pi-ui-extend 1.0.42 → 1.0.45

package/skills/spec-wiki/evals/files/bootstrap/requirements/payment-behavior.md

@@ -1,17 +0,0 @@
-# Payment capture requirements
-
-## Behavior
-
-- A capture request with the same idempotency key must not charge twice.
-- A successful capture returns the existing capture result when the same key is
-  replayed.
-- A different amount with a reused idempotency key is rejected.
-
-## Related files
-
-- `src/payments.ts`
-- `tests/payments.test.ts`
-
-## Verification
-
-Cover first capture, exact replay, and conflicting replay.

package/skills/spec-wiki/evals/files/bootstrap/src/payments.ts

@@ -1,12 +0,0 @@
-const captures = new Map<string, { amount: number; id: string }>();
-
-export function capture(key: string, amount: number): { amount: number; id: string } {
-  const existing = captures.get(key);
-  if (existing) {
-    if (existing.amount !== amount) throw new Error("idempotency conflict");
-    return existing;
-  }
-  const result = { amount, id: `capture-${captures.size + 1}` };
-  captures.set(key, result);
-  return result;
-}

package/skills/spec-wiki/evals/files/bootstrap/tests/payments.test.ts

@@ -1,11 +0,0 @@
-// @ts-nocheck
-import assert from "node:assert/strict";
-import { test } from "node:test";
-
-import { capture } from "../src/payments.js";
-
-test("idempotent replay returns the same capture", () => {
-  const first = capture("key-1", 100);
-  assert.deepEqual(capture("key-1", 100), first);
-  assert.throws(() => capture("key-1", 200), /idempotency conflict/);
-});

package/skills/spec-wiki/evals/files/focused/src/session.ts

@@ -1,5 +0,0 @@
-export type TokenLoader = () => Promise<string>;
-
-export async function refreshSession(loadToken: TokenLoader): Promise<string> {
-  return loadToken();
-}

package/skills/spec-wiki/evals/files/focused/tests/session.test.ts

@@ -1,17 +0,0 @@
-// @ts-nocheck
-import assert from "node:assert/strict";
-import { test } from "node:test";
-
-import { refreshSession } from "../src/session.js";
-
-test("refreshSession performs one token-load attempt", async () => {
-  let attempts = 0;
-  await assert.rejects(
-    refreshSession(async () => {
-      attempts += 1;
-      throw new Error("transient");
-    }),
-    /transient/,
-  );
-  assert.equal(attempts, 1);
-});

package/skills/spec-wiki/evals/files/maintenance/low-signal/incoming/capture.txt

@@ -1,2 +0,0 @@
-Duplicate capture requests with the same idempotency key return the original result.
-This rule describes current production behavior for capture requests.

package/skills/spec-wiki/evals/files/maintenance/low-signal/src/capture.ts

@@ -1,3 +0,0 @@
-export function capture(idempotencyKey: string, previous?: string): string {
-  return previous ?? `capture:${idempotencyKey}`;
-}

package/skills/spec-wiki/evals/files/maintenance/move/docs/old-job-contract.md

@@ -1,18 +0,0 @@
-# Background job cancellation contract
-
-## Type
-
-As-is
-
-## Lifecycle
-
-Active
-
-## Behavior
-
-Cancelling a running job is idempotent. Running jobs become cancelled; already
-terminal jobs remain terminal.
-
-## Related files
-
-- `src/job.ts`

package/skills/spec-wiki/evals/files/maintenance/move/src/job.ts

@@ -1,5 +0,0 @@
-export type JobState = "running" | "done" | "failed" | "cancelled";
-
-export function cancelJob(state: JobState): JobState {
-  return state === "running" ? "cancelled" : state;
-}

package/skills/spec-wiki/evals/files/maintenance/new-relation/docs/session-refresh.md

@@ -1,24 +0,0 @@
-# Session refresh contract
-
-## Type
-
-As-is
-
-## Lifecycle
-
-Active
-
-## Behavior
-
-Interactive and background session refresh retry a transient token-acquisition
-failure exactly once. A second token-acquisition failure propagates to the
-caller. Successful acquisition is not retried.
-
-## Related files
-
-- `src/session.ts`
-
-## Verification
-
-- Interactive refresh covers success, retry-success, and second-failure paths.
-- Every additional refresh entrypoint must preserve the same retry contract.

package/skills/spec-wiki/evals/files/maintenance/new-relation/incoming/refresh-worker.ts

@@ -1,7 +0,0 @@
-export async function refreshInBackground(loadToken: () => Promise<string>): Promise<string> {
-  try {
-    return await loadToken();
-  } catch {
-    return await loadToken();
-  }
-}

package/skills/spec-wiki/evals/files/maintenance/new-relation/src/session.ts

@@ -1,7 +0,0 @@
-export async function refreshSession(loadToken: () => Promise<string>): Promise<string> {
-  try {
-    return await loadToken();
-  } catch {
-    return await loadToken();
-  }
-}

package/skills/spec-wiki/evals/files/maintenance/no-impact/docs/auth-contract.md

@@ -1,17 +0,0 @@
-# Authentication contract
-
-## Type
-
-As-is
-
-## Lifecycle
-
-Active
-
-## Behavior
-
-Authentication accepts a valid token and rejects an empty token.
-
-## Related files
-
-- `src/auth.ts`

package/skills/spec-wiki/evals/files/maintenance/no-impact/incoming/telemetry-format.ts

@@ -1,3 +0,0 @@
-export function formatTelemetryName(name: string): string {
-  return name.trim().toLowerCase();
-}

package/skills/spec-wiki/evals/files/maintenance/no-impact/src/auth.ts

@@ -1,3 +0,0 @@
-export function authenticate(token: string): boolean {
-  return token.length > 0;
-}

package/skills/spec-wiki/evals/files/query/docs/how-to-run-jobs.md

@@ -1,4 +0,0 @@
-# How to run jobs locally
-
-Start the worker, enqueue a sample job, and tail the development log. This is a
-usage guide and does not define cancellation semantics.

package/skills/spec-wiki/evals/files/query/docs/spec-overview.md

@@ -1,9 +0,0 @@
-# Job specifications overview
-
-The primary job behavior documents are linked below.
-
-- [Cancellation contract](../features/jobs/cancellation-contract.md) — jobs can
-  be stopped safely.
-- [Retry spec](../features/jobs/retry-spec.md) — retry timing and limits.
-
-This page is a catalog, not a behavioral source of truth.

package/skills/spec-wiki/evals/files/query/features/jobs/cancellation-contract.md

@@ -1,15 +0,0 @@
-# Background job cancellation contract
-
-## Behavior
-
-- Cancellation requests are idempotent.
-- A running job transitions to `cancelled` when cancellation is accepted.
-- Jobs already in `completed`, `failed`, or `cancelled` remain in that terminal
-  state when cancellation is requested again.
-- Cancellation does not imply that already-committed external side effects are
-  rolled back.
-
-## Verification
-
-Tests should cover repeated cancellation, running-to-cancelled transition, and
-all terminal states.

package/skills/spec-wiki/evals/files/query/features/jobs/retry-spec.md

@@ -1,6 +0,0 @@
-# Background job retry behavior
-
-## Behavior
-
-Failed jobs retry at most three times with exponential backoff. Cancellation is
-outside this document's scope.

package/skills/spec-wiki/evals/maintenance_regression.py

@@ -1,311 +0,0 @@
-#!/usr/bin/env python3
-"""Deterministic regression checks for Spec Wiki maintenance primitives."""
-
-from __future__ import annotations
-
-import json
-import subprocess
-import tempfile
-from pathlib import Path
-
-
-CLI = Path(__file__).resolve().parents[1] / "scripts" / "spec_wiki.py"
-
-
-def run(cwd: Path, *args: str) -> str:
-    proc = subprocess.run(
-        ["python3", str(CLI), "--root", str(cwd), *args],
-        cwd=cwd,
-        text=True,
-        stdout=subprocess.PIPE,
-        stderr=subprocess.PIPE,
-        check=False,
-    )
-    if proc.returncode != 0:
-        raise AssertionError(f"command failed {args}:\nstdout={proc.stdout}\nstderr={proc.stderr}")
-    return proc.stdout
-
-
-def run_json(cwd: Path, *args: str) -> dict:
-    return json.loads(run(cwd, *args, "--json"))
-
-
-def git(cwd: Path, *args: str) -> None:
-    subprocess.run(["git", *args], cwd=cwd, check=True, stdout=subprocess.DEVNULL)
-
-
-def write(root: Path, rel: str, text: str) -> None:
-    path = root / rel
-    path.parent.mkdir(parents=True, exist_ok=True)
-    path.write_text(text, encoding="utf-8")
-
-
-def init_repo(root: Path) -> None:
-    git(root, "init", "-q")
-    git(root, "config", "user.email", "spec-wiki-eval@example.com")
-    git(root, "config", "user.name", "Spec Wiki Eval")
-
-
-def commit_all(root: Path, message: str = "baseline") -> None:
-    git(root, "add", ".")
-    git(root, "commit", "-qm", message)
-
-
-def test_relation_map_invalidates_verification(tmp: Path) -> None:
-    root = tmp / "relations"
-    root.mkdir()
-    init_repo(root)
-    write(root, "src/a.ts", "export const a = 1\n")
-    write(root, "src/b.ts", "export const b = 2\n")
-    write(
-        root,
-        "docs/contract.md",
-        """# Contract
-
-## Type
-As-is
-
-## Lifecycle
-Active
-
-## Behavior
-A is part of the current contract.
-
-## Related files
-- `src/a.ts`
-""",
-    )
-    commit_all(root)
-    run(
-        root,
-        "record",
-        "--path",
-        "docs/contract.md",
-        "--classification",
-        "spec",
-        "--type",
-        "as-is",
-        "--lifecycle",
-        "active",
-        "--confidence",
-        "high",
-        "--summary",
-        "Contract for A.",
-    )
-    run(root, "verify", "--path", "docs/contract.md")
-    audit = run_json(root, "audit")
-    assert audit["freshCount"] == 1
-
-    run(root, "relate", "--path", "docs/contract.md", "--add-code", "src/b.ts")
-    audit = run_json(root, "audit")
-    assert audit["freshCount"] == 0
-    status = next(item for item in audit["statuses"] if item["path"] == "docs/contract.md")
-    assert status["status"] == "inputs-changed"
-    assert "relations:changed" in status["reasons"]
-
-    run(root, "verify", "--path", "docs/contract.md")
-    assert run_json(root, "audit")["freshCount"] == 1
-    run(root, "relate", "--path", "docs/contract.md", "--remove-code", "src/b.ts")
-    status = next(item for item in run_json(root, "audit")["statuses"] if item["path"] == "docs/contract.md")
-    assert status["status"] == "inputs-changed"
-    assert status["reasons"] == ["relations:changed"]
-
-
-def test_impact_detects_coverage_gaps_and_moves(tmp: Path) -> None:
-    root = tmp / "impact"
-    root.mkdir()
-    init_repo(root)
-    write(root, "src/session.ts", "export function session() { return 'ok' }\n")
-    write(
-        root,
-        "docs/session-contract.md",
-        """# Session contract
-
-## Type
-As-is
-
-## Lifecycle
-Active
-
-## Behavior
-Session loading returns the current session.
-
-## Related files
-- `src/session.ts`
-""",
-    )
-    commit_all(root)
-    run(
-        root,
-        "record",
-        "--path",
-        "docs/session-contract.md",
-        "--classification",
-        "spec",
-        "--type",
-        "as-is",
-        "--lifecycle",
-        "active",
-        "--confidence",
-        "high",
-        "--summary",
-        "Session loading contract.",
-    )
-    run(root, "verify", "--path", "docs/session-contract.md")
-
-    git(root, "mv", "docs/session-contract.md", "docs/session-lifecycle.md")
-    with (root / "docs/session-lifecycle.md").open("a", encoding="utf-8") as handle:
-        handle.write("\nMoved without changing the behavioral contract.\n")
-    write(root, "src/refresh-worker.ts", "export const retryOnce = true\n")
-    write(root, "notes.txt", "deliberately low signal document\n")
-
-    impact = run_json(root, "impact")
-    assert ".spec-wiki/state.json" not in impact["changedPaths"]
-    assert any(
-        item.get("status") == "R"
-        and item.get("oldPath") == "docs/session-contract.md"
-        and item.get("path") == "docs/session-lifecycle.md"
-        for item in impact["changes"]
-    )
-    assert "src/refresh-worker.ts" in impact["uncoveredPaths"]
-    changed_docs = {item["path"]: item for item in impact["changedDocuments"]}
-    assert changed_docs["docs/session-lifecycle.md"]["requiresClassification"] is True
-    assert changed_docs["notes.txt"]["score"] == 0
-    assert "docs/session-contract.md" in impact["missingTrackedSpecs"]
-    assert impact["semanticSweepRequired"] is True
-
-
-def test_deep_audit_finds_low_signal_committed_docs(tmp: Path) -> None:
-    root = tmp / "deep"
-    root.mkdir()
-    init_repo(root)
-    write(
-        root,
-        "notes/capture.txt",
-        "Duplicate capture requests with the same key return the original result.\n",
-    )
-    write(root, "skills/example/evals/fixture.md", "# Contract\n\n## Behavior\nFixture only.\n")
-    commit_all(root)
-
-    normal = run_json(root, "discover", "--limit", "100")
-    assert "notes/capture.txt" not in {item["path"] for item in normal["candidates"]}
-    deep = run_json(root, "discover", "--all-unclassified", "--limit", "100")
-    paths = {item["path"] for item in deep["candidates"]}
-    assert "notes/capture.txt" in paths
-    assert "skills/example/evals/fixture.md" not in paths
-
-
-def test_relate_restores_source_explicit_relation(tmp: Path) -> None:
-    root = tmp / "supersession"
-    root.mkdir()
-    init_repo(root)
-    write(root, "docs/b.md", "# B\n")
-    write(
-        root,
-        "docs/a.md",
-        """# A contract
-
-## Type
-As-is
-
-## Lifecycle
-Active
-
-## Behavior
-A references B.
-
-[B](./b.md)
-""",
-    )
-    commit_all(root)
-    run(
-        root,
-        "record",
-        "--path",
-        "docs/a.md",
-        "--classification",
-        "spec",
-        "--type",
-        "as-is",
-        "--lifecycle",
-        "active",
-        "--confidence",
-        "high",
-        "--summary",
-        "A contract.",
-    )
-    run(root, "verify", "--path", "docs/a.md")
-    run(root, "relate", "--path", "docs/a.md", "--add-supersedes", "docs/b.md")
-    state = json.loads((root / ".spec-wiki/state.json").read_text(encoding="utf-8"))
-    relations = state["entries"]["docs/a.md"]["relations"]["specs"]
-    assert any(
-        item.get("path") == "docs/b.md"
-        and item.get("source") == "explicit"
-        and item.get("kind") == "related"
-        for item in relations
-    )
-    assert any(item.get("path") == "docs/b.md" and item.get("kind") == "supersedes" for item in relations)
-    run(root, "relate", "--path", "docs/a.md", "--remove-supersedes", "docs/b.md")
-    state = json.loads((root / ".spec-wiki/state.json").read_text(encoding="utf-8"))
-    relations = state["entries"]["docs/a.md"]["relations"]["specs"]
-    assert any(
-        item.get("path") == "docs/b.md"
-        and item.get("source") == "explicit"
-        and item.get("kind") == "related"
-        for item in relations
-    )
-    assert not any(item.get("kind") == "supersedes" for item in relations)
-
-
-def test_changed_non_primary_requires_reclassification(tmp: Path) -> None:
-    root = tmp / "reclassification"
-    root.mkdir()
-    init_repo(root)
-    write(root, "docs/info.md", "# Operator notes\n\nRun the service with --help.\n")
-    commit_all(root)
-    run(
-        root,
-        "record",
-        "--path",
-        "docs/info.md",
-        "--classification",
-        "guide",
-        "--summary",
-        "Operator notes.",
-    )
-    write(
-        root,
-        "docs/info.md",
-        """# Request contract
-
-## Type
-As-is
-
-## Lifecycle
-Active
-
-## Behavior
-Duplicate requests return the original result.
-""",
-    )
-    impact = run_json(root, "impact", "docs/info.md")
-    changed = next(item for item in impact["changedDocuments"] if item["path"] == "docs/info.md")
-    assert changed["knownClassification"] == "guide"
-    assert changed["requiresClassification"] is False
-    assert changed["requiresReclassification"] is True
-    assert impact["semanticSweepRequired"] is True
-
-
-def main() -> None:
-    with tempfile.TemporaryDirectory(prefix="spec-wiki-maint-") as directory:
-        tmp = Path(directory)
-        test_relation_map_invalidates_verification(tmp)
-        test_impact_detects_coverage_gaps_and_moves(tmp)
-        test_deep_audit_finds_low_signal_committed_docs(tmp)
-        test_relate_restores_source_explicit_relation(tmp)
-        test_changed_non_primary_requires_reclassification(tmp)
-    print("maintenance regression: 5/5 passed")
-
-
-if __name__ == "__main__":
-    main()

package/skills/spec-wiki/evals/retrieval-eval.md

@@ -1,37 +0,0 @@
-# Retrieval regression protocol
-
-Use `retrieval-evals.json` to check the query router after search/index changes.
-The cases are intentionally mixed: direct keywords, natural paraphrases,
-Russian requests, synonyms, exact code paths, and symbols.
-
-## Guard against answer leakage
-
-The retrieval agent must receive only each case `id` and `query`. Do not expose
-`expectedAny` until after it returns candidate paths.
-
-## Stage 1: deterministic candidate generation
-
-Run `spec_wiki.py search` against the current `.spec-wiki/state.json` and record
-top-1/top-3/top-5 plus `confidence` / `catalogFallbackRecommended`. This stage is
-expected to be imperfect for semantic paraphrases; its job is cheap candidate
-generation and reliable low-confidence detection.
-
-## Stage 2: full skill workflow
-
-Have a fresh agent read only `SKILL.md` and `references/query.md`, then for each
-query:
-
-1. create the in-model retrieval bundle;
-2. call relation-aware `search` with `--also` aliases;
-3. semantically rerank compact result metadata;
-4. when fallback is recommended or intent coverage is visibly incomplete, use
-   only `.spec-wiki/index.md` as the semantic catalog fallback;
-5. return up to three primary spec paths without opening the primary specs.
-
-Compare those paths with `expectedAny` only after the run. A case passes top-k
-when any expected path appears in the first k selections.
-
-Current target for this repository: at least 95% top-1 and 100% top-3/recall.
-Do not tune expected paths after seeing a run merely to improve the score; change
-an expectation only when independent review shows that more than one primary
-spec is genuinely defensible.

package/skills/spec-wiki/evals/retrieval-evals.json

@@ -1,37 +0,0 @@
-{
-  "version": 1,
-  "description": "Live retrieval regression cases for the host repository. Expected paths are chosen before running the search; use expectedAny when more than one primary spec is defensible.",
-  "cases": [
-    {"id":"kw-dcp","category":"keyword","query":"dcp persistence recovery","expectedAny":["specs/dcp.md"]},
-    {"id":"kw-desktop-markdown","category":"keyword","query":"desktop markdown mermaid file preview","expectedAny":["specs/desktop-markdown-rendering.md"]},
-    {"id":"kw-project-agents","category":"keyword","query":"project local agents .pi/agents","expectedAny":["specs/project-agents-dir.md"]},
-    {"id":"kw-question-tab","category":"keyword","query":"question inactive tab","expectedAny":["specs/question-inactive-tab.md"]},
-    {"id":"kw-retry-toast","category":"keyword","query":"retry toast action","expectedAny":["specs/retry-toast-action.md"]},
-    {"id":"kw-openai-usage","category":"keyword","query":"openai codex token refresh usage","expectedAny":["specs/openai-codex-usage-refresh.md"]},
-
-    {"id":"para-recovery","category":"paraphrase","query":"recover old conversation after compaction","expectedAny":["external/pi-tools-suite/docs/session-recovery.md"]},
-    {"id":"para-browser-qa","category":"paraphrase","query":"browser qa credentials origin isolation redaction","expectedAny":["external/pi-tools-suite/docs/browser-qa-subagent.md"]},
-    {"id":"para-concurrency","category":"paraphrase","query":"async callback updates the wrong session tab","expectedAny":["docs/concurrency.md"]},
-    {"id":"para-cache","category":"paraphrase","query":"keep tool output stable for prompt caching","expectedAny":["specs/dcp-provider-cache-stability.md"]},
-    {"id":"para-window","category":"paraphrase","query":"window opens offscreen after monitor change","expectedAny":["specs/desktop-window-state.md"]},
-    {"id":"para-copy","category":"paraphrase","query":"copy last assistant response slash command","expectedAny":["specs/desktop-slash-commands.md"]},
-    {"id":"para-task-jsonc","category":"paraphrase","query":"task storage comments jsonc","expectedAny":["docs/desktop-task-manager.md"]},
-    {"id":"para-wrapped-link","category":"paraphrase","query":"click file uri in wrapped markdown","expectedAny":["specs/clickable-markdown-links.md"]},
-    {"id":"para-zed","category":"paraphrase","query":"open source file in zed at line","expectedAny":["specs/file-link-opening.md"]},
-
-    {"id":"ru-recovery","category":"russian","query":"как восстанавливается история после сжатия контекста","expectedAny":["external/pi-tools-suite/docs/session-recovery.md"]},
-    {"id":"ru-zed","category":"russian","query":"как открыть ссылку на файл в zed","expectedAny":["specs/file-link-opening.md"]},
-    {"id":"ru-question-tab","category":"russian","query":"что происходит с вопросом в неактивной вкладке","expectedAny":["specs/question-inactive-tab.md"]},
-    {"id":"ru-tasks","category":"russian","query":"где хранятся задачи проекта и поддерживаются ли комментарии","expectedAny":["docs/desktop-task-manager.md"]},
-    {"id":"ru-provider-seen","category":"russian","query":"почему старый tool result нельзя удалять до отправки провайдеру","expectedAny":["specs/dcp-provider-cache-stability.md"]},
-
-    {"id":"syn-auth","category":"synonym","query":"oauth account rotation when quota exhausted","expectedAny":["specs/antigravity-auth.md"]},
-    {"id":"syn-background-agent","category":"synonym","query":"cancel a running background agent","expectedAny":["specs/async-subagents.md"]},
-
-    {"id":"path-openai","category":"path","query":"src/app/model/model-usage-status.ts","expectedAny":["specs/openai-codex-usage-refresh.md"]},
-    {"id":"path-dcp-journal","category":"path","query":"external/pi-tools-suite/src/dcp/journal.ts","expectedAny":["specs/dcp.md","specs/dcp-reliability-mechanisms.md","specs/dcp-provider-cache-stability.md"]},
-
-    {"id":"symbol-sidebar","category":"symbol","query":"WorkspaceSidebar.svelte tasks","expectedAny":["docs/desktop-task-manager.md"]},
-    {"id":"symbol-question-bridge","category":"symbol","query":"ui-request-bridge question","expectedAny":["specs/desktop-question-tool.md"]}
-  ]
-}