validator defects as the working surface.
• In debug telemetry mode, summarize_telemetry.py writes telemetry.debug.summary.json and root run.trace.jsonl; use those for efficiency, stall, fallback, and token-pressure analysis before raw event logs.
• Run the model catalog phase before choosing or launching model routes.
• Prompt audit must pass before branch worktree creation or branch dispatch.
• For resumable or interrupted runs, run reconcile_goal_run.py first to materialize orchestration.state.json/resume.report.json, then launch only work that is safe to resume.
• Keep branch orchestrator slots saturated up to max_active_branch_agents; waves are scheduling order, not dependency barriers.
• Before launching a branch, run render_branch_worktree_commands.py --branch Bxx --delegation-report /abs/bundle/branches/Bxx.delegation.json and use the report's bounded launch plan. Run scheduler_launch_command immediately before starting a native branch agent or CLI branch fallback. For CLI fallback, run the worktree command, launch_prompt_command, scheduler_launch_command, and launch_command in that order. The CLI launch_command defaults branch-control Codex to gpt-5.4-mini, redirects stdout/stderr to branches/Bxx.codex.log, and writes --output-last-message to branches/Bxx.codex.final.md; do not stream branch CLI output into the main model context.
• Wait on active branch agents; do not poll branch worktrees, worker packets, reviewer packets, or logs while branch agents are active.
• Branch/reviewer status promotion and launch-config integrity are validator-enforced; treat assemble_branch_status.py/validate_branch_status.py output as the source of truth before relaunching branch work.
• After each dispatch loop and before final reporting, reconcile terminal evidence and run validate_main_status.py on a freshly assembled main.status.json.
• Do not finish with only an in-progress prose summary while launched branches are still active or unresolved. Before any final answer, run scheduler close-from-artifacts, telemetry summarization, assemble_main_status.py, validate_main_status.py, and reconcile_goal_run.py; if validation cannot pass, preserve the blocker in validator-visible main.status.json evidence.
• Use goal-plan-amender only after validated terminal branch evidence and only for future unstarted work.
• Do not read or search skills/*/scripts/*.py during normal runtime, 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/prompt-audit-contract.md for prompt-audit artifact semantics.
• references/main-runtime-contract.md for validate_main_status.py defects.
• references/lite-advisor-contract.md before creating Lite packets.