From agent-sequencer
A mechanism for step-driving AI agents according to a classical program (sequencer program)
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-sequencer:agent-sequencerThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
This skill is an MCP server that implements an architecture where "a classical program drives an AI agent."
README.mdREADME_ja.mddocs/authoring-programs.mddocs/authoring-programs_ja.mdprograms/README.mdprograms/README_ja.mdprograms/hello.pyprograms/review_rounds.pyprograms/review_rounds/README.mdprograms/review_rounds/README_ja.mdprograms/review_rounds/agents/prompt-sensei.mdprograms/review_rounds/agents/python-sensei.mdprograms/review_rounds/agents/sequencer-sensei.mdprograms/review_rounds/scripts/fetch-diff.shprograms/review_rounds/scripts/render-review.pyprograms/review_rounds/scripts/rm-tmp.shprograms/review_rounds/skills/sequencer-review-resolve.mdprograms/review_rounds/skills/sequencer-review-respond.mdprograms/review_rounds/skills/sequencer-review.mdThis skill is an MCP server that implements an architecture where "a classical program drives an AI agent." Decision logic is contained within the program, and the agent acts purely as a driver.
hello — Minimal sample / smoke-test program. Greets each name in params["names"] (default ["world"]) one at a time.review-rounds — Self-review helper for sequencer programs. Three specialists (python-sensei / sequencer-sensei / prompt-sensei) review → respond → verify, iterating up to N rounds until convergence.For how to write new programs, see docs/authoring-programs.md.
sequencer_list_programs to identify the appropriate program.sequencer_start, then memorize the returned instance_id as the highest priority.last_yield.text (using your own tools).expect_schema and submit it via sequencer_next(for_step_no=<current value>, result=...).state is awaiting_result, repeat steps 3 and 4.sequencer_current.state becomes completed / aborted / failed, report the final result to the user and call sequencer_close.sequencer_current must not be used as a basis for deciding what to do next. Decisions must always follow the last_yield instruction itself.~/.claude/sequencer/state/<id>.jsonl or the directory specified by the AGENT_SEQUENCER_STATE_DIR environment variable) using Read. Do not stream history into context via MCP.sequencer_list_programs — List available programs.sequencer_start — Start an instance.sequencer_current — Re-fetch the most recent yield (for re-sync).sequencer_next — Submit result and obtain the next yield.sequencer_resume — Restore from JSONL.sequencer_close — Release (recommended path).sequencer_list — List active instances.For details including setup and usage, see README.md.
Guides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub opensphere-inc/agent-sequencer --plugin agent-sequencer