@imfusion/web-ui 0.6.4-dev.51.gf38b0107 → 0.6.4-dev.57.g664ba66a

package/dist/llms/evals/runner/affected.d.ts

@@ -0,0 +1,2 @@
+#!/usr/bin/env -S npx tsx
+export {};

package/dist/llms/evals/runner/benchmark.d.ts

@@ -0,0 +1,12 @@
+import { Benchmark, RunResult } from './types';
+type AggregateOptions = {
+    skillFamily: string;
+    runsPerScenario: number;
+    executorModel: string;
+    graderModel: string;
+    graderVotes: number;
+    tier: string;
+};
+export declare function aggregateBenchmark(runs: RunResult[], options: AggregateOptions): Benchmark;
+export declare function mergeBenchmarkUpdate(previous: Benchmark | undefined, update: Benchmark, updatedEvalIds: string[]): Benchmark;
+export {};

package/dist/llms/evals/runner/compare.d.ts

@@ -0,0 +1,2 @@
+#!/usr/bin/env -S npx tsx
+export {};

package/dist/llms/evals/runner/config.d.ts

@@ -0,0 +1,29 @@
+export declare const EVALS_DIR: string;
+export declare const NO_SKILLS: boolean;
+export declare const EVAL_ROOT: string;
+export declare const TEMPLATE: string;
+export declare const CODEX_HOME_DIR: string;
+export declare const RESULTS_DIR: string;
+export declare const REPORT_PATH: string;
+export declare const SCENARIOS_DIR: string;
+export declare const FIXTURES_DIR: string;
+export declare const SKILLS_DIR: string;
+export declare const SKILL_FAMILY: string;
+export declare const WORKSPACE_LABEL: string;
+export declare const DEFAULT_FIXTURE = "fresh-install";
+export declare const ROUND_TIMEOUT_MS = 600000;
+export declare const MAX_TURNS = 60;
+export declare const HOST_PLANS_DIR: string;
+type TierName = string;
+type Tier = {
+    runs: number;
+    graderVotes: number;
+};
+type EvalConfig = {
+    executorModel?: string;
+    graderModel?: string;
+    defaultTier?: TierName;
+    tiers?: Record<TierName, Tier>;
+};
+export declare const CONFIG: EvalConfig;
+export {};

package/dist/llms/evals/runner/evidence.d.ts

@@ -0,0 +1,7 @@
+import { HostPlan, ToolUse, TranscriptEntry } from './types.ts';
+export declare function isHostPlanPath(path: string): boolean;
+export declare function extractToolUses(transcript: TranscriptEntry[]): ToolUse[];
+export declare function extractSkillLoads(toolUses: ToolUse[]): string[];
+export declare function findHostPlans(toolUses: ToolUse[]): HostPlan[];
+export declare function renderConversation(transcript: TranscriptEntry[], maxChars?: number): string;
+export declare function renderWorkspaceFiles(workdir: string, hostPlans?: HostPlan[], maxChars?: number): string;

package/dist/llms/evals/runner/execute.d.ts

@@ -0,0 +1,32 @@
+import { TranscriptEntry } from './types.ts';
+export type RoundOutput = {
+    session_id: string | null;
+    result?: string;
+    duration_ms?: number;
+    total_cost_usd?: number;
+    usage?: {
+        input_tokens?: number;
+        output_tokens?: number;
+    };
+    modelUsage?: Record<string, unknown>;
+    subtype?: string;
+    num_turns?: number;
+    timedOut?: boolean;
+    timeoutMs?: number;
+};
+export declare function runClaude(prompt: string, { cwd, resumeSessionId, maxTurns, model, permissionMode, timeoutMs }: {
+    cwd: string;
+    resumeSessionId?: string | null;
+    maxTurns?: number;
+    model?: string | null;
+    permissionMode?: string;
+    timeoutMs?: number;
+}): RoundOutput;
+export declare function runCodex(prompt: string, { cwd, resumeSessionId }: {
+    cwd: string;
+    resumeSessionId?: string | null;
+}): RoundOutput;
+export declare function parseJsonl(raw: string): Record<string, unknown>[];
+export declare function readTranscript(workdir: string, sessionId: string): string;
+export declare function findCodexRollout(threadId: string): string;
+export declare function codexEntries(lines: Record<string, unknown>[]): TranscriptEntry[];

package/dist/llms/evals/runner/grade.d.ts

@@ -0,0 +1,10 @@
+import { AssertionResult, Scenario, ToolUse } from './types.ts';
+export declare function gradeRouting(label: "Skill" | "Tool", config: {
+    expect?: string[];
+    forbid?: string[];
+} | undefined, actual: string[]): AssertionResult[];
+export declare function gradeTranscript(scenario: Scenario, rawTranscript: string): AssertionResult[];
+export declare function gradeWorkspace(scenario: Scenario, workdir: string): AssertionResult[];
+export declare function gradeWrites(scenario: Scenario, toolUses: ToolUse[], workdir: string): AssertionResult[];
+export declare function gradeReport(scenario: Scenario, workdir: string, initialReport: string | null): AssertionResult[];
+export declare function gradeWithLlm(scenario: Scenario, conversation: string, workspaceFiles: string, voteCount: number): AssertionResult[];

package/dist/llms/evals/runner/report-md.d.ts

@@ -0,0 +1,2 @@
+#!/usr/bin/env -S npx tsx
+export declare function writeReportMarkdown(): string | null;

package/dist/llms/evals/runner/report.d.ts

@@ -0,0 +1,4 @@
+import { Benchmark } from './types.ts';
+export declare function printReport(benchmark: Benchmark, { allEvidence }?: {
+    allEvidence?: boolean | undefined;
+}): void;

package/dist/llms/evals/runner/run.d.ts

@@ -0,0 +1,2 @@
+#!/usr/bin/env -S npx tsx
+export {};

package/dist/llms/evals/runner/scenario.d.ts

@@ -0,0 +1,3 @@
+import { Host, RunResult, Scenario } from './types.ts';
+export declare function evalIdFor(scenario: Scenario, host: Host): string;
+export declare function runScenario(scenario: Scenario, host: Host, runNumber: number, runsRoot: string, model: string | null, graderVotes: number): RunResult;

package/dist/llms/evals/runner/types.d.ts

@@ -0,0 +1,140 @@
+export declare const HOSTS: readonly ["claude", "codex"];
+export type Host = (typeof HOSTS)[number];
+export type PathRule = {
+    pattern: string;
+    text: string;
+    flags?: string;
+    /** Named paths to search instead of the `src/**` default. */
+    files?: string[];
+};
+export type Scenario = {
+    id: string;
+    description: string;
+    rounds: string[];
+    /** Skills this scenario belongs to without asserting they loaded. */
+    covers?: string[];
+    skills?: {
+        expect?: string[];
+        forbid?: string[];
+    };
+    tools?: {
+        expect?: string[];
+        forbid?: string[];
+    };
+    writes?: {
+        only?: string[];
+        expect?: string[];
+    };
+    workspace?: {
+        forbidPatterns?: PathRule[];
+        expectPatterns?: PathRule[];
+    };
+    transcript?: {
+        expectStrings?: string[];
+        forbidStrings?: string[];
+    };
+    report?: {
+        path: string;
+        headings?: string[];
+        preserveFrom?: string;
+    };
+    assertions?: string[];
+    fixture?: string;
+    setup?: string;
+    hosts?: Host[];
+    maxTurns?: number;
+    roundTimeoutMinutes?: number;
+    permissionMode?: string;
+};
+export type AssertionResult = {
+    text: string;
+    passed: boolean;
+    method: "deterministic" | "llm";
+    evidence: string;
+};
+export type ToolUse = {
+    type: "tool_use";
+    name: string;
+    input?: Record<string, unknown>;
+};
+export type TranscriptEntry = {
+    type: string;
+    message?: {
+        content?: unknown;
+    };
+};
+export type HostPlan = {
+    path: string;
+    content: string;
+};
+export type RoundMeta = {
+    prompt: string;
+    session_id: string | null;
+    duration_ms?: number;
+    total_cost_usd?: number;
+    usage?: {
+        input_tokens?: number;
+        output_tokens?: number;
+    };
+    models?: string[];
+    subtype?: string;
+    num_turns?: number;
+    timed_out?: boolean;
+};
+export type RunResult = {
+    eval_id: string;
+    executor_host: Host;
+    fixture: string;
+    description: string;
+    run_number: number;
+    models: string[];
+    result: {
+        pass_rate: number;
+        passed: number;
+        failed: number;
+        total: number;
+        time_seconds: number;
+        turns: number;
+        tokens: number;
+        cost_usd: number;
+    };
+    expectations: {
+        text: string;
+        passed: boolean;
+        evidence: string;
+    }[];
+    truncated: boolean;
+    timed_out: boolean;
+    notes: string;
+};
+export type PerEvalSummary = {
+    pass_rate_mean: number;
+    pass_rate_stddev: number;
+    time_seconds_mean: number;
+    turns_mean: number;
+    tokens_mean: number;
+    cost_usd_total: number;
+};
+export type Benchmark = {
+    metadata: BenchmarkMetadata;
+    runs: RunResult[];
+    run_summary: {
+        per_eval: Record<string, PerEvalSummary>;
+        overall_pass_rate: number;
+        total_cost_usd: number;
+    };
+    notes: string[];
+};
+export type BenchmarkMetadata = {
+    skill_family: string;
+    timestamp: string;
+    evals_run: string[];
+    runs_per_scenario: number;
+    executor_model: string;
+    executor_hosts: string[];
+    grader_model: string;
+    grader_votes: number;
+    tier: string;
+    models_observed: string[];
+    fixtures: string[];
+};

package/dist/llms/evals/runner/workspace.d.ts

@@ -0,0 +1,2 @@
+export declare function copyTemplate(dest: string): void;
+export declare function applyFixture(dest: string, fixture: string): void;

package/dist/llms/evals/runner/write-generated.d.ts

@@ -0,0 +1 @@
+export declare function writeGenerated(path: string, contents: string): string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@imfusion/web-ui",
-  "version": "0.6.4-dev.51.gf38b0107",
+  "version": "0.6.4-dev.57.g664ba66a",
   "description": "The official Web UI component library for ImFusion web apps",
   "author": "ImFusion GmbH",
   "homepage": "https://imfusion.com",
@@ -90,10 +90,11 @@
     "test:unit": "vitest run --project unit",
     "test:stories": "vitest run --project storybook",
     "test:watch": "vitest",
-    "skills:eval": "node src/llms/evals/run.mjs",
-    "skills:eval:dev": "WEB_UI_SKILL_EVAL_ROOT=/private/tmp/web-ui-dev-skill-evals WEB_UI_SKILL_EVAL_RESULTS_DIR=.agents/evals/results WEB_UI_SKILL_EVAL_REPORT_PATH=.agents/evals/REPORT.md WEB_UI_SKILL_EVAL_SKILL_FAMILY=web-ui-dev WEB_UI_SKILL_EVAL_WORKSPACE_LABEL='development repository' node src/llms/evals/run.mjs --scenarios-dir .agents/evals/scenarios --setup .agents/evals/setup-env.sh",
-    "skills:eval:compare": "node src/llms/evals/compare.mjs",
-    "skills:eval:report": "node src/llms/evals/report.mjs",
+    "skills:eval": "tsx src/llms/evals/runner/run.ts",
+    "skills:eval:dev": "WEB_UI_SKILL_EVAL_ROOT=/private/tmp/web-ui-dev-skill-evals WEB_UI_SKILL_EVAL_RESULTS_DIR=.agents/evals/results WEB_UI_SKILL_EVAL_REPORT_PATH=.agents/evals/REPORT.md WEB_UI_SKILL_EVAL_SKILL_FAMILY=web-ui-dev WEB_UI_SKILL_EVAL_WORKSPACE_LABEL='development repository' tsx src/llms/evals/runner/run.ts --scenarios-dir .agents/evals/scenarios --setup .agents/evals/setup-env.sh",
+    "skills:eval:compare": "tsx src/llms/evals/runner/compare.ts",
+    "skills:eval:affected": "tsx src/llms/evals/runner/affected.ts",
+    "skills:eval:report": "tsx src/llms/evals/runner/report.ts",
     "git:config": "git config core.hooksPath .githooks && git config pull.rebase true && git config merge.ff only"
   },
   "peerDependencies": {

package/src/llms/skills/imf-web-ui/SKILL.md

@@ -13,27 +13,30 @@ copy of every convention.
 
 ## 1. Decide whether guidance is needed
 
-Skip a companion when the task is explicit and small, an existing local pattern already solves it, and the library usage is
-already correct. Do the work. Look up an API silently only when you are unsure.
+Skip a companion when the task is explicit and small, an existing local pattern already solves it in the file being changed,
+and the library usage is already correct. Do the work. Look up an API silently only when you are unsure.
 
-Use a companion when the task involves a choice, a missing setup piece, a new file, or a library convention the project may
-not already have.
+Use a companion when the task involves a choice, a missing setup piece, a new UI file, or a library convention the project
+may not already have. Open a companion by invoking its skill; reading one of its files directly does not load its workflow.
 
 ## 2. Route the task
 
-| Task                                                              | Open                                                   |
-| ----------------------------------------------------------------- | ------------------------------------------------------ |
-| Understand library setup, brand assets, theming, or usage         | Packaged user guides below                             |
-| Look up a component, part, prop, default, or icon                 | `imf-web-ui-components`                                |
-| Choose components or shape a screen or flow                       | `imf-web-ui-ux`                                        |
-| Write or update documentation                                     | `/documentation-writer`, then `imf-web-ui-conventions` |
-| Write a wrapper, custom UI, CSS, data layer, validation, or tests | `imf-web-ui-conventions`                               |
-| Install the library or bootstrap project tooling                  | `imf-web-ui-setup`                                     |
-| Inspect an existing project without changing it                   | `imf-web-ui-audit`                                     |
-| Update the package, skills, or hooks                              | `imf-web-ui-update`                                    |
-
-A screen often needs both `imf-web-ui-ux` and `imf-web-ui-components`, in that order. Setup and audit are for project-wide
-questions, not every one-file edit.
+| Task                                                                 | Open                                                   |
+| -------------------------------------------------------------------- | ------------------------------------------------------ |
+| Understand library setup, brand assets, theming, or usage            | Packaged user guides below                             |
+| Look up a component, part, prop, default, or icon                    | `imf-web-ui-components`                                |
+| Choose components or shape a screen, form, or flow                   | `imf-web-ui-ux`                                        |
+| Write or update documentation                                        | `/documentation-writer`, then `imf-web-ui-conventions` |
+| Write a wrapper, custom UI, CSS, data layer, validation, or tests    | `imf-web-ui-conventions`                               |
+| Choose a routing, data, form, or validation library for the project  | `imf-web-ui-conventions`                               |
+| Install the library or bootstrap project tooling                     | `imf-web-ui-setup`                                     |
+| Diagnose a broken install: unstyled output, missing tokens, no theme | `imf-web-ui-setup`                                     |
+| Inspect an existing project without changing it                      | `imf-web-ui-audit`                                     |
+| Update the package, skills, or hooks                                 | `imf-web-ui-update`                                    |
+
+A new form opens `imf-web-ui-ux`, even when its fields and action are already specified. A screen often needs both
+`imf-web-ui-ux` and `imf-web-ui-components`, in that order. Setup and audit are for project-wide questions, not every
+one-file edit.
 
 ### Read a packaged user guide
 
@@ -58,8 +61,9 @@ props and the conventions `tokens` topic for exact token names and defaults.
 The host project's existing conventions win. The companion skills fill gaps; they do not justify refactoring a working
 styling system, state library, or folder structure.
 
-When a task needs TanStack Router, Query, Form, Table, or Store and the project has no incumbent, propose the matching
-library and read its current documentation with `npx @tanstack/cli` before using it.
+When a task needs TanStack Router, Query, Form, Table, or Store and the project has no incumbent, open
+`imf-web-ui-conventions` for the house position on that layer, then read the library's current documentation with
+`npx @tanstack/cli` before using it. The recommendation lives in the conventions topics, not here.
 
 ## 4. Ask only when the choice depends on missing context
 

package/src/llms/skills/imf-web-ui-audit/SKILL.md

@@ -10,22 +10,35 @@ allowed-tools: Read Glob Grep
 
 # Audit a consumer project
 
-This is a read-only audit. Compare the project with the selected `imf-web-ui-conventions` topics, report evidence, and end
-with work the human can approve. The baseline is an ImFusion default, not a universal law; a deliberate project choice is a
-deviation, not a defect.
+This is a read-only audit. Its job is to align a project with its conventions: compare the project against the selected
+`imf-web-ui-conventions` topics, report evidence, and end with work the human can approve.
+
+A gap against a convention is high severity. The project agreed to those conventions, so the finding carries no argument
+about whether it matters, and a deliberate reason to differ is the only thing that lowers it. Say what that reason is.
+
+Anything else you notice belongs under `Suggestions`: what running the code revealed, what an investigation turned up, what a
+reader would tidy on the way past. Report it, keep it low, and do not let it crowd the sections a consumer has to act on.
 
 ## Workflow
 
 1. Resolve the argument. Bare means `full`; an unknown topic is an error, not a reason to widen the scope.
 2. Enter plan mode unless this audit is being called as verification by another skill or the host has no plan mode.
-3. Assign one investigator to each in-scope topic. An investigator checks every section in that topic and writes nothing. If
-   the host cannot dispatch agents, inspect the topics inline in the same order.
+3. Read the report template and existing reviewer notes, then dispatch one investigator per in-scope topic. Tell each
+   investigator to use only `Read`, `Glob`, and `Grep`, with no shell commands or writes. An investigator checks every
+   section in that topic and writes nothing. Once dispatch starts, the host performs no repository inspection: the returned
+   reports are the evidence, and the host merges them without reading or searching the files again. Gathering the evidence
+   twice spends context on the second pass and reports what the host still remembers. Inspect a topic inline only after its
+   dispatch has been tried and refused, and say which attempt failed.
 4. Merge the evidence into the shared report format in
-   [`templates/REPORT.md`](../imf-web-ui-conventions/templates/REPORT.md).
-5. Put each finding's `Next action` into an ordered plan, grouped by topic and cheapest first. Present the report and wait
-   for approval; an audit does not edit the project.
+   [`templates/REPORT.md`](../imf-web-ui-conventions/templates/REPORT.md). Before presenting or writing it, check that every
+   Broken, Missing, Deviation, and Unverified entry has a severity, evidence, impact, and next action; do not compact
+   Unverified entries.
+5. Put each finding's `Next action` into an ordered plan, grouped by topic and cheapest first. The plan is separate from the
+   report contract: present it in the host plan or response, never as another heading in `AUDIT_REPORT.md`. Present the
+   report and wait for approval; an audit does not edit the project.
 
-If the caller needs a durable report, write `AUDIT_REPORT.md` and preserve everything under `## Reviewer notes` verbatim.
+If the caller needs a durable report, write `AUDIT_REPORT.md` with exactly the template's H2 headings and preserve everything
+under `## Reviewer notes` verbatim.
 
 ## Finding format
 
@@ -37,8 +50,9 @@ Classify every result as one of these:
 - **Present**: the rule is met, with evidence.
 - **Unverified**: static inspection cannot establish it.
 
-Use the report contract's severity, `path:line` evidence, concrete impact, and smallest next action. Do not promote an
-optional tool or a working alternative to a missing finding.
+Use the report contract's severity, `path:line` evidence, concrete impact, and smallest next action. Unverified entries use
+that same structure: unavailable runtime evidence, the impact of the unknown, and the check that resolves it. Do not promote
+an optional tool or a working alternative to a missing finding.
 
 ## Topics
 

package/src/llms/skills/imf-web-ui-components/SKILL.md

@@ -87,7 +87,7 @@ A component whose identity entry says it comes from `@imfusion/web-ui/integratio
 listed optional peer explicitly, then import from that path:
 
 ```tsx
-import { Code } from "@imfusion/web-ui/integrations/code-highlight";
+import { ImageDisplayOptions } from "@imfusion/web-ui/integrations/image-display-options";
 ```
 
 ## Data grids

package/src/llms/skills/imf-web-ui-conventions/templates/REPORT.md

@@ -12,8 +12,23 @@ Broken, Missing, Deviations, and Unverified entries use:
   - Impact: concrete consequence
   - Next action: smallest selectable follow-up
 
-Present entries name the topic and evidence path. Deviations are selectable follow-up work, not defects. Optional tools are not
-missing findings.
+Present entries name the topic and evidence path. Optional tools are not missing findings.
+
+Where a finding belongs follows from where it came from, not from how strongly you hold it.
+
+A convention names it:
+- Broken: the convention is followed but does not work. Something is unmounted, unwired, or fails at runtime.
+- Missing: a convention expects it and the project has nothing there.
+- Deviations: it works, and differs from what a convention says.
+
+These three are high severity by default. The project agreed to the conventions; a gap against them is not a preference.
+Drop the severity only when the project shows a deliberate reason to differ, and say what the reason is.
+
+Nothing names it:
+- Suggestions: real improvements that no convention asks for. What running the code revealed, what an investigation turned
+  up, what a reader would tidy on their way past. Never high severity, and never a reason to hold up the work.
+
+- Unverified: it could not be checked from the files available.
 -->
 
 ## Verdict
@@ -32,6 +47,10 @@ None.
 
 None.
 
+## Suggestions
+
+None.
+
 ## Present
 
 None.

package/src/llms/skills/imf-web-ui-conventions/topics/data.md

@@ -85,7 +85,9 @@ export function getDetails(options?: QueryOptions<User>) {
 }
 ```
 
-Callers choose `ensureQueryData`, `useSuspenseQuery`, or another Query API.
+Callers choose `ensureQueryData`, `useSuspenseQuery`, or another Query API. Mutation factories likewise return
+`mutationOptions`, own their mutation key and parsing, and are passed to `useMutation`; do not inline those options at the
+call site.
 
 ## Router context access
 

package/src/llms/skills/imf-web-ui-conventions/topics/tokens.md

@@ -8,5 +8,7 @@ Read:
 node_modules/@imfusion/web-ui/src/llms/tokens.gen.json
 ```
 
-It contains the shipped token names and authored defaults. Look up the exact entry instead of guessing a plausible
-`--imf-ui-*` name or copying a token list into the project.
+It contains the shipped token names and authored defaults. Before writing CSS, check every concrete `--imf-ui-*` name the
+file will use against that index or the relevant component stylesheet. A family with one entry does not imply numbered
+variants. When no matching token exists, use an existing documented seam or say that the token is absent instead of guessing
+a plausible name.

package/src/llms/skills/imf-web-ui-setup/SKILL.md

@@ -5,7 +5,7 @@ description:
   needs library wiring, tooling, git, npm, authentication, structure, data, testing, docs, or agent tooling. Inspect first,
   propose concrete files, and write only what the user approves. Use imf-web-ui-audit for read-only checks."
 argument-hint: "[full|library-setup|tooling|git|npm-project|authentication|project-structure|docs-structure|data|testing|agent-tooling]"
-allowed-tools: Read Glob Grep Write
+allowed-tools: Read Glob Grep Write Skill
 ---
 
 # Set up a consumer project
@@ -15,16 +15,25 @@ working choices.
 
 ## Workflow
 
-1. Resolve the argument. Bare means `full`. A named topic limits the assessment to that topic. If the argument is unknown,
-   list the supported topics and point to `imf-web-ui-audit`.
+1. Resolve the scope. A named topic limits the assessment to that topic. A bare invocation means `full`, unless the request
+   is about one topic, in which case that topic is the scope: answer what was asked and offer the wider sweep rather than
+   running it. If the argument is unknown, list the supported topics and point to `imf-web-ui-audit`.
 2. Enter the host's plan mode before inspecting or proposing changes.
 3. Read the selected convention topics and inspect the repository with `Read`, `Glob`, and `Grep` only.
 4. Put a complete proposal in the host plan, using [`templates/REPORT.md`](../imf-web-ui-conventions/templates/REPORT.md).
    For new or updated documentation, invoke `/documentation-writer` before writing and use the `docs-structure` topic for
    repository-specific boundaries. Name every file to create or change, include its intended content, and cite the repository
    evidence behind the choice.
-5. Wait for explicit approval. After approval, write only the listed files.
-6. Run `imf-web-ui-audit full` as a follow-up and review its findings before calling setup complete.
+5. Wait for explicit approval. After approval, write only the listed files. Before calling setup complete, ensure every
+   package imported by those files and every tool named by their scripts is declared in `package.json`.
+6. Offer `imf-web-ui-audit full` as an optional next step for checking the wider project, and run it only if the user says
+   yes. Setup is complete once the approved files are written; the audit reviews ground setup did not touch.
+
+## Audit handoff
+
+When the approved request includes an audit, the next tool call after the final setup write invokes `imf-web-ui-audit`
+through the host's skill tool. Do not read audit files, inspect the project again, or dispatch audit investigators first;
+reproducing the audit workflow does not load the skill.
 
 ## Greenfield `full` setup
 

package/src/llms/skills/imf-web-ui-update/SKILL.md

@@ -20,8 +20,10 @@ Never overwrite user changes or commit without fresh approval.
    if another lockfile is the active one.
 3. Record `git status --short`, the installed Web UI version, the installed skill targets, hooks, registrations, and
    `AGENTS.md` fence.
-4. Check for another npm process and stop on a conflict. A dirty file this workflow needs is also a conflict. Do not reset,
-   clean, restore, or hide it.
+4. Check for another npm process and stop on a conflict. Uncommitted work is also a conflict: report it and stop before the
+   first mutating command, rather than deciding it falls outside this workflow's paths. What the update touches is not
+   knowable until it runs, and the person can commit, stash, or wave it through in one reply. Never reset, clean, restore, or
+   hide their work.
 5. With `--dry-run`, stop after reporting what would happen. Do not install, format, stage, commit, or ask for commit
    approval.
 

package/src/llms/skills/imf-web-ui-ux/SKILL.md

@@ -15,7 +15,9 @@ when it is used across many call sites so API movement stays local.
 
 ## Ask only what matters
 
-If the request is vague and a human can answer, ask these questions one at a time and stop once the choice is clear:
+When the request is vague and a human is there to answer, the first reply is a question and nothing else. No default, no
+options table, no recommendation held in reserve: any of those end the conversation, because the person now has an answer and
+no reason to reply. Ask one at a time, and stop once the choice is clear:
 
 1. Who uses the screen, and how often?
 2. What is its one primary action?
@@ -23,8 +25,8 @@ If the request is vague and a human can answer, ask these questions one at a tim
 4. What should users see when it is empty, loading, or failing?
 5. Is it a full page or part of another flow?
 
-If the prompt already answers a question, do not ask it again. If nobody can answer, make the conservative choice and state
-the assumption.
+If the prompt already answers a question, do not ask it again. Only when nobody is there to answer, or a question has gone
+unanswered, make the conservative choice and state the assumption.
 
 ## Choose a surface