Task Orchestrator

task-orchestrator

This is a meta-skill. It controls HOW other tasks are executed, not WHAT they do. It must be loaded before any non-trivial work begins.

Pipeline (mandatory for every atomic task)

graph LR
    R["🔍 RESEARCHaudit, gatherevidence"] --> D["🎨 DESIGNdesign/optimizeselect template"]
    D --> P["📋 PLANnumbered stepsverification per step"]
    P --> E["⚡ EXECUTEfollow planrecord results"]
    E --> V["✅ VERIFYcheck evidencerun CI, tests"]
    V -->|fail| R
    V -->|pass| DONE["Done"]

Phase 1: RESEARCH

• Audit current state: read relevant files, check CI logs, inspect git status
• Identify all stakeholders: which files, systems, or processes are affected
• Gather evidence: do NOT assume, do NOT guess, do NOT rely on memory
• Record findings in .task-orchestrator/research-${task}.md

Phase 2: DESIGN

• Design the solution approach (or select from existing template)
• For known task types: load the subskill template from skills/.task-orchestrator-templates/
• For novel tasks: design from first principles
• If the model/agent suggests a BETTER approach than the template: flag it for template upgrade
• Record design decisions in .task-orchestrator/design-${task}.md

Phase 3: PLAN

• Decompose into numbered, verifiable steps
• Each step must have: action, expected outcome, verification method
• If using a template: adapt the template steps to the specific context
• Use update_plan tool when available
• Record plan in .task-orchestrator/plan-${task}.md

Phase 4: EXECUTE

• Follow the plan step by step
• Do NOT skip verification steps
• Do NOT claim completion without evidence
• Record completion status per step

Phase 5: VERIFY

• For each step: confirm the expected outcome with evidence
• Run structural checks (validator, linter, test suite)
• Check CI status if applicable
• If verification fails: return to RESEARCH phase
• Record verification in .task-orchestrator/verify-${task}.md

Model/Agent Capability Profiling

The orchestrator adapts its strictness based on detected capabilities:

| Profile | Agent | Model | Strategy | |---|---|---|---| | High-capability | Claude Code | Opus 4.6+, GPT-5+ | Templates are REFERENCE only. Agent can self-derive approach. Light audit at end. | | Medium-capability | Codex | GPT-4, Claude 3.5 | Templates are GUIDELINES. Follow pipeline but allow adaptation. Full verify. | | Low-capability | Codex | DeepSeek, Gemini, local | Templates are MANDATORY. Strict pipeline enforced. No deviation without explicit approval. Extra verification gates. |

Model-tier sub-agent routing (token thrift, task 10)

Model tiers per agent are declared once in adapters/models.config.json (high/mid/small) and resolved by scripts/resolve_model.mjs:

node scripts/resolve_model.mjs              # e.g. claude research -> claude-opus-4-8
node scripts/resolve_model.mjs   --effort   # e.g. claude mid      -> max

Each tier also carries a reasoning effort (adapters/models.config.json → effort). The mid tier is claude-sonnet-5 at max effort — a strong Sonnet-5 run costs far less than Opus but, at max effort, holds quality on implement/verify. So when you spawn a mid sub-agent, pass model: claude-sonnet-5, effort: max.

Use it to cut token cost WITHOUT losing quality on the hard steps:

• Claude Code cannot switch model mid-session, so run the MAIN loop on high and delegate the cheap,

mechanical, or verify steps to SUB-AGENTS resolved to mid/small (e.g. resolve_model.mjs claude verify → claude-sonnet-5 + --effort → max). Guard: only delegate when the sub-agent's token-in is less than the tokens saved by the smaller model — spawning a sub-agent that re-reads a huge context can cost MORE. Measure before defaulting.

• opencode switches natively; its model/small_model in opencode.json are generated from this same

config, so the tiers already apply.

• Codex switches via external tooling (cc-switch / provider config). NEVER auto-configure provider

switching for Claude Code (account-ban risk).

Map task-kinds → tiers via the config's task_tier (research/design → high, implement/verify → mid, mechanical → small) so the RESEARCH→DESIGN→IMPLEMENT→VERIFY pipeline can pick a sensible model per stage.

Capability detection

# The orchestrator detects capability from:
# 1. Environment variables: CODEX_MODEL, CLAUDE_MODEL
# 2. Config files: ~/.codex/config.toml, ~/.claude/settings.json
# 3. Self-assessment: ask model to rate its own capability on known benchmarks

Subskill Plan Templates

Templates live in skills/.task-orchestrator-templates/. Each template is a subskill with frontmatter that records:

• template-for: what task type this covers
• last-updated: when the template was last improved
• min-capability: minimum model capability needed to skip this template
• improvements-from: which model/agent contributed improvements

Available templates

Each is a directory under skills/.task-orchestrator-templates/ containing a SKILL.md.

| Template | Task | Auto-selects when | |---|---|---| | feature-audit/ | Audit existing features for gaps | task mentions "audit" or "check" or "verify all" | | workflow-fix/ | Fix CI/CD workflow failures | task mentions "fix workflow" or "fix CI" or "fix actions" |

For task types without a template, derive the plan from the RESEARCH→DESIGN→IMPLEMENT→VERIFY pipeline directly — do not reference a template that has no directory.

Planned templates (NOT yet shipped — must NOT be auto-selected)

On the roadmap but no SKILL.md exists yet, so the orchestrator must not load or advertise them as available: hook-add (add a hook, Claude + Codex), skill-create (Claude source + Codex wrapper), rename-project (rename across all files), dependency-update (safe dependency upgrades). Add a directory + SKILL.md before wiring any of these in — otherwise it is a dead knob.

Template self-improvement protocol

When any model/agent discovers a BETTER approach than the template:

1. Record the improvement in .task-orchestrator/improvements/${task}-${date}.md
2. Flag with @template-upgrade in the improvement record
3. The maintainer reviews flagged improvements periodically
4. Approved improvements are merged into the template
5. Template last-updated and improvements-from fields are updated

Template version tracking

Each template has a version history in its frontmatter:

template-version: 3
version-history:
  - version: 1
    date: 2026-07-01
    author: claude-code/opus-4.6
    changes: Initial template
  - version: 2
    date: 2026-07-05
    author: codex/deepseek-v3
    changes: Added OpenSSL fix, gitleaks allowlist step
  - version: 3
    date: 2026-07-08
    author: codex/gpt-5
    changes: Added pre-commit-ci.yml multi-workflow audit step

Quickstart

For any task

1. Load this skill: /skills → task-orchestrator
2. State your task. The orchestrator will:

a. Detect model/agent capability b. Select appropriate template (if exists) c. Guide you through Research → Design → Plan → Execute → Verify

1. The orchestrator creates artifacts in .task-orchestrator/

For maintainers adding new templates

# Use the new-skill meta-skill, selecting "task-orchestrator-template" type
/new-skill
# Answer: type=task-orchestrator-template, task=

Template files go in skills/.task-orchestrator-templates//SKILL.md.

Anti-patterns (what this skill prevents)

• "I'll just fix it" without checking CI logs → RESEARCH phase catches this
• "This looks right" without running verification → VERIFY phase catches this
• "I already fixed this" based on memory → evidence requirement catches this
• "The workflow should pass now" without checking CI → CI check in VERIFY catches this
• "All the other features are fine" without auditing → systematic audit template catches this

Integration with code-verifier

When EXECUTE or VERIFY phases involve code quality claims, code-verifier is invoked automatically (via allow_implicit_invocation). The orchestrator does not replace code-verifier; it ensures code-verifier is called at the right time.

Integration with review-gate

After all tasks complete and the session ends, review-gate fires automatically (via Stop hook). The orchestrator's .task-orchestrator/ directory provides the review-gate with structured evidence of what was done and verified.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jajupmochi
• Source: jajupmochi/agent-harness
• License: MIT

Install and usage instructions live in the source repository linked above.