git:20261009.e1a8e27 to git:20261009.55f76ee

---

name: fit-setup

description: "Setup and upgrade Expectation Fit project config in .expectation-fit/ — full wizard for greenfield monoliths and multi-repo workspaces (placement, conventions.auto roots, severity_align, pattern preference), plus upgrade that merges missing capabilities without clobbering notes. Optional calibrate measures repo distributions and proposes thresholds. Stack-aware; idempotent. Use to set up, extend, or upgrade a project's .expectation-fit/ config."

argument-hint: "[all | ways | patterns | thresholds | fresh | upgrade | multi-repo | roots:a,b | calibrate [p90|p75|…]] (blank = wizard)"

---

# Expectation Fit — Init / setup / upgrade

Scaffolds and upgrades project config under `.expectation-fit/` so a team can declare ways of

working (lens toggles, conventions, auto sources, severity align), design-pattern

policy (preferred / allow / block / approved), and architecture thresholds — then

commit them. Once present, `.expectation-fit/` supersedes the plugin defaults

(`${CLAUDE_PLUGIN_ROOT}/references/config-resolution.md`) for every `fit-*` run in that

tree (via walk-up).

**Modes** (auto-detected when `$ARGUMENTS` is blank; overridable by token):

| Mode | When | What it does |

|------|------|--------------|

| **fresh** | no project `.expectation-fit/` yet (or user said `fresh` / `all` / file tokens) | Full setup wizard → write files |

| **upgrade** | `.expectation-fit/` exists (or user said `upgrade`) | Diff capabilities vs defaults; merge only missing keys with confirm |

| **calibrate** | user said `calibrate` [percentile] | Measure distributions; propose thresholds with evidence |

Also supports legacy file tokens: `all`, `ways`, `patterns`, `thresholds` (skip profile

questions when non-interactive; still run stack-aware copy rules).

## Interactive tool

If presenting menus interactively in Claude Code, **pre-load `AskUserQuestion`**

(deferred tool): call `ToolSearch` with `select:AskUserQuestion` once before the first

menu. If the harness has no blocking-question tool (ToolSearch returns nothing / call

fails), fall back to a numbered list and wait for the user's reply — never silently pick.

---

## Procedure

### 0. Resolve mode + detect layout (always)

#### 0a. Mode

1. If `$ARGUMENTS` starts with token `calibrate` → **calibrate** (optional `p90` /

`p75` / `p95`; default **p90**). Jump to Step 2c.

2. Else if `$ARGUMENTS` has a **whole-token** `upgrade` (space-separated; not a

substring of another word) → **upgrade**. If `ways` / `patterns` / `thresholds` also

appear, limit upgrade merges to those files only. Jump to Step U after a light

detect pass (0b–0c still run so the report names stacks/roots).

3. Else if `$ARGUMENTS` has `fresh` or a file token (`all` / `ways` / `patterns` /

`thresholds`) **or** `multi-repo` / `roots:…` → **fresh** with that selection

4. Else (blank): discover nearest usable `.expectation-fit/` via walk-up

(`${CLAUDE_PLUGIN_ROOT}/references/config-resolution.md`).

- Found → recommend **upgrade**; still offer fresh re-scaffold (overwrite only with

confirm). Present: `1) upgrade 2) fresh (re-scaffold) 3) calibrate`.

- Not found → **fresh**.

#### 0b. Profile: monolith vs multi-repo

From the **intended project base** (cwd's git root, or a parent workspace if cwd is a

child app):

**Multi-repo signals** (any two → default profile **multi-repo**):

- Two or more immediate child dirs each with their own `.git`, or with `AGENTS.md` /

`CLAUDE.md` / `Gemfile` / `package.json` / `pyproject.toml`.

- Child names look like product halves (`*_be`, `*_fe`, `backend`, `frontend`, `api`,

`web`, `mobile`, `e2e`).

- Cwd is inside a child app but the parent has multiple app-like siblings.

Otherwise default **monolith**.

Always **confirm** with the user (AskUserQuestion): monolith vs multi-repo stack.

Recommend the detected default.

#### 0c. Stack detection

- **Read `${CLAUDE_PLUGIN_ROOT}/references/stack-catalog.md`** — the registry is the

source of truth for detection and which pack a stack carries. Match the tree against

`Detection signals`. Do not hardcode detection here.

- For multi-repo: detect **per root** under the project base (and under each

`conventions.auto.roots` candidate). Collect the set of stacks (e.g. rails + react).

- Note each stack's **Arch pack** status and **Threshold ns** from the catalog **column

only** (never a hardcoded "today" list):

- **Arch pack ✅** — scaffold that stack's threshold namespace + pattern ids.

- **Arch pack ⬜** — convention-only; no threshold/pattern pack yet.

- **Unknown / no match** — ways-of-working only for that root.

#### 0d. Surface inventory (informational; feeds later questions)

Under the project base and each multi-repo root, note presence of:

- `AGENTS.md` / `CLAUDE.md`

- `.github/copilot-instructions.md` or `**/copilot-instructions.md`

- `.github/instructions/**`

- `.github/workflows/*` that look like PR gates (name/body signals in

`${CLAUDE_PLUGIN_ROOT}/references/config-resolution.md` → Convention auto-sources)

Report a one-line inventory in the wizard summary (e.g. "agents=2, copilot=1,

instructions packs=3, gate-like workflows≈8").

---

### 1. Placement (fresh + multi-repo; skip for pure calibrate)

**Write root** for config files is `{project_base}/.expectation-fit/` unless per-repo is chosen.

| Profile | Default project base | Confirm |

|---------|----------------------|---------|

| **monolith** | git root of cwd | Usually no question; write `.expectation-fit/` there |

| **multi-repo** | **workspace root** that contains the sibling apps (parent of BE/FE/…), **not** deep inside one app | Ask: shared workspace `.expectation-fit/` (Recommended) vs per-repo `.expectation-fit/` only |

**Shared workspace config (recommended for multi-repo):**

- One `.expectation-fit/` at the workspace root (`project_base` = workspace).

- Child apps inherit via **walk-up** when agents run from `backend/` etc.

- Set `conventions.auto.roots` to the sibling dir names (relative to project base).

**Per-repo only:** when the user insists stacks diverge sharply:

1. List each sibling app root that will get its own `.expectation-fit/`.

2. Confirm the list; write **once per app** under `{app_root}/.expectation-fit/` (not only

workspace).

3. Do **not** set shared `auto.roots` across apps unless the user also wants a parent

config; warn that notes/severity will not share unless duplicated.

If cwd is already inside a child app and **shared** multi-repo is chosen, **change writes**

to the workspace root after confirmation — never use a relative `mkdir .expectation-fit` in the

child without an absolute `project_base`.

For **upgrade**, resolve existing `PROJECT_CONFIG` via walk-up; do not move it unless

the user explicitly asks to migrate placement.

**Legacy folder (`.intense/`, before 0.9.0).** In upgrade, always check for a legacy

`.intense/`: either `PROJECT_CONFIG` ends in `/.intense`, or a sibling `.intense/` sits

next to the resolved `.expectation-fit/`. Set `LEGACY_PARENT` to the absolute folder that

holds `.intense/` (not cwd). "Yaml" below means one of the three files

(`ways-of-working.yaml`, `patterns.yaml`, `thresholds.yaml`). Ask once, show each command,

and never move or remove without a yes.

- **Both folders hold yaml:** stop and list both. Config resolution uses

`.expectation-fit/` and ignores the legacy one. Offer to merge the keys that are

missing from `.expectation-fit/` out of the legacy files, then to remove `.intense/`.

Remove it only on a yes.

for a tracked file, else `mv` with both paths under `$LEGACY_PARENT`). Do not migrate

---

### 2. Choose what to scaffold (fresh) / menu

Parse `$ARGUMENTS`:

| Token | Effect |

|-------|--------|

| `all` | ways + patterns + thresholds (full set; not calibrate) |

| `ways` / `patterns` / `thresholds` | that file only |

| `fresh` | full wizard; default file set = all |

| `upgrade` | Step U |

| `calibrate` [p…] | Step 2c |

| blank | mode from 0a; then menu if still needed |

**Fresh interactive menu** (multi-select) when selection is still open:

| Option | File | What it controls |

|--------|------|------------------|

| Ways of working | `.expectation-fit/ways-of-working.yaml` | lenses, tools, severity, conventions (incl. auto), severity_align, confidence, artifacts |

| Pattern policy | `.expectation-fit/patterns.yaml` | preferred / allowed / blocked / approved |

| Thresholds | `.expectation-fit/thresholds.yaml` | architecture metric limits (Arch pack ✅ stacks only) |

| **Calibrate** | updates thresholds | measure + propose (after files exist) |

| All | all three | full config set (not calibrate) |

Recommend **All** for first run. Recommend **Calibrate** after All when any Arch pack ✅

is present.

Non-interactive / `$ARGUMENTS`-driven runs: skip menus; scaffold documented defaults

(and any values already known from arguments). File comments explain later edits.

---

### 2b. Capability questions (fresh interactive; also offered in upgrade)

When scaffolding or upgrading **ways-of-working** and/or **patterns** interactively,

ask short questions and write answers into the files. Defaults differ by profile:

| Question | Monolith default | Multi-repo default |

|----------|------------------|--------------------|

| Which files? | All | All |

| Lens toggles | predict/convention/simplicity **on**; experience/architecture **auto** | same |

| `tools.architecture` | `enrich` | same |

| `conventions.auto.mode` | `curated` | `curated` + **roots** filled from detected siblings |

| `severity_align.mode` | `curated_gates` | `curated_gates` |

| Prefer interactor over service_object? | **ask** (Rails Arch pack only) | **ask** (if any rails root) |

| Grandfather `app/services/**`? | only if blocking services | same |

| Artifact paths | defaults from config-resolution | same |

| Calibrate now? | offer after write | offer per Arch pack stack |

#### Questions (use AskUserQuestion; skip if non-interactive)

1. **Which lenses run?** Default as table. User may turn any **off**. Write `lenses.*`.

2. **External-tool preference** (if architecture not off): enrich / prefer / report / off.

Write `tools.architecture`.

3. **Convention auto-sources** — "Discover Copilot packs, path-scoped instructions, and

PR-gate workflows?" → `curated` (Recommended) / `all` / `off`. Write

`conventions.auto.mode`.

- Multi-repo: confirm **roots** list (detected siblings; user can edit). Write

`conventions.auto.roots: [be, fe, …]`. Empty roots = project base only.

- Keep default `include` and `exclude` from the template unless the user trims packs.

4. **Severity align with CI gates** — "Promote finding severity when a discovered PR

workflow matches the theme (e.g. callback check)?" → `curated_gates` (Recommended) /

`off`. Write `severity_align.mode` (and keep `min_severity: P1` unless changed).

5. **Pattern stance** (when scaffolding patterns **and** a stack catalog has

`interactor` + `service_object`, typically Rails):

- Prefer interactors for new multi-step domain work? → if yes, set **only**

`preferred: [{ id: interactor, instead_of: [service_object], when: "…", note: "…" }]`.

Do **not** also add `service_object` to `blocked` for the same stance (preferred

already P1s new instead_of use; double-listing creates duplicate findings).

- Grandfather legacy services? → if yes,

`approved: [{ path: app/services/**, reason: "legacy; no NEW services" }]`.

- Optional advanced: block without preferred only when the team wants a ban with no

replacement id.

- If no Rails (or user declines), leave preferred/blocked empty (or stack seed only).

6. **Where do reports go?** Confirm `artifacts.*` defaults from

`${CLAUDE_PLUGIN_ROOT}/references/config-resolution.md`. Say that reports stay local

by default (`.expectation-fit/reports/`, which ignores itself). Name

`artifacts.report_dir` as the setting that commits them (for example

`docs/expectation-fit`). Warn: reports name unfixed defects, and reports committed

under a published `docs/` site are public. Write overrides only if the user changes

them. Always restate resolved paths in the final summary.

7. **Seed `conventions.sources`?** If AGENTS/CLAUDE (or triage docs) exist under roots,

offer to add those **paths** to `sources` (not paste file bodies). Preview list;

user accepts / edits. Explicit sources always win over exclude.

Non-interactive: scaffold defaults with the Step 3 header (auto mode `curated`, severity_align on,

preferred empty). Set `conventions.auto.roots` when `$ARGUMENTS` includes

`multi-repo` and/or `roots:a,b,c` (comma-separated sibling dirs). If multi-repo

**signals** are strong but neither token is present, **do not write** `.expectation-fit/` only

inside a child app: refuse with the required tokens, or write at the detected workspace

root and print a hard warning in the report. Without tokens and without multi-repo

signals, place at the current git root (monolith).

---

### 2c. Calibrate thresholds (`calibrate` / menu)

When the user asks to calibrate (or `$ARGUMENTS` starts with `calibrate`):

1. Detect stack(s) via stack-catalog; require at least one **Arch pack ✅**. If none, stop

and explain.

2. Load metric keys from each stack's namespace in

`${CLAUDE_PLUGIN_ROOT}/config/defaults/thresholds.yaml` and path globs from the stack

architecture doc / pattern catalog (same units the architecture lens measures:

LOC, public methods, associations, callbacks, method length, etc.).

3. Measure distributions over the relevant tree (exclude `vendor`/`node_modules`/build):

for each metric compute **n, median, p75, p90, p95, max** (simple sort + index is fine;

Bash or a short Ruby/Python one-liner is OK). Multi-repo: measure under each root that

matches that stack.

4. Target percentile = argument (`p90` default) or user choice. Propose

`threshold = ceil(percentile)` (or just above p90 so the fat tail flags, not the bulk).

5. **Present** a table: metric | median | p75 | p90 | max | default | proposed. Flag any

metric where the shipped default is far outside the distribution (either direction).

6. On confirmation, write/update `.expectation-fit/thresholds.yaml` for those stack namespaces.

**Never clobber without confirmation.** Put **evidence in comments**:

```yaml

max_callbacks: 2 # median 0, p75 0, p90 2, max 29 (measured 2026-07-29, n=515, target p90)

```

7. If the user declines write, print the proposed YAML block for copy-paste.

If full measurement is too heavy for the session, fall back to the smaller form: run

the same measurement summary into the next `/fit-audit` Observations section and leave

thresholds unchanged until the human pastes numbers.

---

### 3. Copy / write templates (fresh; idempotent, stack-aware)

Source templates are `${CLAUDE_PLUGIN_ROOT}/config/defaults/<file>`. Write to

`{project_base}/.expectation-fit/<file>`. **Never overwrite an existing `.expectation-fit/<file>`**

without explicit confirmation — a surprising clobber of committed team config is the

failure mode to avoid (least astonishment / no data loss). If a target exists, report

it and ask whether to overwrite, diff, or skip; default to **skip**.

**Project header.** Every file this skill writes starts with the header below. It

replaces the template title line `(GLOBAL DEFAULTS)` and, in `ways-of-working.yaml`, the

template's copy instructions. Put no absolute path in the header: a path shows a home

folder and goes stale. No written file says `(GLOBAL DEFAULTS)`.

```yaml

# Expectation Fit: project config for <repo>.

# Values here override the plugin defaults. Omitted keys keep the default value.

# Written by /fit-setup <version> on <date>.

```

Before write on interactive fresh: show a **short preview** of the values that differ

from raw defaults (roots, auto mode, severity_align, preferred patterns, lens offs).

Confirm once, then write.

#### `ways-of-working.yaml`

Start from the default template, then apply capability answers:

- `lenses.*`, `tools.architecture`, `artifacts.*` from questions 1–2 and 6.

- `conventions.auto.mode`, `conventions.auto.roots`, optional `include`/`exclude` tweaks.

- `conventions.sources` from question 7 (paths that exist).

- `severity_align.mode` (+ keep `min_severity` / `themes` structure from template).

- Do **not** invent long `notes` at init; leave `notes: []` (use `/fit-from-pr-learnings`

later for PR-mined notes).

```bash

# PROJECT_BASE is the resolved placement root (workspace or monolith git root), absolute.

mkdir -p "$PROJECT_BASE/.expectation-fit"

SRC="${CLAUDE_PLUGIN_ROOT}/config/defaults/ways-of-working.yaml"

DST="$PROJECT_BASE/.expectation-fit/ways-of-working.yaml"

# Prefer: copy then edit keys, or emit merged YAML with comments preserved where practical.

if [ -e "$DST" ]; then echo "EXISTS: $DST"; else …; fi

```

#### `thresholds.yaml`

**Scaffold only the detected stack namespace(s)** (union for multi-repo), not every stack

in the global defaults file. Config-resolution deep-merges a partial file over defaults.

Emit the file header comment + each `<stack>:` block for detected Arch pack ✅ stacks:

```bash

SRC="${CLAUDE_PLUGIN_ROOT}/config/defaults/thresholds.yaml"

DST=".expectation-fit/thresholds.yaml"

# For each STACK in detected Arch-pack stacks, extract that top-level block with awk

# (same pattern as single-stack extract; append blocks for multi-stack).

```

For **Arch pack ⬜** / unknown only: skip `thresholds.yaml` and say no architecture pack

exists yet.

#### `patterns.yaml`

Copy default policy shape, then:

1. Seed `allowed:` with pattern ids from **each** detected stack's catalog

(`${CLAUDE_PLUGIN_ROOT}/resources/patterns/<stack>.yaml`) when that helps the team —

or leave empty with a top comment naming the stacks (empty allowed = no explicit

allow-list; lens still classifies). Prefer **commented** examples over a huge

allow-list unless the user asked for full seed.

2. Apply **preferred / blocked / approved** from capability question 5 when yes.

3. Note stacks in the file's top comment.

4. Pattern ids must exist in a catalog; never invent ids.

The written files keep explanatory comments so the team can edit in place.

---

### U. Upgrade mode (`upgrade`)

Goal: add **missing capabilities** from current plugin defaults without wiping hand-tuned

`notes`, `sources`, `severity_overrides`, roots, or pattern lists.

#### U1. Load

2. If none: say so and offer **fresh** instead.

3. Load project yaml + defaults from `${CLAUDE_PLUGIN_ROOT}/config/defaults/`.

4. Re-run light detect (0b–0d) for roots/stacks recommendations.

#### U2. Capability table

Build and **show** a table (do not write yet):

| Capability | In project? | Plugin default / recommendation | Proposed action |

|------------|-------------|----------------------------------|-----------------|

| `conventions.auto` block | present / missing / partial | mode curated; roots for multi-repo | add missing keys only |

| `conventions.auto.roots` | empty / set | detected siblings | set if empty and multi-repo |

| `severity_align` | present / missing | curated_gates | add if missing |

| `artifacts.*` | present / legacy `report_dir` only / old default values | two-layer defaults | carry legacy `report_dir` over as `artifacts.report_dir`; offer the new defaults for old values (U4) |

| `patterns.preferred` | empty / set | optional interactor preference | ask if rails + empty |

| `patterns.blocked` / approved | … | … | ask if preferred chosen |

| Missing config **files** | ways / patterns / thresholds | stack-aware scaffold | create missing files only |

| Thresholds never calibrated | n/a | offer calibrate | handoff Step 2c |

**Deep-merge reminder:** maps merge key-by-key. If the project has `conventions.notes`

but omits `conventions.auto`, defaults already supply `auto` at runtime. Upgrade still

**materializes** important blocks into the project file when the user wants them

visible/editable (least surprise for humans reading git).

#### U3. User picks

Interactive multi-select: enable missing / leave alone / customize. Default: enable

safe missing blocks (`auto`, `severity_align`, missing files) **without** overwriting

existing non-empty lists or notes.

#### U4. Apply rules (strict)

- **Never** delete or replace existing `conventions.notes` lines.

- **Never** clear non-empty `sources`, `preferred`, `blocked`, `approved`, or

`severity_overrides` without explicit overwrite confirm.

- **Add** missing top-level or nested keys from defaults (e.g. insert `severity_align:`

block; insert `conventions.auto:` if absent; fill `roots` only when empty and user

accepted).

- **Create** missing of the three files with stack-aware scaffolding (same as Step 3).

- **Legacy** top-level `report_dir`: carry its value over as `artifacts.report_dir`

(it is already an alias at runtime). Remove the legacy key only on confirm. Run

scratch then uses `artifacts.run_dir` (document this in the summary).

Name the legacy folder in the summary. Say that its old run folders stay. The user can

- **Old default values** (before 0.9.0): `artifacts.report_dir: docs/intent-engineering`,

`artifacts.run_dir: .intense/runs`, or a top-level `report_dir`. Name each one and

offer the new defaults: reports go to `.expectation-fit/reports/` and runs to

`.expectation-fit/runs/`, both git-ignored. Keep the old value when the user wants

committed reports. Change nothing without a yes.

- **Copied title:** when a file still starts with the template title `(GLOBAL DEFAULTS)`,

offer to replace that title with the Step 3 project header.

- Show a **diff-style** summary before write; confirm once.

#### U5. Optional handoffs

After upgrade: offer `/fit-setup calibrate` and `/fit-from-pr-learnings` (do not run them

unless the user asks).

---

### 4. Gitignore for run scratch (optional; re-offered until the path is ignored)

Resolve the **active** run-scratch path first (same rules as

`${CLAUDE_PLUGIN_ROOT}/references/config-resolution.md`):

- If project `.expectation-fit/ways-of-working.yaml` has `artifacts.run_dir` → use that.

- Else, only when the resolved config folder IS the legacy `.intense/` (not yet moved),

and its `ways-of-working.yaml` has `artifacts.run_dir` → use that.

- Else → default `.expectation-fit/runs/`.

New run folders ignore themselves (config-resolution writes a `*` `.gitignore` into

each one). This offer covers folders left by older runs. After scaffolding or upgrading ways-of-working (or on any init that mentions artifacts),

if the project `.gitignore` does **not** already ignore the resolved scratch root,

**offer** to append it (example for the default):

```

# Expectation Fit: ephemeral lens run scratch (from older runs; new ones ignore themselves)

.expectation-fit/runs/

```

This is **not** remembered across declines — if the line is still missing, the offer

reappears on the next `/fit-setup`. That is intentional (least surprise: no hidden

"don't ask again" flag). Never force; never add `.expectation-fit/` itself (config YAML must

stay committable). Do not invent a `wip/` ignore for the plugin default — `wip/` is no

longer the plugin report home.

For multi-repo **shared** config: prefer the workspace root `.gitignore`. If only a

child has a gitignore, note that runs may live at workspace `.expectation-fit/runs/` and the

ignore should be where that path is tracked.

---

### 5. Report

List what was created, updated, skipped, or proposed. Then tell the user:

- **Profile + placement:** monolith/multi-repo; absolute path of `.expectation-fit/`; roots if any.

- **Capabilities enabled:** auto mode, severity_align, preferred patterns, stacks.

- The files are **meant to be committed** (project config, not artifacts) — do **not**

add `.expectation-fit/` to `.gitignore` (only `.expectation-fit/runs/` if they accepted Step 4).

- Reports land under `.expectation-fit/reports/` by default and stay out of git (the

folder ignores itself). Set `artifacts.report_dir` to commit them. Run scratch is

cleaned up after each report step when `cleanup_runs: true`.

- Edit config to taste; every `fit-*` run merges project over plugin defaults (project

wins). Nested maps merge recursively at every depth. Only lists replace, unless the block sets `extends: true`.

- **Natural next steps:**

- `/fit-audit` for posture under the new config

- `/fit-setup calibrate p90` if Arch pack ✅ and thresholds still generic

- `/fit-from-pr-learnings` when PR triage exists (smarter notes over time; not required

at first install)

- For multi-repo: run agents from a child cwd and confirm Coverage shows

`Config: project:…` via walk-up

- **Git state:** after a write, end with the output of

`git -C "<resolved .expectation-fit dir>" status -sb .` and this line: "Commit these

config files in their own commit, not inside feature work." Outside a git repo, skip the

status line.

---

## Argument quick reference

| Invocation | Mode |

|------------|------|

| `/fit-setup` | Wizard: detect fresh vs upgrade; profile; capabilities |

| `/fit-setup all` | Fresh full file set (still capability questions if interactive) |

| `/fit-setup ways` | Ways only |

| `/fit-setup patterns` | Patterns only |

| `/fit-setup thresholds` | Thresholds only (stack namespaces) |

| `/fit-setup fresh` | Force fresh wizard |

| `/fit-setup upgrade` | Capability diff + merge missing |

| `/fit-setup calibrate` | Threshold measurement (default p90) |

| `/fit-setup calibrate p75` | Calibrate at p75 |

| `/fit-setup multi-repo` | Fresh with multi-repo profile default |

| `/fit-setup roots:backend,frontend` | Seed `conventions.auto.roots` (with or without multi-repo) |

---

## Reference files (read at runtime)

- `${CLAUDE_PLUGIN_ROOT}/references/config-resolution.md` — walk-up, merge, auto sources,

severity_align, artifacts

- `${CLAUDE_PLUGIN_ROOT}/references/stack-catalog.md` — stack detection + Arch pack

- `${CLAUDE_PLUGIN_ROOT}/config/defaults/` — source templates

- `${CLAUDE_PLUGIN_ROOT}/resources/patterns/<stack>.yaml` — valid pattern ids

- `${CLAUDE_PLUGIN_ROOT}/skills/fit-from-pr-learnings/SKILL.md` — post-init learnings loop

(not part of first install)