agents-orchestration-pipeline
A semi-autonomous multi-agent pipeline that takes a development task from raw request to merged pull request: an orchestrator supervises per-task pipelines of first-class agents (refiner, designer, implementer, code reviewer, documenter, documentation reviewer, submitter, PR shepherd), each in its own git worktree, with configurable human gating and a decision journal.
The design document lives at docs/agent-pipeline-design.md.
The staged implementation plan (build order, per-step verification, and technical design for each
step) lives at docs/implementation/. Implementation is underway;
Step 1 (plugin scaffold and the
configuration model) is complete — see config/, lib/resolve_config.py, and tests/.
Step 2 (orchestrator core: routing, state,
worktrees) is also complete — see agents/orchestrator.md, config/transition_table.yaml,
lib/{graph_validate,worktree,state}.py, commands/, skills/, and fixtures/.
Step 3 (refiner + designer, gates G1/G2, the
decision journal, and the autonomy gradient) is also complete — see agents/{refiner,designer}.md,
lib/journal.py, commands/decisions.md, skills/decisions/, and fixtures/tasks/.
Step 4 (implementer + inner green loop) is also
complete — see agents/implementer.md, lib/checks.py, and fixtures/sample-project/.
Step 5 (code reviewer + rework loops & loop
budgets) is also complete — see agents/code_reviewer.md and lib/loop_budget.py.
Step 6 (documenter + documentation
reviewer, gates G5/G6, the L3 docs-rework loop) is also complete — see agents/documenter.md and
agents/documentation_reviewer.md.
Step 7 (submitter + permissions/sandboxing) is also
complete — see agents/submitter.md and hooks/sandbox_guard.py. The real linear spine now runs
end-to-end from the refiner through a live PR (G7).
Step 8 (PR shepherd) is also complete — see
agents/pr_shepherd.md and fixtures/pr-events/. Every node now has a real agent; the
stub-agent harness (fixtures/stub_agent.md + fixtures/stub-outcomes/) remains only for
deterministic routing tests, per agents/orchestrator.md's "Spawning a node".
Step 9 (resource budgets + model selection) is
also complete — see lib/budget.py, hooks/budget_meter.py, and agents/orchestrator.md's
"Resource budgets and model selection".
Step 10 (ticketing integration: none/github_issues/
jira) is also complete — see lib/ticketing/, tests/test_ticketing.py, and the "Ticketing"
sections of agents/orchestrator.md, agents/submitter.md, and agents/pr_shepherd.md.
Step 11 (init skill & packaging
finalization) is also complete — see skills/init/, commands/init.md, reference/ (the
bundled design doc), and tests/test_init.py. This closes out the plan: all 9 agents, the 4
skills/commands, the hooks, and the config schema/defaults ship as one versioned plugin.
Installing
This repository is both the plugin and its own plugin marketplace
(.claude-plugin/marketplace.json). From within Claude Code:
/plugin marketplace add sylmarien/agents-orchestration-pipeline
/plugin install agent-pipeline@agent-pipeline-marketplace
Optionally run /pipeline:init first to generate a project .agents/pipeline.yaml (see
skills/init/SKILL.md) — this is facultative, so you can skip straight to running a task.
Then invoke it with /pipeline:run <task description> (see commands/run.md).