CLAUDE.md@plugins/wise · git:20260906.8b0eb38 · 2026-09-06 · sha256 2cfb9578dd290d50
CLAUDE.md@plugins/wise git:20260906.8b0eb38A
Immutable. This exact content is served forever at /api/v1/blob/2cfb9578dd290d50.
# CLAUDE.md — operational context for the `wise` plugin
Loaded automatically when a developer opens this plugin in Claude Code.
This file records **invariants** — things that must be true whenever the
plugin is at rest. Procedures (how to add an action, how to test) live
in the root [`CONTRIBUTING.md`](../../CONTRIBUTING.md); do not duplicate
them here.
**Before making any non-trivial change, read `CONTRIBUTING.md`.** It
documents the decisions that led to the current shape of the plugin —
most "obvious improvements" have already been considered and rejected
for reasons that are listed there.
---
## What this plugin is
This directory is the **`wise` plugin for Claude Code** — the
canonical, hand-edited source. Edit files here directly and validate
with `just check`.
`wise` is a standalone copilot plugin in the `wise-claude`
marketplace: a workflow engine, shared scripts, and a set of
tech-neutral action skills (`/wise-init`, `/wise-workflow-*`,
`/wise-skills-*`, `/wise-pr-*`, `/wise-commit-*`) plus the
`wise-estimation`, `wise-markitdown`, and `wise-code-comments`
reference skills, the
`wise-human-writing` and `wise-tickets` hybrid skills (auto-consulted
references + `/wise-human-writing` rewrite and `/wise-tickets`
restructure commands), and the
`/wise` natural-language helper.
Every user-facing skill is a **flat, autocomplete-visible slash
command**. No dispatcher-style routing: typing `/wise-` fans out in
Claude Code's slash menu to every action. The plugin also exposes the
`/wise` bare command as a **natural-language helper** — typing `/wise`
alone prints the catalog of every command, and typing `/wise <free-form
text>` (e.g. `/wise open a PR`) classifies the request via LLM
judgement, proposes the matching `/wise-*` command, and offers to run
it through `AskUserQuestion`.
Current actions (all standalone):
- `/wise-init` — first-time dep-install wizard; walks the user through
bun or Node 24 (the engine runtime), the `claude` CLI login, gh,
markitdown and Python (v1 scripts, until plan M3.4), caches results
for the workflow skills' fast-path.
- `/wise-skills-create` — scaffold a new action skill via `skill-creator`.
- `/wise-skills-edit` — modify an existing action skill via `skill-creator`.
- `/wise-workflow-list` — list bundled + user workflow definitions.
- `/wise-workflow-create` — wizard to scaffold a new user workflow.
- `/wise-workflow-run` - start a run on the wise engine: pre-flight
questions, run context, one line per event, gates (this conversation
is a thin conductor over the `wise_*` MCP tools).
- `/wise-workflow-resume` - resume a paused or failed engine run, or
answer a gated one, then follow it.
- `/wise-workflow-status` - list engine runs, or show one run and its
open gate.
- `/wise-workflow-remove` — delete a user workflow definition.
- `/wise-feedback` — file a feedback issue against the marketplace repo.
- `/wise-profile` — set the session's token-budget profile
(`low|medium|max`, default `medium` = the standard behavior). Stored
per session; profile-sensitive skills (`wise-pr-watch-auto`)
read it via `references/profile-read.md` and
degrade silently to `medium`. Workflows never read it: the engine's
pre-flight asks harness, model and effort per tuning group instead.
Budget only — model tiers, optional-step scope, panel size, retry
caps; NEVER correctness rules.
- `/wise-fork` — reorient a forked session. Inherited context becomes
read-only background; every in-flight task, plan, and promise from
before the fork is dropped (the original session owns them);
remembered file state is distrusted (the original session may touch
the same tree in parallel) and a fresh git baseline is snapshotted.
Optionally starts straight on a new goal passed as the argument.
- `/wise-report` — verified session status report. Recalls claims from
context/memory, verifies each against git, `gh` PR/CI state, workflow
runs, and files on disk, then emits a compact ref-coded report
(F/A/Q/R/L/P) where every line carries an evidence tag or is marked
`(unverified)`. `--full` expands detail; `--save` writes a handoff
file that survives compaction. Read-only apart from that file.
- `/wise-insights-mine` — the self-improvement loop (the "harvest"
pass). Mines the local insights ledger (fed by the SessionEnd hook)
for recurring task patterns, frequency-gates them, and drafts the
strongest ones into user-global `~/.claude/skills/` after
per-candidate approval.
- `/wise-insights-refine` — the "garden" pass. Enumerates the learned
skills, finds overlapping ones, and (with approval) merges them into
one aggregated skill and retires the originals — reversibly (backed
up first). Acts only on wise-managed skills (marker-tagged); never
deletes hand-written ones.
- `/wise-insights-reset` — reversible cleanup + rollback. Snapshots
then removes the auto-created skills and/or the insights index into
`snapshots/<ts>/`, and restores any snapshot. Only managed skills;
never hard-deletes (that's `insights.py purge`).
- `/wise-commit-message`, `/wise-commit`, `/wise-commit-push`,
`/wise-pr-create`, `/wise-pr-add-reviewers`, `/wise-pr-watch` —
standalone PR / git helpers. The three commit skills are a graded
trio: `/wise-commit-message` is read-only (drafts and hands back),
`/wise-commit` drafts + commits locally, `/wise-commit-push` drafts
+ commits + pushes.
- `/wise-pr-create-auto`, `/wise-pr-request-review-auto`,
`/wise-pr-watch-auto`, `/wise-implement-plan-auto`,
`/wise-simplify-auto` — the autonomous (`-auto`) building blocks:
decision-free, `AskUserQuestion`-free variants of the PR / implement /
quality steps, each a thin reader of a shared fragment or reference.
`/wise-simplify-auto` (the lightweight per-commit tier — the
`code-simplifier` agent) and the `code-review` workflow (the
heavyweight branch gate — three reviewer children, a curator, an
optional verifier and a fixer, each with its own tuning group) are
the two quality passes; the `ticket-auto` workflow's engine-side
review phase follows the same discipline.
- `/wise-supervise` — attach a watchdog / supervisor loop to a running
team of background agents and keep them on task: probe each member,
nudge the idle-but-unfinished or off-goal ones, escalate the
persistently stuck (the automation of manually typing "ping all your
subagents, are you on track?"). Reads the shared
`references/supervise-loop.md`, as does the `-auto` implement phase
(`SUPERVISE=yes`). The v2 engine has its own stale policy
(`stale_after`: nudge, then kill) and does not read it.
- `/wise-revise` — the proactive planner: investigates a scope (folder /
component / whole project) against a free-form improvement intent,
read-only, via a panel of roster lenses; ranks findings by leverage and
writes self-contained `PLAN-*.md` plans + an index into `docs/plans/`
for the user to execute later. Writes only under `docs/plans/`; never
edits source and never runs a plan (execution is delegated). Reads its
own skill-local `references/audit-lenses.md` + `references/plan-format.md`.
- `/wise-grill` — the subject-understanding pass: classifies its input
(tracker ticket, doc link, free-form prompt, or question), then
deep-researches it across every reachable source (tracker comments +
screenshots, wiki, Slack, Drive, design, codebase + git history),
gap-checks the evidence, and forks by type — a ready
`docs/plans/PLAN-<ref>.md`; a `BLUEPRINT-<ref>.md` with targeted
questions (per-person for tickets / docs, asked inline when the user
is the one who can answer — prompts); or a researched
`ANSWER-<ref>.md` for a pure question (re-run with answers to
upgrade a blueprint). Facts get researched; only decisions get
asked. Writes only under `docs/plans/`; read-only against every
external system. Reads the shared `references/grill/*` routines —
the same ones the `ticket-plan` / `ticket-auto` workflows run in
their plan phases (ticket subjects only there).
- `/wise` — the natural-language helper (bare = catalog; with free-form
text = intent classifier).
Two model-invoked document-authoring skills round out the plugin —
they carry no `argument-hint`, so they auto-trigger on matching prose
rather than being typed as a flat command:
- `wise-prd-architect` — drives a structured multi-phase process for
writing Product Requirements Documents; auto-triggers on "PRD",
"product spec", "feature spec", and similar.
- `wise-trd-architect` — the engineering counterpart; auto-triggers on
"TRD", "technical design", "architecture document", "system design".
Each ships its own `agents/` and `references/` files *inside its
skill directory* (the skill spawns those agents via `Task` and reads
the reference templates). These skill-local agents are distinct from the
**plugin-level agent roster** under `agents/` (see below) — the
skill-local ones are private to one skill; the roster is shared across
every workflow.
The plugin also ships a **plugin-level `agents/` roster** — a set of
SDLC role subagents (`wise:ceo`, `wise:cto`, `wise:architect`,
`wise:software-engineer`, `wise:qa-engineer`, `wise:security-engineer`,
`wise:devops-engineer`, `wise:sre`, `wise:code-reviewer`, …), catalogued
in `AGENTS.md`. They are real Claude Code plugin subagents (invocable as
`subagent_type: wise:<name>`); a Claude child the workflow engine spawns
reaches them through its own `Task` / `Agent` tool when the step prompt
asks for a role.
---
## Layout
```
plugins/wise/
├── .claude-plugin/plugin.json # manifest (no `dependencies:` — see CONTRIBUTING §2.3)
├── .mcp.json # bundled MCP servers: `wise-engine` (bash engine/engine.sh mcp)
├── engine/ # the workflow engine: TypeScript run as source on bun or Node 24
│ ├── engine.sh # entry: bun, else node >= 24, execs src/cli.ts
│ ├── package.json # deps (@modelcontextprotocol/sdk, yaml, zod); `npm run check`
│ ├── src/ # defs (YAML v2 schema), scheduler, executor, ledger, daemon, mcp,
│ │ │ # unit-mcp, channel, resolve, render, preflight, migrate, auth
│ │ ├── adapters/ # claude, codex, grok (gemini landing) + clean-env spawn
│ │ ├── steps/ # agent, bash, gate step runners
│ │ ├── phases/ # unit pipeline phases (claim, worktree, model phases, push, pr, ...)
│ │ └── prompts/units/ # the model-phase prompt templates + schemas
│ └── test/ # node --test / bun test suite
├── hooks/ # the ONE sanctioned hook (see CONTRIBUTING §2.4)
│ ├── hooks.json # auto-discovered; registers the SessionEnd hook
│ └── session-end-ingest.sh # SessionEnd → insights.py ingest (stdlib-only, exit 0, no LLM)
├── agents/ # plugin-level SDLC agent roster (auto-discovered; wise:<name>)
│ ├── ceo.md cto.md product-manager.md engineering-manager.md
│ ├── architect.md software-engineer.md qa-engineer.md
│ ├── security-engineer.md devops-engineer.md sre.md
│ └── ux-designer.md technical-writer.md code-reviewer.md
├── CLAUDE.md # this file (invariants)
├── AGENTS.md # catalog/index of the plugin-level agent roster
├── README.md # slim overview + links into /docs/wise/*
├── LICENSE
├── .gitignore # keeps .wise-init-registry.yaml out of source control
├── .wise-init-registry.yaml # RUNTIME ONLY — written by `/wise-init`, wiped on reinstall
├── scripts/
│ ├── engine.sh # thin bash bootstrap → execs engine.py (skill catalog only; not the workflow engine)
│ ├── engine.py # skill-catalog emitter (`list-skills` subcommand) — consumed by the /wise helper
│ ├── bootstrap-deps.sh # full dep probe (python3 + pyyaml/ulid/typing_extensions, bun or node ≥24, gh + auth); cold-start fallback
│ ├── init.sh # bash-only per-dep probes used by `/wise-init` (works before Python is installed)
│ ├── init-registry.py # YAML I/O for .wise-init-registry.yaml + fast-path `check` for the workflow skills
│ ├── workflows.py # LEGACY v1 workflow scripts: list/create/remove skills, profile store, legacy conductor; deleted in plan M3.4
│ └── insights.py # self-improvement engine: ingest/mine/gate sessions → skill candidates (STDLIB ONLY)
├── workflows/ # bundled workflow definitions (shipped defaults)
│ └── <name>/ # folder form: workflow.yaml + sibling artifacts
│ ├── workflow.yaml # the definition (YAML v2, validated by engine.sh compile-check)
│ ├── README.md # summary, flow, steps, inputs, outputs (kept in sync with the YAML)
│ ├── templates/ # optional — workflow-shipped templates; addressable as {{workflow.dir}}/templates/…
│ └── prompts/ # optional — e.g. watch-pipelines-auto.md; addressable as {{workflow.dir}}/prompts/…
├── references/ # cross-skill shared prose (addressed as ${CLAUDE_PLUGIN_ROOT}/references/<file>.md)
│ ├── subject-drafting.md # Conventional-Commits scope / type / subject rules
│ ├── branch-naming.md # the ticket = branch-name rule
│ ├── init-check.md # shared init-registry fast-path protocol
│ ├── profile-read.md # session token-budget profile read (silent degrade to medium); read by profile-sensitive skills
│ ├── dispatch.md # the --on routine: run a skill's procedure as a headless child of any harness (engine.sh models + dispatch); read by wise-pr-watch(-auto), the pr/simplify/implement -auto skills
│ ├── simplify-pass.md # canonical per-commit simplify pass (code-simplifier agent)
│ ├── code-review-pass.md # canonical high-depth branch review (reviewer-subagent panel)
│ ├── report-pass.md # canonical verified status report (recall → verify → emit; read by /wise-report + the ticket-auto / impl-plan-auto report steps)
│ ├── supervise-loop.md # the watchdog routine (idle/hung detection → nudge → escalate); read by the -auto implement phase + /wise-supervise
│ ├── legacy-conductor/ # the v1 prose conductor (run/resume/status/preflight/step-types/roster); followed only for v1 definitions and state.yaml runs until M3.4
│ ├── insights-init-guard.md # /wise-init gate read by wise-insights-mine / -refine
│ ├── grill/ # the subject-understanding routines (context sweep + gap analysis + blueprint schema) — read by /wise-grill (any subject), ticket-plan, ticket-auto (tickets)
│ └── pr/ # shared PR/commit fragments (draft-body, ensure-pr, watch-pipelines, handle-*, commit-from-fix, paged-bulk-mode, comment-surfaces, sonar-fetch) + templates/pr-template.md — read by the wise-pr-* skills + ticket-auto
└── skills/
├── wise/SKILL.md # natural-language helper (bare catalog + intent classifier)
├── wise-init/SKILL.md # dep-install wizard
├── wise-skills-create/SKILL.md
├── wise-skills-edit/SKILL.md
├── wise-workflow-list/SKILL.md
├── wise-workflow-create/SKILL.md
├── wise-workflow-run/SKILL.md # the conductor
├── wise-workflow-resume/SKILL.md
├── wise-workflow-status/SKILL.md
├── wise-workflow-remove/SKILL.md
├── wise-prd-architect/ # model-invoked PRD authoring (SKILL.md + agents/ + references/)
├── wise-trd-architect/ # model-invoked TRD authoring (SKILL.md + agents/ + references/)
├── wise-feedback/SKILL.md # file a feedback issue
├── wise-profile/SKILL.md # session token-budget profile (low|medium|max; budget only, never correctness)
├── wise-fork/SKILL.md # reorient a forked session (context = background; pre-fork work dropped)
├── wise-report/SKILL.md # verified session status report (ref-coded, evidence-tagged; --full / --save)
├── wise-insights-mine/SKILL.md # self-improvement loop: mine sessions → draft skills
├── wise-insights-refine/SKILL.md # consolidate learned skills: merge overlaps → retire originals
├── wise-insights-reset/SKILL.md # reversible cleanup + rollback (snapshot → clear → restore)
├── wise-commit-message/SKILL.md # Conventional-Commits drafter (read-only)
├── wise-commit/ # draft + commit (no push)
│ ├── SKILL.md
│ └── commit-routine.md # shared draft + commit (+ optional push) procedure
├── wise-commit-push/SKILL.md # draft + commit + push; reads wise-commit/commit-routine.md
├── wise-estimation/SKILL.md # reference skill (Fibonacci SP scale)
├── wise-markitdown/SKILL.md # reference skill (file → markdown text extraction via markitdown)
├── wise-code-comments/SKILL.md # reference skill (code-comment rules: ELI5, concise, present-tense, no history)
├── wise-human-writing/SKILL.md # hybrid: human-first rules for all outbound tracker/PR/doc/chat writing + /wise-human-writing rewrite
├── wise-tickets/SKILL.md # hybrid: ticket structure/scoping/breakdown rules for any tracker + /wise-tickets restructure
├── wise-pr-create/SKILL.md # create or refresh a PR
├── wise-pr-add-reviewers/SKILL.md # attach Copilot + extras
├── wise-pr-watch/SKILL.md # drive pipelines + comments to green
├── wise-pr-create-auto/SKILL.md # autonomous PR create (no prompts)
├── wise-pr-request-review-auto/SKILL.md # autonomous Copilot attach (no prompts)
├── wise-pr-watch-auto/SKILL.md # autonomous CI watch + fix loop (no prompts)
├── wise-implement-plan-auto/ # autonomously implement a PLAN-*.md
│ ├── SKILL.md
│ └── agents/executor.md # fresh-context per-task executor persona
├── wise-simplify-auto/SKILL.md # autonomous simplify + commit (no prompts)
├── wise-supervise/SKILL.md # attach the watchdog loop to a running team of background agents
├── wise-revise/ # proactive planner: audit a scope → executable PLAN-*.md backlog
│ ├── SKILL.md
│ └── references/ # audit-lenses.md (the panel) + plan-format.md (the plan + index schema)
└── wise-grill/SKILL.md # subject-understanding pass (ticket / doc / prompt / question): multi-source research → PLAN, BLUEPRINT-with-questions, or ANSWER
```
No `commands/` directory is present, and one must not be added without
the discussion called for in `CONTRIBUTING.md`
[§2](../../CONTRIBUTING.md#2-conventions-that-apply-to-every-plugin). An
`agents/` directory **IS** present — the plugin-level SDLC role roster
(`AGENTS.md` + `agents/*.md`), auto-discovered by Claude Code and
addressable as `subagent_type: wise:<name>`. It is a deliberate part of
the plugin's design (Claude children spawned by the workflow engine
delegate to it); adding or editing a role follows the procedure in
`AGENTS.md`. A
`hooks/` directory IS present, holding **exactly one** sanctioned hook —
the SessionEnd insights-ingest hook (`hooks/session-end-ingest.sh` +
`hooks/hooks.json`). It is the single documented exception to the
no-hooks default; its rationale and hard constraints live in
`CONTRIBUTING.md` [§2.4](../../CONTRIBUTING.md#24-hooks). No other hook
(and no `SessionStart` hook) may be added without that same discussion.
`.mcp.json` IS present — it bundles the MCP servers wise skills depend
on: today the `wise-engine` server (`bash
${CLAUDE_PLUGIN_ROOT}/engine/engine.sh mcp`, tool timeout 660 s), a thin
client that starts the `wise-engined` daemon on demand. See the
bundled-tooling convention in `CONTRIBUTING.md`
[§2.2](../../CONTRIBUTING.md#22-bundled-tooling-convention). An
`engine/` directory IS present: the TypeScript workflow engine, run as
source (no build step; `cd engine && npm run check` for typecheck, lint,
format and tests).
---
## Invariants
Keep these true. Each one has a rationale in `CONTRIBUTING.md`; the
one-liners below are the rule, not the argument for it.
- **Two skill shapes — standalone and reference.** Every skill's
frontmatter lands in one of two buckets. Pick the right one when you
author; mixing fields breaks discovery.
- **Standalone slash-command skills** (the default shape for every
action, e.g. `wise-workflow-run`, `wise-commit-message`) —
user-invocable, shown in the slash menu as `/wise:<skill-name>`
with a bare `/<skill-name>` alias when unambiguous. Frontmatter
sets `argument-hint:` (may be an empty string). No `command:` /
`subcommand:` / `user-invocable: false` / `arguments:` fields —
those were v1 dispatcher-routing fields with no meaning in v2.
The skill body reads `$ARGUMENTS` as a raw string and parses its
own positionals.
- **Reference / guidance skills** (e.g. `wise-estimation`) —
description-triggered docs. Frontmatter has NO `argument-hint:`
and no action-oriented fields. Claude auto-consults them when the
user's prose matches the `description:`. Body is reference
content, not action logic.
- **Exactly one `disable-model-invocation: true` skill — `wise`.** It's
the natural-language helper. Users invoke it explicitly; Claude
never auto-runs it.
- **The `/wise` helper classifies intent; it does not act directly.**
Action logic lives in the action skills. The helper's job is to
(a) print the catalog when called bare, or (b) pick the best
matching `/wise-*` command and offer to invoke it via the `Skill`
tool after `AskUserQuestion` confirmation.
- **`scripts/engine.py` is a catalog emitter, nothing more.** Its
only supported subcommand is `list-skills`, which walks `skills/`
and emits a JSON document the `/wise` helper consumes. No routing,
no fuzzy matching, no argument parsing. `engine.sh` stays as the
bash bootstrap; every skill that calls `engine.sh` grants the bash
path in its `allowed-tools`.
- **Action skills never invoke other action skills.** The `/wise`
helper is the only place that calls the `Skill` tool on a
wise-namespaced action skill (and only after user confirmation).
Action-to-action work-sharing happens through `scripts/` helpers.
The exception for `wise-workflow-run` and `wise-workflow-resume`
(which compose other skills as workflow steps) is below.
- **All persistent state lives in `${CLAUDE_PLUGIN_DATA}`.** Never
write elsewhere — with these narrow exceptions:
(a) the init registry, see below;
(b) workflow run state, which is per-workspace by design and lives
under `~/.local/share/wise/runs/<cwd-slug>/` (off-tree, off
`.claude/**`, never auto-cleaned); and
(c) the **insights store** under `~/.local/share/wise/insights/`
(`ledger.jsonl` + `candidates.json` + `decisions.json` +
`skill-backups/<ts>/<name>/` + `snapshots/<ts>/{index,skills}/`), the
self-improvement loop's per-user state. `decisions.json` records
`promoted` / `dismissed` / `retired` (the last set by
`/wise-insights-refine` when it merges a skill away; `mine` resurrects
it if the merged skill is later deleted). `snapshots/` holds
`/wise-insights-reset` restore points (reversible cleanup); `purge
--yes` is the only irreversible wipe; and
(d) the **report handoff store** under
`~/.local/share/wise/reports/<cwd-slug>/` — output of the report
pass (`/wise-report --save` today; any `SAVE=yes` caller of
`references/report-pass.md`), a per-workspace sibling of the runs
tree (same slug, same XDG rules), kept off-tree for the same
reasons as run state; and
(e) the **session profile store** under
`~/.local/share/wise/profile/<session-id>` — one word
(`low|medium|max`) written atomically by `/wise-profile`
(`workflows.py profile-set`), read via
`references/profile-read.md` / `profile-get` with silent degradation
to `medium` (the engine keeps `engine/src/profile.ts` for the data
root only; workflows run at `medium`), and
GC'd opportunistically (files from sessions older than 30 days) on
each write. Routed through `wise_data_root()` (engine: `paths.ts`).
New per-user persistent state
that doesn't fit `${CLAUDE_PLUGIN_DATA}` MUST route through the
`wise_data_root()` helper in `scripts/workflows.py` — never
hard-code paths so future relocations are one-function changes.
(`insights.py` mirrors that helper with a stdlib-only fallback,
because the SessionEnd hook may run before pyyaml is installed; the
canonical helper is still used whenever importable.) When
`${CLAUDE_PLUGIN_DATA}` is unset (e.g. the scripts are run directly,
outside Claude Code), `plugin_data_root()` falls back to
`$WISE_DATA_DIR` then `wise_data_root()`; under Claude Code the
Claude vars win, unchanged.
- **Init registry — the one file we write inside `${CLAUDE_PLUGIN_ROOT}`.**
`/wise-init` caches probe results at
`${CLAUDE_PLUGIN_ROOT}/.wise-init-registry.yaml`. It lives in the
install dir on purpose: every `/plugin install wise@…` wipes it,
giving natural invalidation. Workflow engine skills read this as a
fast-path via `scripts/init-registry.py check` before falling back
to the full `bootstrap-deps.sh` probe. Two writers: `/wise-init`
(full payload, interactive) and `bootstrap-deps.sh` itself, which
auto-populates a successful-probe subset on its way out. `.gitignore`
keeps the registry out of source control.
- **No persisted project registry.** The `{{project.*}}` template
variables a workflow run operates on are derived from the current
context - the engine's `detectProject` derives them from the run
`cwd` (`project-selection: current`). There is no `projects.yaml`
and no skill that writes one.
- **`allowed-tools` in each skill is narrowly scoped.** Expanding it
should be a deliberate decision, not an incidental fix-up.
- **`model` / `effort` frontmatter follows the work, not the skill.**
Lightweight, mechanical, or read-only skills (the commit-drafting
trio, `wise-workflow-list` / `-status` / `-remove`, `wise-feedback`)
pin `model: opus` + `effort: low` for snappy turnaround. Skills that
do real reasoning or orchestration (`wise-pr-watch`, the workflow
conductor / resume, the wizards, the PRD/TRD architects) omit both
and inherit the session model — `effort: low` would hurt them. Set
the knobs to match the skill's cognitive load.
- **The agent roster is plugin-level; the engine has no roster field.**
The `agents/*.md` roster files are real Claude Code plugin subagents -
frontmatter is limited to `name` / `description` / `tools` / `model` /
`effort` / `color`; plugin subagents **ignore** `hooks` / `mcpServers` /
`permissionMode`, so never add those. Roster `model:` is `inherit`;
`effort:` is the role's default reasoning level. In YAML v2 an `agent`
step is one headless harness child (`claude -p`, `codex exec`,
`grok -p`, `gemini -p`) spawned by the engine under the user's own
login, with the step's `harness` / `model` / `effort` (or its tuning
group) passed as real CLI flags; a Claude child delegates to a roster
role through its own `Task` / `Agent` tool when the prompt says so.
There is no `agent:` / `agents:` field, no teams, no conductor-side
synthesis. Model resolution (retired-id swap, capability clamp,
policy ceiling; the low-profile Opus rule stays dormant because
workflows run at `medium`) is `engine/src/resolve.ts`; the model
catalog pre-flight offers is `engine/src/models.ts`. Keep
`AGENTS.md`'s catalog table in sync with `agents/*.md`, the same way
workflow READMEs stay in sync with YAML.
- **The roster is canonical; never hand-maintain a divergent copy.**
`agents/*.md` is the single source. The repo-root `AGENTS.md` and
`plugins/wise/AGENTS.md` document it as project-*instructions* (not
loadable registries); their roster tables mirror `agents/*.md` and are
kept in sync the same way workflow READMEs track their YAML.
- **Cross-skill shared prose lives in `plugins/wise/references/`.** A
rule or routine read by more than one skill — the Conventional-Commits
`subject-drafting.md` (read by the commit routine, `wise-commit-message`,
and `draft-body.md`), the workflow `init-check.md` (read by
`wise-workflow-run` / `-resume` / `-list` / `-status`), and the two
quality passes `simplify-pass.md` (read by the commit routine, the
implement phase, and `wise-simplify-auto`) and `code-review-pass.md`
(the discipline the `code-review` workflow's prompts and the PR
watcher's review fallback follow), the
verified status report `report-pass.md` (read by `/wise-report`;
parameterized by `SCOPE` / `MODE` / `SAVE` so every caller runs the
identical routine), the watchdog routine `supervise-loop.md` (read
by the `-auto` implement phase and `wise-supervise`), the
`references/grill/` subject-understanding routines
(`research-sources.md` + `gap-analysis.md` + `blueprint-format.md`,
read by `/wise-grill` and the `ticket-plan` / `ticket-auto` plan
phases), and the
`references/pr/` PR/commit fragments (`draft-body.md`, `ensure-pr.md`,
`ensure-reviewers.md`, `propose-reviewers.md`, `watch-pipelines.md`,
the `handle-*.md` queue handlers, `paged-bulk-mode.md`,
the shared `comment-surfaces.md` / `sonar-fetch.md` fetch spines,
`commit-from-fix.md`, read by the `wise-pr-*` skills and the
`ticket-auto` workflow's `prompts/`) - has a single home there, addressed as
`${CLAUDE_PLUGIN_ROOT}/references/<file>.md` and read at run time.
Skill-*local* `references/` (the architects' own templates) stay
inside the skill dir. Change a shared rule in the reference, never in
a copy.
- **Action skill directory name IS the slash command.** A skill at
`skills/wise-workflow-run/SKILL.md` with frontmatter `name:
wise-workflow-run` is invocable as `/wise-workflow-run` (bare
alias) or `/wise:wise-workflow-run` (canonical namespaced). There
is no translation layer. Rename the dir → rename the command; do
both in the same commit, and flag the change as breaking.
- **Workflow run state lives under `~/.local/share/wise/runs/<cwd-slug>/<run-ulid>/`**
(honours `XDG_DATA_HOME`). Per-workspace scoping is preserved via
the `<cwd-slug>` segment. Why off the project tree: Claude Code's
sensitive-path heuristic flags `.claude/**` paths; putting state in
the project tree also adds gitignore management and IDE noise. Why
user-scoped rather than `/tmp`: `/tmp` auto-cleans and breaks the
resume contract. Never under `${CLAUDE_PLUGIN_DATA}`. Workflow
*definitions* are the opposite — they live under a `workflows/` root
(user-authored at `${CLAUDE_PLUGIN_DATA}/workflows/definitions/`,
shipped at `${CLAUDE_PLUGIN_ROOT}/workflows/`) in one of two layouts:
- `<root>/<name>/workflow.yaml` — **folder form, preferred**. The
workflow can ship its own artifacts as siblings (`templates/`,
`prompts/`) and address them from steps via `{{workflow.dir}}`.
- `<root>/<name>.yaml` — **legacy flat form**. Still accepted; no
artifacts dir.
Folder form wins on same-root collision. User root still wins over
bundled root on cross-root collision. See
[`docs/wise/workflows.md`](../../docs/wise/workflows.md) for
the full reference.
- **Exception to "action skills never invoke other action skills" —
the legacy v1 conductor only.** On the v2 engine the conductor skills
(`wise-workflow-run`, `wise-workflow-resume`, `wise-workflow-status`)
call only the `wise_*` MCP tools and the engine CLI; a workflow step
that runs a skill is an `agent` step with `skill: <name>` (the prompt
`Run /<name>` on the `claude` harness), executed by the engine's child,
never by the conductor. The `references/legacy-conductor/` prose,
followed only for `version: 1` definitions and `state.yaml` runs until
plan M3.4, keeps the old narrow exception: it may call `Skill` on a
wise-namespaced action skill as part of a `type: skill` step, never
re-entering the `wise` helper.
- **Workflow README.md stays in sync with `workflow.yaml` +
`prompts/`.** When you touch a bundled workflow's YAML or any of
its `prompts/*.md` fragments, update the workflow's `README.md` in
the SAME commit — Flow mermaid, Steps table, Inputs/Outputs tables,
and Related-links section must reflect the new shape.
- **The unit pipelines are idempotent on resume.** `ticket-auto` and
`impl-plan-auto` are one `units` step; the loop is engine code
(`engine/src/units.ts`, `engine/src/phases/`), not prose. Its `claim`
phase must *ensure* (create, re-attach, or adopt) each unit's worktree
from the per-unit ledger under `<run dir>/units/` plus live `git` /
`gh` probes - never reintroduce a collide-and-fail `git worktree add
-b`. Live state is the source of truth; the ledger is a hint; a
worktree or branch the run did not claim is skipped, never adopted.
`.worktreeinclude` files are carried over once per worktree
(`includes-done` ledger key). Phase prompts live under
`engine/src/prompts/units/`; change a rule there, never in a workflow.
- **All workflow YAML, scheduling and state handling lives in
`engine/`.** `engine/src/defs.ts` owns the v2 schema (with v1 hints),
`scheduler.ts` the DAG, `executor.ts` the run loop, `ledger.ts` the run
directory. SKILL.md bodies call the `wise_*` MCP tools or
`engine/engine.sh`; they never parse YAML or state themselves.
`scripts/workflows.py` is the v1 engine, kept only for
`/wise-workflow-list` / `-create` / `-remove`, the profile store and
the legacy conductor; it is deleted in plan M3.4 and gains no new
behaviour. Validate a definition with `engine.sh compile-check`; the
repo validator runs it on every bundled `version: 2` workflow.
- **External-tool dependencies — bundle the static ones, probe the
open-ended ones.** When a skill needs a third-party tool, prefer
declaring it so the install is one step; but when the *set* of
possible tools is open-ended, probe and propose at run time instead.
Concretely:
- Plugin-to-plugin deps are currently NOT declared in
`.claude-plugin/plugin.json`'s `dependencies: [...]` array — a
marketplace-qualified entry silently breaks plugin loading in the
Claude desktop app (CONTRIBUTING §2.3). Optional plugins (the
`code-simplifier` agent the per-commit simplify pass dispatches)
are documented in the README's Bundled-tooling table instead, and
the consuming skill degrades gracefully when they are absent.
- MCP server deps go in `.mcp.json` (today: `wise-engine`). MCP tool
ids are derived from the plugin name, so moving an MCP between
plugins is a breaking rename.
- CLI / environment deps that neither mechanism can install (Python,
brew packages) are probed at run time by
`scripts/bootstrap-deps.sh` and surface a one-shot install prompt.
- Open-ended deps — where the workflow or skill cannot know up
front *which* tool the user needs — are probed and proposed
dynamically. The `ticket-plan` / `ticket-auto` workflows are
the reference case: they work with any task tracker, so they
detect the tracker, probe for a matching MCP / CLI, and
web-search + propose install options when none is found, rather
than pre-declaring a specific tracker plugin.
See CONTRIBUTING.md [§2.2](../../CONTRIBUTING.md#22-bundled-tooling-convention) for the full convention.
---
## Pointers for common tasks
For the full procedure on each of these, read the linked section of
`CONTRIBUTING.md`:
| Task | See |
|---|---|
| Add a new `/wise-<action>` command | `CONTRIBUTING.md` [§4](../../CONTRIBUTING.md#4-adding-an-action-to-a-plugin) |
| Install the plugin from this clone | `CONTRIBUTING.md` [§6.1](../../CONTRIBUTING.md#61-install-the-plugin-from-a-clone) |
| Validate JSON / bash / naming before commit | `CONTRIBUTING.md` [§6.2](../../CONTRIBUTING.md#62-syntax-and-structural-checks) |
| Commit and PR style | `CONTRIBUTING.md` [§7](../../CONTRIBUTING.md#7-commit-and-pr-conventions) |
| Versioning rules for this plugin | `CONTRIBUTING.md` [§8](../../CONTRIBUTING.md#8-versioning) |
| Add or modify workflow-subsystem behaviour | `CONTRIBUTING.md` [§9](../../CONTRIBUTING.md#9-workflow-subsystem) |
---
## Why `CLAUDE.md`, `README.md`, and `CONTRIBUTING.md` all exist
Three audiences, three stability contracts:
- `README.md` — users landing on GitHub or running `/plugin info`.
Describes what the plugin does and how to use it.
- `CLAUDE.md` (this file) — an agent working inside the plugin tree.
Short list of invariants and a pointer to the full procedures.
- `CONTRIBUTING.md` — the full procedural manual. Single source of
truth for *how* to change things.
Overlap between the three is limited on purpose. If you find the same
rule documented in two of them and they disagree, the code is the truth
and both files are stale — fix both.