coding-loop plugin
The evidence-gated coding loop: orient (map), build (feature), fix (bugfix), improve (refactor), gate (done), and run it all on one yes (autopilot)
How to install the coding-loop plugin
Run both commands inside Claude Code. The first adds the marketplace once; the second installs the plugin.
/plugin marketplace add duthaho/skillhub
/plugin install coding-loop@skillhubWhat's inside coding-loop
8 parts read from the plugin's own files.
Skills 8
Claude loads it when the task matches- autopilot
The whole coding loop on one deliberate yes: map, spec, critic-hardened plan, GitHub issues, parallel worktree implementers, fresh-context review loops, one PR with the evidence. Acts only on its itemized charter; parks what it can't settle honestly. Commands: /autopilot <task> [--merge], resume, status. It signs a charter and pushes branches, so it never auto-triggers — explicit invocation only.
- blueprint
Human-facing documentation, architecture docs, and mermaid diagrams for a codebase, a module, or a feature — "draw me the architecture, with receipts." Fans out read-only sub-agents to build a grounded component graph first: every node is a real path, every edge backed by a grep-verified import or call — unverifiable edges are dropped, not drawn. Diagrams render from that verified table per a fixed checklist (never freehand mermaid), capped at 5–8 top-level components with drill-down pages; every doc section cites its sources as file:line. Drafts in chat, writes docs/blueprint/ only on your yes, and stamps a manifest with date + commit so /blueprint refresh regenerates only what the diff touched. Use for "generate an architecture diagram", "document this codebase or module for people", "create a mermaid diagram of this feature" — e.g. "/blueprint", "/blueprint the payment flow", "/blueprint refresh". For agent-facing orientation (AGENTS.md) use map; blueprint is the human-facing sibling.
- bugfix
Lightweight bug-fixing loop — reproduce → root-cause → fix test-first → verify — deliberately separate from the heavyweight feature workflow so small fixes don't drag a spec-and-plan process behind them. Demands a deterministic reproduction before any code changes, a stated root-cause hypothesis confirmed with evidence (never a shotgun fix), a failing test that captures the bug, and the original repro re-run as proof. Keeps a one-line-per-bug log in out/dev/bugfix-log.md and checks it first, so recurring bugs get recognized instead of re-diagnosed. Use when the user reports something broken, failing, throwing, or flaky, or asks to fix a bug — e.g. "/bugfix login 500s on empty password", "fix this error", "this test is flaky", "why is X crashing". If the fix turns out to need real design work, hand off to feature; the shipping gate is done.
- done
End-of-session shipping gate — "prove it works, then ship it." Runs an evidence checklist (full test suite, lint/typecheck, build, and actually running the change — output quoted, never asserted), then a fresh-context two-stage review by a sub-agent that sees only the diff and the spec: spec-compliance first, correctness second, style never — joined by a cross-model second opinion when a codex/gemini CLI is installed. Flags fake-green tripwires (edited or deleted tests, new TODOs or skips, hardcoded expectations) and ends with a verdict — SHIP, FIX FIRST, or NEEDS HUMAN — before drafting the commit/PR for the user's approval. Use at the end of any coding session or when the user asks "is this done", "ready to ship?", "/done", "wrap this up", "review and commit this". It is the closing step of both feature and bugfix, but works standalone on any uncommitted or branch diff.
- feature
End-to-end feature workflow around the agent — spec by interview, a plan of 2–5-minute tasks, test-first execution with checkpoint commits, then an evidence-gated finish. Effort-scaled: a one-sentence diff skips straight to implementation; a real feature gets spec.md → plan.md → task-by-task TDD, with all state in out/dev/<change>/ so any later session resumes at the first unchecked task. Human-in-the-loop: the spec and the plan are approved (and editable) before any code is written, and nothing ships without the done gate. Use when the user wants to build, add, or implement a feature or start non-trivial coding work — e.g. "/feature add rate limiting", "implement X", "build a Y that Z", "resume the <change> feature". For fixing something broken use bugfix; for just the end-of-session shipping gate use done; for this whole loop end-to-end without approval stops, suggest the user run /autopilot (explicit invocation only).
- map
Codebase orientation for an unfamiliar or large repo — "map this before I work in it." Fans out read-only Explore sub-agents that navigate the code (keyless — grep/glob/read, no embeddings to go stale) along the order that matters: entry points, build/test/lint commands, architecture and key abstractions, conventions and gotchas, test layout. Distills the findings into a lean AGENTS.md at the repo root — the cross-tool standard file many agents auto-load — where every claim points to a real file and nothing survives that isn't load-bearing. Human-in-the-loop: it drafts, you approve, and it never writes or commits into your repo without an explicit yes. Stamps the map with date + commit so /map refresh regenerates only what went stale. Use before working in a repo you or the agent don't know — e.g. "/map", "/map the auth subsystem", "generate an AGENTS.md", "/map refresh". Run it upstream of feature or bugfix; for the end-of-session shipping gate use done.
- newskill
Scaffold a new skillhub skill, or repair one that drifted, across all four registration surfaces in one motion — folder, README table row, marketplace plugin, and trigger evals — then prove it with validate-skills.py. User-invoked.
- refactor
Evidence-driven refactoring — where does structural improvement actually pay off, then make that change without changing behavior. Survey mode diagnoses by evidence, not aesthetics — git churn × complexity hotspots, the recurring-bug log, upcoming work — and verdicts each candidate REFACTOR/PREP/LEAVE/WATCH, where LEAVE is the default and "the textbook recommends it" is never a reason. Execute mode takes one approved candidate in small behavior-preserving moves, suite green after each. Keeps a debt register in out/refactor/tracker.md so repeat surveys open with what changed. Use for "clean up this module", "improve the architecture", "pay down tech debt" — e.g. "/refactor", "/refactor <target>". For new behavior use feature; for something broken use bugfix.
Using coding-loop in Claude Code
Skills need no command. Claude reads them when your request matches what they describe.
Update, turn off or remove
- Update the plugin (shell)
- claude plugin update coding-loop@skillhub
- Refresh the marketplace
- /plugin marketplace update skillhub
- Turn it off for now
- /plugin disable coding-loop@skillhub
- Remove it
- /plugin uninstall coding-loop@skillhub
Skills in coding-loop
- blueprintduthaho/skillhub
Human-facing documentation, architecture docs, and mermaid diagrams for a codebase, a module, or a feature — "draw me the architecture, with receipts." Fans out read-only sub-agents to build a grounded component graph first: every node is a real path, every edge backed by a grep-verified import or call — unverifiable edges are dropped, not drawn. Diagrams render from that verified table per a fixed checklist (never freehand mermaid), capped at 5–8 top-level components with drill-down pages; every doc section cites its sources as file:line. Drafts in chat, writes docs/blueprint/ only on your yes, and stamps a manifest with date + commit so /blueprint refresh regenerates only what the diff touched. Use for "generate an architecture diagram", "document this codebase or module for people", "create a mermaid diagram of this feature" — e.g. "/blueprint", "/blueprint the payment flow", "/blueprint refresh". For agent-facing orientation (AGENTS.md) use map; blueprint is the human-facing sibling.
3 files9stars - refactorduthaho/skillhub
Evidence-driven refactoring — where does structural improvement actually pay off, then make that change without changing behavior. Survey mode diagnoses by evidence, not aesthetics — git churn × complexity hotspots, the recurring-bug log, upcoming work — and verdicts each candidate REFACTOR/PREP/LEAVE/WATCH, where LEAVE is the default and "the textbook recommends it" is never a reason. Execute mode takes one approved candidate in small behavior-preserving moves, suite green after each. Keeps a debt register in out/refactor/tracker.md so repeat surveys open with what changed. Use for "clean up this module", "improve the architecture", "pay down tech debt" — e.g. "/refactor", "/refactor <target>". For new behavior use feature; for something broken use bugfix.
2 files9stars - autopilotduthaho/skillhub
The whole coding loop on one deliberate yes: map, spec, critic-hardened plan, GitHub issues, parallel worktree implementers, fresh-context review loops, one PR with the evidence. Acts only on its itemized charter; parks what it can't settle honestly. Commands: /autopilot <task> [--merge], resume, status. It signs a charter and pushes branches, so it never auto-triggers — explicit invocation only.
9stars - bugfixduthaho/skillhub
Lightweight bug-fixing loop — reproduce → root-cause → fix test-first → verify — deliberately separate from the heavyweight feature workflow so small fixes don't drag a spec-and-plan process behind them. Demands a deterministic reproduction before any code changes, a stated root-cause hypothesis confirmed with evidence (never a shotgun fix), a failing test that captures the bug, and the original repro re-run as proof. Keeps a one-line-per-bug log in out/dev/bugfix-log.md and checks it first, so recurring bugs get recognized instead of re-diagnosed. Use when the user reports something broken, failing, throwing, or flaky, or asks to fix a bug — e.g. "/bugfix login 500s on empty password", "fix this error", "this test is flaky", "why is X crashing". If the fix turns out to need real design work, hand off to feature; the shipping gate is done.
9stars - doneduthaho/skillhub
End-of-session shipping gate — "prove it works, then ship it." Runs an evidence checklist (full test suite, lint/typecheck, build, and actually running the change — output quoted, never asserted), then a fresh-context two-stage review by a sub-agent that sees only the diff and the spec: spec-compliance first, correctness second, style never — joined by a cross-model second opinion when a codex/gemini CLI is installed. Flags fake-green tripwires (edited or deleted tests, new TODOs or skips, hardcoded expectations) and ends with a verdict — SHIP, FIX FIRST, or NEEDS HUMAN — before drafting the commit/PR for the user's approval. Use at the end of any coding session or when the user asks "is this done", "ready to ship?", "/done", "wrap this up", "review and commit this". It is the closing step of both feature and bugfix, but works standalone on any uncommitted or branch diff.
9stars - featureduthaho/skillhub
End-to-end feature workflow around the agent — spec by interview, a plan of 2–5-minute tasks, test-first execution with checkpoint commits, then an evidence-gated finish. Effort-scaled: a one-sentence diff skips straight to implementation; a real feature gets spec.md → plan.md → task-by-task TDD, with all state in out/dev/<change>/ so any later session resumes at the first unchecked task. Human-in-the-loop: the spec and the plan are approved (and editable) before any code is written, and nothing ships without the done gate. Use when the user wants to build, add, or implement a feature or start non-trivial coding work — e.g. "/feature add rate limiting", "implement X", "build a Y that Z", "resume the <change> feature". For fixing something broken use bugfix; for just the end-of-session shipping gate use done; for this whole loop end-to-end without approval stops, suggest the user run /autopilot (explicit invocation only).
9stars - mapduthaho/skillhub
Codebase orientation for an unfamiliar or large repo — "map this before I work in it." Fans out read-only Explore sub-agents that navigate the code (keyless — grep/glob/read, no embeddings to go stale) along the order that matters: entry points, build/test/lint commands, architecture and key abstractions, conventions and gotchas, test layout. Distills the findings into a lean AGENTS.md at the repo root — the cross-tool standard file many agents auto-load — where every claim points to a real file and nothing survives that isn't load-bearing. Human-in-the-loop: it drafts, you approve, and it never writes or commits into your repo without an explicit yes. Stamps the map with date + commit so /map refresh regenerates only what went stale. Use before working in a repo you or the agent don't know — e.g. "/map", "/map the auth subsystem", "generate an AGENTS.md", "/map refresh". Run it upstream of feature or bugfix; for the end-of-session shipping gate use done.
9stars - newskillduthaho/skillhub
Scaffold a new skillhub skill, or repair one that drifted, across all four registration surfaces in one motion — folder, README table row, marketplace plugin, and trigger evals — then prove it with validate-skills.py. User-invoked.
9stars
coding-loop plugin questions
How do I install the coding-loop plugin?
In Claude Code run /plugin marketplace add duthaho/skillhub, then /plugin install coding-loop@skillhub.
What does the coding-loop plugin include?
coding-loop bundles 8 skills.
How do I update or remove the coding-loop plugin?
Update it with claude plugin update coding-loop@skillhub in your shell, or from the Installed tab of /plugin. Remove it with /plugin uninstall coding-loop@skillhub.