agent-plugins

Practical methods for AI-assisted work: diagnose a bug, plan a change, research a claim or hand work over with evidence someone else can use.

Start with the task. You can ask for a method explicitly and inspect how it works when you need more control. Shared scope, authority and evidence rules have one owner; a plugin supplies the detailed method.

Try a first task · Choose by need · Installation · Follow the guides

Try it

Open a disposable clone in Codex:

git clone https://github.com/JRichlen/agent-plugins.git agent-plugins-first-task
cd agent-plugins-first-task

Ask it to read AGENTS.md and plugins/find-before-build/skills/find-before-build/SKILL.md, then:

Before adding another gallery capture helper, find the existing one. Show the searches and what can be reused. Do not edit files or run models.

The first-task guide shows actual search results from source-loaded Codex, how to check and undo the exercise, and host limits. No plugin or global instruction is installed by this route.

Prefer a Claude session trial?

From the same clone, with a Claude version that supports the flag:

claude --plugin-dir ./plugins/find-before-build

This loads one local plugin for that session. Native Claude use was not tested in the recorded Codex demonstration. See installation choices or give your agent the guided prompt.

See what happens

flowchart LR
  Task[Describe the outcome] --> Method[Select the smallest useful method]
  Method --> Work[Do the work within authority]
  Work --> Check[Check the result]
  Check --> Return[Return evidence and remaining limits]

Loading

For example, the capture-launch fix follows a real scheduled-run failure through its source, regression controls and remaining limits. The small-feature guide shows source candidate packaging. The handoff guide explains what another session needs to continue from the actual commits.

Start with the result, then open the explanation, evidence or customization links as needed. Required permissions, material costs and verification limits stay visible even when the explanation is short.

How the pieces fit together

One system with clear responsibilities

Layer	Responsibility	Source
Core	Intent, scope, authority, shared task state and method-selection boundaries	Operating contract and ownership index
Shared services	Bounded delegation, continuity and reusable verification operations	Delegate work, continue or hand off, check evidence
Playbooks	The task procedure for planning, research, debugging, implementation or instruction maintenance	Choose by need
Techniques	A focused approach within a task, such as a retained end-to-end slice, comparison or review perspective	Tracer bullets and the selected playbook's references
Host adapters	Reviewable projections of the core for a host's discovery, tools and instructions	Adapters

Evidence, knowledge and decisions connect these layers through existing task, project and decision records. They are not another coordinator or an autonomous self-rewriting subsystem. Each method should finish or hand off without leaving its entire procedure active for the next stage.

Choose by need

What you need	Start here
Clarify a consequential plan	grill-me
Map work across sessions	wayfinder
Diagnose a failure	diagnosing-bugs
Build one useful end-to-end slice	tracer-bullets
Choose an interface	codebase-design
Find existing code before adding more	find-before-build
Test a contested claim	wave-research
Delegate independent work	delegated-intelligence
Check completion evidence or use an explicit acceptance contract	verify-before-claim
Resume or hand off	context-handoff
Audit stale instruction files	docs-hygiene
Write or revise prose	Ordinary authoring, with voice when useful
Learn from completed work	dev-diary and recurrence-detector

Choose the depth you want

Get the result: describe the outcome and constraints; receive the result, essential limits and any decision you need to make.

Understand the work: ask why a method was selected, what changed, which alternatives were considered or where uncertainty remains.

Inspect the evidence: follow actual sources, artifacts, diffs, test results and worker handbacks.

Customize the system: inspect the core, activation/exit contracts, host adapters and checks before changing behavior.

These choices apply per task. Explanation length does not determine verification depth or require a heavier workflow.

How methods enter, finish and hand off

Core distinguishes an explicit request, task match, dependency, state/risk check, always-on invariant and offer-only capability. Selection never grants authority.

A playbook owns its task procedure. Shared services return evidence and gaps to their caller. The ownership index records service entry and exit conditions. A perspective changes what is inspected; it does not gain control over the whole task.

Use the existing task record when tracking is needed. On a stop, account for workers and in-flight actions; some actions may be non-cancellable or have an uncertain outcome. Report that state rather than implying rollback.

How decisions and knowledge improve without collecting duplicate rules

Record an observation with its source and scope. Compare it with prior evidence. Revise the affected decision or method, check the change and retain a restore path. Recurrence-detector proposes candidates from repeated sourced observations; it does not silently install a lesson as universal policy.

Prefer improving or retiring an existing responsibility before adding another. Keep temporary task state distinct from established knowledge and behavioral instructions. Reuse existing records and the compiler where they fit.

What verification can and cannot establish

Claim checking ties a statement to checks actually run and preserves uncertainty. Its optional verify-work method adds executable acceptance criteria and independent verification when that contract is selected. It is not a mandatory lifecycle for every task.

Structural checks, model evaluations and real-host runs provide different evidence. A skipped job is not a completed evaluation. A file or report does not by itself prove the result correct. A sequential self-review is not independent verification.

See the testing inventory, evaluation instructions and example gallery. Some gallery entries are ungraded seeds; inspect their provenance.

Voice's second-opinion requires a direct request or accepted offer and a real worker tool. It cannot simulate dispatched validation when that tool is absent.

Keep one method or inspect another host

Use the installation chooser for source loading, a session trial, Codex project skills, native marketplaces and APM. It separates package availability from discovery, tools, hooks and actual host qualification. The guided-install prompts help your agent inspect existing installs, choose scope, verify a read-only invocation and preserve undo.

Installing a method does not adopt Core globally or supply worker tools. See the host adapters when you need deeper integration.

Current status and migration

The catalog has 24 registered plugins. Redgate and Jori are retired public products; useful mechanisms have one owner in the existing verification and delegation packages. The fleet curator is retired without a replacement for its scheduled cross-repository service.

Existing caller	Current path	Compatibility boundary
Jori coordination	delegated-intelligence	Replace old global activation fragments explicitly; installation alone does not activate Core.
Redgate acceptance helpers	verify-before-claim / verify-work	Three old script paths remain thin adapters; existing .redgate/ records remain readable.
Fleet curator deployment	No new marketplace installation	Existing copied repositories, schedules and credentials are unchanged and uninspected; inventory them before any separate cleanup.

Keep migration notices and adapters only while a known caller needs them. Remove them when caller inventories confirm migration and active source-path checks have moved to current entry points. Repository retirement does not uninstall copies, edit global instructions or stop external jobs. To roll back, restore source, catalog, callers and generated outputs together from the pre-migration revision; restoring only an unlisted plugin directory is incomplete.

Qualification is still in progress. Focused structural checks are not proof of behavioral improvement or complete host parity. Inspect current checks and their limits; installed host behavior and general outcome improvement remain separately qualified. The frozen PR129 experiment remains separate from the synthetic re-entry pilot.

Full catalog

Browse the 24 registered plugins

Plugin	Purpose
graveyard	Archive repositories as restorable bundles and prepare guarded deletion scripts.
plugin-factory	Scaffold a plugin with manifests, entry points and checks that start red.
tailscale-wif	Set up secretless GitHub Actions authentication to Tailscale.
dev-diary	Discuss what mattered in the day's work and record or revise a journal entry.
voice	Improve prose and machine-facing output; includes explicitly requested second-opinion review.
grill-me	Examine a plan through questions scaled to uncertainty and stakes.
tracer-bullets	Build one real end-to-end slice and expand it.
semver-gate	Classify action risk and rehearse restoration before irreversible changes.
verify-before-claim	Check completion evidence; optionally establish and verify executable acceptance criteria.
diagnosing-bugs	Test possible causes, make a focused fix and demonstrate regression coverage.
docs-hygiene	Audit instruction files against current repository behavior and resolve contradictions.
context-handoff	Choose how to continue at a phase boundary and pass source references onward.
codebase-design	Compare materially different interface designs before choosing one.
wayfinder	Map a larger effort into decisions and dependent work across sessions.
scope-fence	Keep diffs tied to the agreed task and record unrelated findings.
find-before-build	Assess existing implementations before introducing another one.
egress-gate	Inspect outbound content and its destination before transmission.
stop-rule	Bound repeated attempts and return an evidence-backed stop report.
recurrence-detector	Propose improvements from repeated sourced failure patterns.
agent-compiler	Compose versioned behaviors into reproducible definitions with provenance and effect limits.
eval-ladder	Choose proportionate checks and describe their coverage gaps.
consolidation-map	Account for every entity in a consolidation and reconcile its disposition.
wave-research	Examine a contested claim through rival explanations and sourced adversarial review.
delegated-intelligence	Brief real workers, own outputs, integrate evidence and recover only invalid work.

Maintaining the system

Start with AGENTS.md. Change the canonical owner of a responsibility, update its callers, and run the required checks. core/sync.py generates the packaged contract copies; do not edit them independently.

The documentation site is generated from committed sources. Keep current usage, ownership and evidence guidance in the tree; use commits and PR discussion to retrieve prior designs and decisions.

License

MIT. See LICENSE.