If CONDUCTOR is too heavy weight for you try SPAE

[mystilleef]

https://github.com/mystilleef/gemini-cli-spae

Why

Conductor is awesome. But, it's token heavy, pollutes context, and its tightly coupled to Gemini CLI.

I use multiple harnesses, and I needed a simple, efficient and effective workflow to resume work at any point on any harness. Good context hygiene and token efficiency was also paramount.

Enter SPAE.

What SPAE does

SPAE prevents context drift, hallucination, and scope creep by:

• Decomposing requests into atomic, independently verifiable tasks
• Persisting state in external artifacts (not conversational memory)
• Enforcing phase boundaries so each agent reads only what it needs
• Allowing any harness (Gemini, Claude, OpenAI) to resume work at any
  point

Framework design

• Agent-first workflow—combat context pollution and excessive token consumption
  by executing all SPAE phases using sub-agents.
• Harness agnostic—use any combination of Gemini, Claude, OpenAI, or
  Codex at each phase
• Zero-knowledge resumption—any agent resumes any workstream by
  reading the artifacts
• Atomic execution—every task stays small enough to verify in a
  single cycle
• Human oversight—you invoke each phase; nothing progresses
  automatically

Workflow overview

spec -> plan -> review -> build

1. Use Scout agent to create SPEC.md
2. Use Planner agent to create PLAN.md from SPEC.md
3. Use Reviewer agent to optimize and address gaps in PLAN.md
4. Use Builder agent to execute a single task from PLAN.md

The framework creates 3 artifacts in .spae

• SPEC.md: fleshed out requirements.
• PLAN.md: a list of atomic tasks to implement.
• STATE.json: a task tracker.

The SPAE workflow

Run each phase in order. Agents read only their designated inputs and
write only their designated outputs.

Phase 1: Spec

The scout agent distills your request into unambiguous requirements.

/run @scout implement <your feature or request>

Phase 2: Plan

The planner decomposes SPEC.md into a directed acyclic graph (DAG)
of atomic tasks.

/run @planner

Phase 3: Review

The reviewer performs a gap analysis on PLAN.md and optimizes it.

/run @reviewer

Phase 4: Build

Implement tasks from PLAN.md

# builds only one task at a time for PLAN.md
/run @builder 

or

# using tdd, builds only one task at a time for PLAN.md
/run @tdder 

or

# builds all tasks from PLAN.md
/run @executor 

Repeat /run @builder or /run @tdder until PLAN.md marks all tasks
complete.

Other non-spae agents

• Troubleshooter: fix issues.
• Refactor: simplify changes.
• Coverage: address testing gaps.
• Commit: perform autonomous atomic commits.
• Worker: run ad-hoc tasks for better context hygiene.

Conclusion

Gemini CLI is very effective using a structured workflow. You're welcome to take any ideas from the project to experiment with workflows.

https://github.com/mystilleef/gemini-cli-spae