gpt55-auto-orchestrator

name	gpt55-auto-orchestrator
description	Automatically choose Project Governor skills, subagents, model strategy, context budget, and quality gates for Codex GPT-5.5 without requiring the user to specify the workflow.

GPT-5.5 Auto Orchestrator

Purpose

Use GPT-5.5's stronger tool use, completion persistence, and coding/research ability without increasing user prompt burden.

This skill should be used automatically when the user asks to implement, fix, refactor, review, research, upgrade, initialize, refresh, or clean a Project Governor project.

Goal

The user should not need to know which Project Governor skill, subagent, model, or context workflow to invoke. The orchestrator should turn the request into the smallest safe governed workflow that can finish with useful evidence.

Success means:

• the request type, route, context budget, skill sequence, model plan, and quality gate are selected from existing Project Governor policies;
• subagents are selected only when they materially reduce risk, and host-runtime authorization status is surfaced before spawning;
• execution-policy contexts are surfaced when the user selects a command tool or transport;
• the plan preserves project boundaries and never copies plugin-global assets into target projects;
• the workflow stops once the core request can be answered or completed with enough evidence.

Model policy

Prefer:

• gpt-5.5 for implementation, hard reasoning, tool-heavy workflows, code review, research synthesis, migrations, and risk assessment.
• gpt-5.4-mini for fast read-only scouting, context retrieval, simple lint-style classification, summaries, and low-risk subagents.
• gpt-5.4 as fallback when gpt-5.5 is unavailable.

Start GPT-5.5 planning at medium reasoning for standard, research, migration, and upgrade work. Use low for fast docs, micro, and tiny routes. Escalate to high only for risky_feature, strict routes that are not ordinary upgrade/research/refactor planning, or when validation shows medium reasoning is not enough.

Context policy

Do not read all initialization docs at session start.

Instead:

1. read AGENTS.md only when available and small,
2. read .project-governor/context/DOCS_MANIFEST.json if present,
3. read .project-governor/context/SESSION_BRIEF.md if present,
4. query .project-governor/context/CONTEXT_INDEX.json,
5. read returned recommended_sections line ranges before full documents,
6. read full files only when sections are insufficient or confidence is low,
7. escalate to context-indexer if the index is missing or stale.

Workflow policy

Run the deterministic helper first when possible:

python3 skills/gpt55-auto-orchestrator/scripts/select_runtime_plan.py examples/gpt55-runtime-standard-feature.json

Then follow the returned skill_sequence and subagents. If subagent_authorization.status is needs_explicit_user_authorization, ask once for user consent before calling any spawn tool. The user should not have to name individual subagents.

If execution_policy.required is true, carry its quality_gate_input into the final quality-gate input and record the user-selected command constraint in the task plan or evidence.

Output

Return:

• selected workflow,
• selected model plan,
• selected subagents,
• subagent authorization status,
• execution policy context and quality-gate input hints,
• context budget,
• route doc pack,
• skill sequence,
• skipped skills with reasons,
• quality gate level,
• clean reinstall / latest mode actions when relevant.