self-improve
claudeTurns verified corrections and hard-won workflows into durable, reviewable Claude instructions and skills, applied only with explicit user authorization.
4 skills · hooks
General-purpose agent catalog: git, pull requests, review, CI, formatting, container escalation, and an IWE knowledge-graph planning workflow.
Plugin agentdev from marketplace agent-devcontainer in plume-works/agent-devcontainer. Commands read from the repo's manifest — 2 days ago.
code-review-standardsApply review standards and write pull request descriptions. Use when asked to review code, assess a change for correctness or maintainability, or draft a PR description. Commit creation belongs to /agentdev:git-commit and PR publishing belongs to /agentdev:pr-open.
create-agentCreate or update repository subagents shared by Claude Code and OpenCode. Use when adding an agent in the agentdev plugin catalog or defining its delegation scope, tools, and instructions.
create-skillCreate, update, or review repository skills in the agentdev plugin catalog with concise discovery descriptions, progressive disclosure, and validation. Use when adding or refining a SKILL.md, diagnosing discovery, or packaging a repeatable agent workflow.
devcontainer-sandboxRun build, test, or lint commands through this repository's devcontainer when the host lacks the project toolchain. Use for devcontainer up or exec, Docker-backed sandbox testing, and local toolchain escalation.
exec-summaryTranslate whatever is in focus — the current thread, a plan, bug, feature, or PR — into a one-minute, non-technical executive summary. Use when asked for an executive view, a plain-language summary, a "1 minute version", or a non-technical explanation of the work. Technical PR descriptions belong to
extract-github-actions-logsExtract logs from a GitHub Actions run or job with the GitHub CLI. Use when asked to fetch failing CI logs, inspect a GitHub Actions run, or pull logs from a run or job URL. Keywords: github actions logs, failing ci logs, workflow run logs, job logs, gh run view.
get-codeql-dataFetch CodeQL code scanning data with the GitHub CLI. Use when asked to get CodeQL alerts, inspect code scanning findings for a pull request, list open security findings, or retrieve CodeQL alert metadata from GitHub. Keywords: CodeQL, code scanning, security alerts, gh api, pull request alerts, code
git-commitGenerate conventional commit messages automatically. Use when user runs git commit, stages changes, or asks for commit message help. Analyzes git diff to create clear, descriptive conventional commit messages. Triggers on git commit, staged changes, commit message requests.
git-merge-resolveMerge a Git ref into the current branch and resolve merge conflicts with confidence-based escalation. Use when asked to merge branches or refs, finish a conflicted merge, resolve Git conflicts, or preserve both sides of divergent changes. Keywords: git merge, merge branch, merge conflict, resolve co
git-new-branchCreate a new Git work branch at the freshly fetched remote base, push it at once so it tracks its own upstream, and carry uncommitted changes onto it — optionally in a separate worktree. Use when asked to start a feature branch, create or cut a branch, get off main before working, or when another sk
iwe-adversarial-reviewLook for ways to defeat an IWE plan's design or its implemented code, from a fresh context — report-only. Use only when the user explicitly asks for an adversarial review of a named plan, or when a plan task being implemented runs `/agentdev:iwe-adversarial-review <plan> <plan|code>`; never while pl
iwe-auditAudit documents and code comments for session residue — provenance, defended alternatives, dated tool behavior — and prune it or restate it as a verified durable fact. Use when cleaning up docs, pruning stale rationale, or reviewing comments for noise. Proving a rewrite preserved meaning belongs to
iwe-bootstrapInstall a blank IWE project-memory workspace — `.iwe/` plus `docs/knowledge/` with its schemas, starter graph, knowledge instructions, and license — into a repository root from assets shipped with the installed catalog, never overwriting existing knowledge. Use when a repository has no IWE workspace
iwe-branch-nameName the next work branch for an IWE plan — iwe/<slug>/iter-<N>-plan, -impl, or -ship, one iteration counter across every phase — counting every iteration already used locally, on the remote, and in pull requests of any state, and read the plan's state in the checkout to pick the phase. Use when sta
iwe-captureFile a finished inbox item in the IWE graph — a bug, a proposed feature, or a backlog task — after checking for a duplicate and refusing any document that misses a section or field SCHEMA.md requires. Use when Explore settles a defect, feature, or task, when Implement finds a defect or work its plan
iwe-exploreInvestigate problems, compare approaches, and clarify ideas with the user before planning. Read code and the graph. Do not write code. Use when the user says "let's think about ...", "explore <idea>", or "what are our options for ...". Also use when the user wants to discuss an idea before a decisio
iwe-implementImplement a named IWE plan task-by-task with state discipline — verified anchors, tests before checkbox ticks, clean stopping points, deviations written back into the plan — then verify the result and act on the findings. Use only when the request names the plan to implement; a request to build some
iwe-plan-reviewReview one IWE plan draft from three fresh contexts at once — the Durable Knowledge Auditor for durable-knowledge residue, iwe-plan-reviewer for the plan rules, and iwe-ship-reviewer in plan mode for what Ship will need — and return the reports unchanged; report-only. Use when the iwe-planner rulebo
iwe-planCreate or revise implementation planning state — discovery in the real codebase first, then a coherent plan with verified code anchors, spec impact, and dependencies. Never edits implementation code.
iwe-setupBrownfield-first workspace onboarding — scan the codebase and draft data/product.md plus starting spec and guide docs from what the code shows, interview the developer only for what the code can't answer, mark the onboarding tasks done, and delete the example documents. Use when the user says "run t
iwe-shipClose or cancel planned work safely — normal shipping requires a zero-CRITICAL Verify report, merges verified behavior into durable specs, and records idempotent graph transitions; cancellation bypasses implementation and release transitions. Also cuts a release when asked.
iwe-verifyVerify that implementation matches the graph — check a plan's tasks, spec requirements, and scenarios against the actual code before shipping, or sweep the whole workspace for doc↔code drift.
iwe-weeklyCompose a weekly project digest from the workspace — what shipped, active plans and their staleness, open bugs, top backlog, and graph health, ending with the single most leveraged next action. Read-only; prints without writing files.
pr-eval-review-neededDecide whether work just pushed to an open pull request needs a fresh AI review, because the AI reviewer's newest verdict is a pending change request or the work went beyond what a reviewer already saw. Use after implementing, refactoring, or fixing anything on a branch with an open PR that has alre
pr-feedback-resolutionSystematic workflow for resolving pull request feedback, review threads, CI failures, CodeQL findings, and Codecov patch coverage gaps. Use when addressing PR review comments, skipping resolved threads, classifying reviewer intent, fixing failing checks, or preparing a PR for re-review. Keywords: PR
pr-gen-descriptionGenerate comprehensive pull request description following /agentdev:code-review-standards with change analysis, testing strategy, and migration notes. Use when creating a PR, writing PR description, preparing for code review, or documenting technical decisions.
pr-merge-stackMerge a chain of manual stacked pull requests from its top: merge the bottom pull request through pr-merge, restack the rest with one pr-restack-all run, and repeat until the top has merged. Use when asked to merge a PR stack or chain, or a dependent pull request that has unmerged pull requests belo
pr-mergeMerge an open GitHub pull request: monitor CI and AI review, resolve conflicts and remediate failures and feedback, then squash merge it explicitly once every requirement passes. Use when a PR needs to be merged or shepherded through CI and review to merge. Keywords: merge PR, squash merge, monitor
pr-openCreate a GitHub pull request from conversation context — or refresh the branch existing PR in place — with accurate title/body generation, branch sync, and remote push. Use when asked to open/create/submit a PR, draft a pull request, sync or update a PR description, or finalize changes after impleme
pr-request-ai-reviewRequest a fresh AI review of an open pull request by adding the request-ai-review label, then confirm a run actually picked it up. Use when asked to request, re-request, or trigger an AI, Claude, or bot review of a PR, to ask for a re-review after new work landed, or to check whether a requested rev
pr-restack-allRestack a chain of manual stacked pull requests with merges after a lower one changed or squash-merged: merge an open predecessor's head, or reconnect a merged predecessor through its final head and squash commit, bottom to top, from a detached worktree, with plain pushes only. Use after a pull requ
pr-reviewPerform a thorough automated code review of a GitHub pull request, publishing feedback as a single GitHub pull request review with inline comments (a standalone comment only as fallback when inline posting fails). Use when asked to review a pull request, or when a PR is opened/reopened and an automa
pr-syncRefresh an existing GitHub pull request title and body so they match the current branch, then push the branch. Use when asked to sync, update, refresh, or regenerate a PR description or title for the branch in the working tree, including after new commits land on it
semantic-refactor-auditProve that a behavior-preserving change actually preserved behavior, by capturing a ground-truth baseline before editing and comparing it afterwards. Use when a rewrite must not change meaning and no test covers it — rewording instructions or documentation, replacing links with prose, migrating conf
skill-scriptsWrite and document a skill's bundled scripts so their outcome is readable by both a shell caller and an agent — a shared exit-code vocabulary, a RESULT= enum line on stdout, and the shared result_codes.py helpers. Use when adding or editing a Python script under a skill's agent-code/ directory, choo
tddRun the Red, Green, Refactor cycle in this session — one failing test, the minimal code to pass it, then cleanup with tests green. Use only when TDD or Red-Green-Refactor is asked for by name, or when the Principal Engineer agent directs it. Never test text files, only real executable code logic.
template-consumeAdopt or update the agent-devcontainer template in a consuming repository — a first-time setup walkthrough (full-copy or existing-repo merge) and an update mode that diffs the tracked template paths since the last-consumed commit. Use when asked to adopt this devcontainer/catalog in another project,
update-branchUpdate the current Git feature branch: fetch, bring in the commits its own upstream gained on the remote, then merge the remote base branch. Use when asked to sync a feature branch, bring a branch up to date with `origin/main`, refresh a stale PR branch, or unblock CI after base-branch drift. Keywor
Durable Knowledge AuditorAudit a document for durable-knowledge residue from a fresh context that never saw the work that produced it. Use to gate a plan or document edit before validation — reads and reports only, changes nothing. Dispatch it with the file to audit and the audit scope; it invokes /agentdev:iwe-audit and re
iwe-adversarial-reviewerLook for ways to defeat an IWE plan's design or its implemented code from a fresh context — report-only. Dispatched only by the iwe-adversarial-review skill; never on a description match.
iwe-implementerExecute an active IWE plan task by task — verified anchors, tests before checkbox ticks, evidence per task, deviations routed back to the plan. Followed in place by the iwe-implement skill and never dispatched; never on a description match.
iwe-plan-reviewerCheck one IWE plan draft against the plan rules file from a fresh context, returning Conforms or Gaps — report-only. Dispatched only by the iwe-plan-review skill; never on a description match.
iwe-plannerCreate or revise an IWE plan — discovery in the real codebase first, then a coherent plan with verified code anchors, spec impact, and dependencies; never edits implementation code. Followed in place by the iwe-plan skill and, in revise mode only, by the Implement session applying a revision; never
iwe-revision-evaluatorJudge one uncommitted IWE plan revision against the plan at HEAD from a fresh context, returning APPROVE, REJECT, or ESCALATE — report-only. Dispatched only by the Implement session after it drafts a revision; never on a description match.
iwe-ship-reviewerCheck an IWE plan against Ship's obligations from a fresh context — plan mode reports whether the plan gives Ship what it needs; ship mode checks a shipped plan and stamps it with a Ship review checklist or returns its gaps. Dispatched only by the iwe-ship skill, iwe-plan revise after a ship, and th
iwe-shipperShip, cancel, fix Ship review gaps in, or cut a release for an IWE plan from a fresh context — Verify gate, spec merge, idempotent graph transitions. Dispatched only by the iwe-ship skill and iwe-plan revise answering a Ship blocker report or Ship review gaps; never on a description match.
iwe-verifierVerify an IWE plan's tasks, spec requirements, and scenarios against the code, or audit the workspace for doc↔code drift, from a fresh context — report-only. Dispatched only by the iwe-verify skill; never on a description match.
Principal EngineerPrincipal-level engineering guidance for architecture, executable-code implementation and review, with TDD for runtime behavior. Use for engineering decisions and production-code features or bug fixes, not routine docs, config, or workflow edits.
Turns verified corrections and hard-won workflows into durable, reviewable Claude instructions and skills, applied only with explicit user authorization.
4 skills · hooks
This page updates itself from your repo. Add the badge to your README, or get it Featured to pin it above the organic results with a followed link to your site.