Agent pipeline plugins — collaborative robots, human merges.
cobots turns any GitHub repo into an assembly line for AI coding agents:
pm → architect → dev → reviewer ∥ writer → gated merge, with all pipeline
state living on the forge (labels, milestones, PRs) so any harness that can
drive gh can play a role. Industrial cobots
are robots built to share a workspace with humans — same idea here: the agents
do the line work, a human owns the merge button and everything the gates escalate.
/plugin marketplace add tibtof/cobots
/plugin install cobots-core@cobots
/plugin install cobots-java@cobots # if it's a Java project
Then, inside the repo you want on the pipeline:
/cobots-core:pipeline-setup
That detects your build system, writes .claude/pipeline.md (the per-project
config and opt-in marker), creates the six status:* labels and a first
milestone, and recommends the LSP plugin for your language.
Run an issue through the pipeline:
/cobots-core:issue-pipeline 42 # tier chosen by triage
/cobots-core:issue-pipeline 42 light # force light tier
- State machine on the forge (spec):
six
status:*labels are the machine truth and the work-claim lock; comments carry the human narrative; milestones are the roadmap, and a milestone close triggers theauditskill. - Two tiers:
light(dev → reviewer — bug fixes, docs, chores, small projects) andfull(adds PM spec, architect ADR, writer doc review — anything ADR-worthy). Scale effort to complexity. - Deterministic gates: a stop hook blocks the dev agent until the project's
verify:command passes; a PostToolUse hook syncs status labels on PR creation; the merge gate refuses >50-file diffs,--admin, and unresolved threads. All hooks no-op in repos without.claude/pipeline.md. - Structured handoffs: stages exchange schema-bounded payloads (design summary, PR number, verdict, doc flags) — never transcripts.
- Docs stay honest: dev and reviewer flag affected doc surfaces; the writer agent applies exactly those flags on the PR branch.
Two files in each consuming repo, with a strict split:
| file | contains |
|---|---|
CLAUDE.md |
the constitution: locked decisions, architecture and style rules (template) |
.claude/pipeline.md |
operational config: tier, verify command, changelog/ADR paths, language (template) |
cobots defaults apply where the constitution is silent: hexagonal architecture, functional style (immutability, errors as values), conventional commits.
One table, aliases only — retune here when a new model generation ships:
| role | model | why |
|---|---|---|
| pm, architect, reviewer | opus |
judgment-heavy gates |
| dev, writer | sonnet |
implementation throughput |
| triage, mechanical steps | haiku |
cheap and fast |
| milestone audit | fable |
strongest model for the periodic deep look |
A coarse, pipeline-shaped allowlist for consuming projects (instead of accreting
one-off entries): gh issue *, gh pr *, gh api *, gh label *, git *,
your build tool (./gradlew * / go * / npm * / make *).
core/ cobots-core plugin: agents, skills, workflow, hooks, references
lang/java/ cobots-java overlay: idiom skill with JBang-compiled examples
lang/{go,ts,kotlin}/ planned overlays
evals/ structural lint + trigger checks (CI-enforced)
The pipeline content is battle-tested across several private projects and generalized here; the packaging and authoring style borrow deliberately from alexandru/skills (terse skills, compiled examples) and addyosmani/agent-skills (marketplace shape, eval discipline, orchestration constitution).
License: MIT