8000 GitHub - pvginkel/AIWorkflow · GitHub
[go: up one dir, main page]

Skip to content

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

AIWorkflow — the dev plugin marketplace + workshop

This repo is a Claude Code plugin marketplace. It hosts the slice-development workflow as an installable plugin (dev), and it is the workshop where that workflow is measured and improved.

The workflow used to be a template you copied into each project and filled in. It is now a plugin you install once into ~/.claude; each project describes itself through its .kubecoder/project.yaml manifest (read by kc) and an .aiworkflowrc at its root. This version targets the KubeCoder environment — it is kc-native and expects kc on PATH (always true inside a KubeCoder pod).

The dev plugin

The slice pipeline: /dev:triage → /dev:plan-slice → /dev:run-slice.

  • Triage is the operator's own chat skill. It files a batch of cards and findings as slices, the questions first and the slices as their consequence.
  • Planning and the run are stages. Pilot, the project manager, runs each as a work item. The skill puts every ask it has into a document Pilot shows the operator, ends its turn, and carries the answers out when Pilot writes them back (pilot.md). The close-out is the run's continuation, not a step of its own. Without Pilot, the skill prints each document's path and the operator answers in the file.
  • Beside them, /dev:onboard brings a repo onto the pipeline, /dev:inbox sorts every project's tracker inbox with the operator ahead of triage, and /dev:arch-design writes a design with its options.
  • Two skills are no workflow. /dev:document-format is the format of a document Pilot reads, and /dev:writing is how everything the operator reads is written.

plan-slice has the plan-preparer ground the slice and put its requirements asks, then drives a plan-writer/plan-reviewer round (plan_loop.py) to a reviewed phase queue; run-slice launches a kc-native run loop (run_loop.py) that takes each phase through a bounded loop — fetch → branch → code-writer → test gate + test-fixer → consult-funded review rounds against a rising bar → ff-merge → ✅ DONE — then closes the slice out with a loop-tail gate sweep, a completion consult, a test phase and a doc phase, spawning every agent as a headless kc session. Both loops run inside the stage's turn and stop at a question with a document for the stage to register. The gate is kc project test and lint, run by the loop itself: detecting green needs no model, only fixing red does. Files are durable; sessions are ephemeral; scripts drive, agents judge.

The plan is the queue. One plan.md per slice holds phases as ### P<id> — <title> headings, each opening with a Target: line naming a kc project list component or a sibling repo. Document order is authoritative; only the driver stamps a phase done. Every agent in the loop may edit the plan — appending a phase is how work grows, bounded by a generation bar (work the plan owes first, blocking work only after, then the operator). Everything out of the loops' scope goes in one close-out report per slice (close-out.json, rendered at the run's end as close-out.md, the document Pilot reads, and close-out-record.md; created at planning, written by every agent as it goes): nothing from a run is carded per finding. The author of an entry labels what it is, a table in the tool routes it, the operator answers what comes to them in Pilot and the run's continuation carries the answers out.

  • plugins/dev/ — the plugin: 8 skills, 11 agents, the tools (run_loop.py, plan_loop.py, asks.py, close_out.py, checks.py, github_target.py, close_slice.py, slice_cards.py, slice_cost.py, turn_profile.py, preflight.py, project_config.py, redact_secrets.py, spec_check.py, long_run.py, allocate-next-slice.sh, with their suites), and the contract docs (pilot.md / run-loop.md / runner-state.md / plan-loop.md / plan-template.md / agent-dispatch.md / close-out.md / close-out-template.md, plus project-contract.md, preflight.md, spec-tree.md).
  • .claude-plugin/marketplace.json — makes this repo installable.

kc project test and kc project lint gate the plugin from this repo — the suites are the reason the workflow is safe to change here.

Install

/plugin marketplace add <this repo>     # e.g. /plugin marketplace add /work/AIWorkflow
/plugin install dev@aiworkflow

Then make a repo adoptable — author a .kubecoder/project.yaml and an .aiworkflowrc (where the spec repo, the procedure docs and the optional phases live). Preflight enforces the whole contract and tells a new repo exactly what is missing. See docs/ADOPTING.md.

The workshop

  • plugins/dev/tools/slice_cost.py — what a slice cost, priced from the run's own state.json / plan_state.json records and the transcripts they name, and — through turn_profile.py, which replays each transcript turn by turn — what its turns did, per role. Both ship with the plugin rather than living here, because they read a state format the plugin owns and every run writes its own numbers.
  • workflow-improvements/ — the R&D / evidence trail behind the workflow's design: the July 2026 cost investigation and the plan it produced, frozen as written. It describes the task runner that preceded the plugin, not the loops that run today; the story since is docs/rationale/history.md.
  • archive/quality/ — the retired quality capability (quality-improver, quality-issue-finder, refactor-audit) and the code_health grader, one folder per source project because the copies drifted apart. Parked while the tool is rebuilt; /dev:onboard sweeps each project's copies in here. See archive/quality/README.md.

Docs

  • docs/ADOPTING.md — install the plugin and make a repo adoptable (manifest + CLAUDE.md entries), with a worked example.
  • docs/AUTHORING.md — the durable rules for writing/maintaining agents, skills, and docs so they stay lean and drift-free.
  • docs/rationale/ — the explanation layer: why the pipeline is shaped the way it is, with the incidents and numbers behind each rule — the flow in one read, the design principles, the history and reversals, the imp 5831 rovement catalogue, how runs are measured, how the research literature was used, the close-out report's case, and how planning came to its documents.
  • plugins/dev/docs/ — the plugin's own contract: pilot.md (a stage as a Pilot work item: the turns, the brief, the documents, the continuation), run-loop.md (the phase queue and the consult/test/doc ladder) with runner-state.md, plan-loop.md, plan-template.md, agent-dispatch.md, and close-out.md / close-out-template.md (the per-slice report every agent writes its out-of-scope observations to) beside it, plus project-contract.md, preflight.md and spec-tree.md (the rule for every commit into the shared spec tree and for citing a line).
  • Fieldnotes — a cross-project observation store beside the workflow, in a repo of its own: agents post what they now leave in close-out reports, and a scheduled reconciler curates it into a triage doc. It is developed in a KubeCoder environment of its own, with its design note and slices in pvginkel/FieldnotesAppSpecs.
  • CHANGELOG-workflow.md — the plugin's changelog.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0