95 lines
3.6 KiB
Markdown
95 lines
3.6 KiB
Markdown
---
|
|
name: cobb
|
|
description: Dom Cobb orchestrator for the Factory Droid workflow in any repository with explicit model routing, spec confirmation, and verification gates. Trigger word: Cobb. Use when work should be split across explorer, spec, coder, quality, reviewer, and runner droids.
|
|
---
|
|
|
|
# Cobb Orchestrator
|
|
|
|
## Purpose
|
|
|
|
Use this skill when the user asks for `Cobb` (the Inception-inspired orchestrator) to run a consistent multi-droid workflow for coding tasks in a target repo while enforcing the bundled rules in `curated-playbook.md`.
|
|
|
|
## Invocation
|
|
|
|
- Trigger word: `Cobb`
|
|
- Orchestrator name: `Dom Cobb`
|
|
|
|
## Required Inputs
|
|
|
|
- Task goal and success criteria
|
|
- Scope boundaries (allowed files/directories)
|
|
- Risk and autonomy constraints (`--auto` expectations)
|
|
- Size of codegen (`large` vs `small`) to choose coder model
|
|
- Whether any new markdown output is explicitly requested
|
|
|
|
## Policy Precedence
|
|
|
|
When instructions conflict, apply this order:
|
|
|
|
1. Latest explicit user instruction
|
|
2. `SKILL.md`
|
|
3. `curated-playbook.md`
|
|
4. `checklists.md`
|
|
|
|
If conflict remains unresolved, stop and ask the user before code generation.
|
|
|
|
## Canonical Role Routing
|
|
|
|
- Driver and explorer: `custom:Kimi-K2.5`
|
|
- Spec/planning: `custom:Gpt-5.2`
|
|
- Large code generation: `custom:Gpt-5.3-Codex`
|
|
- Small code generation/fixes: `custom:Kimi-K2.5`
|
|
- Quality and run/build/test: `custom:Kimi-K2.5`
|
|
- Review/bug finding: `custom:Opus-4.6`
|
|
|
|
Always verify IDs first with `droid exec --help` before issuing commands.
|
|
|
|
## Workflow
|
|
|
|
1. Preflight
|
|
- Run `droid exec --help` and confirm model IDs.
|
|
- Create/update todos before substantive work.
|
|
- Restate objective, constraints, and expected outputs.
|
|
2. Start with Kimi driver
|
|
- Use `custom:Kimi-K2.5` as entrypoint and coordinator.
|
|
3. Explore in parallel
|
|
- Launch one or more Kimi exploration prompts for relevant parts of the task.
|
|
4. Build a context packet
|
|
- Consolidate: objective, touched paths, findings, open questions, constraints, and accepted assumptions.
|
|
5. Spec stage
|
|
- Send context packet to GPT 5.2 for plan/spec refinement.
|
|
- Confirm spec with user before any code generation.
|
|
6. Code stage
|
|
- Use GPT 5.3 Codex for large codegen; Kimi for small edits.
|
|
- Keep edits scoped to approved plan.
|
|
7. Quality stage
|
|
- Run formatting/lint/type checks with Kimi.
|
|
8. Review stage
|
|
- Run reviewer pass with Opus 4.6 focused on bugs, regressions, and risk.
|
|
9. Run stage
|
|
- Run build/test/runtime checks with Kimi.
|
|
10. Summarize
|
|
- Report what changed, evidence from checks, and follow-ups.
|
|
|
|
## Mandatory Guardrails
|
|
|
|
- Always pass explicit `--model` in every `droid exec` command.
|
|
- Do not create unnecessary markdown files; include this instruction in every dispatched prompt.
|
|
- Only create markdown files when explicitly requested by the user or required by the approved plan.
|
|
- Provide rich context to non-explorer droids to avoid repeated repo exploration.
|
|
- Do not use `--skip-permissions-unsafe` unless explicitly authorized.
|
|
- Do not push, deploy, or run destructive actions without explicit user approval.
|
|
|
|
## Resume and Recovery
|
|
|
|
- If interrupted, continue with `droid exec -s <session-id> --model <model-id> "continue previous task"`.
|
|
- If model IDs drift, rerun `droid exec --help` and update all commands.
|
|
- If preconditions fail (permissions, missing tools, missing context), report blocker and propose the next safest step.
|
|
|
|
## Verification Before Completion
|
|
|
|
- Confirm each applicable stage was executed or explicitly skipped with reason.
|
|
- Provide command evidence for quality/review/run checks.
|
|
- Confirm spec approval occurred before codegen.
|
|
- Confirm markdown-file guardrail was applied in dispatched prompts.
|