sync: merge upstream/master into master / 同步上游 upstream/master 到 master

.agents/skills/neon/SKILL.md

@@ -0,0 +1,79 @@
+---
+name: neon
+description: >-
+  Repository-specific guidance for working with InferenceX's Neon PostgreSQL
+  databases, connection variables, migrations, and query code. Use when Neon,
+  Postgres, DATABASE_URL, database, schema, migration, or backend data access is
+  mentioned.
+---
+
+# Neon PostgreSQL in InferenceX
+
+Use this skill for database work in this repository. InferenceX runs on Vercel
+and uses Neon as PostgreSQL; do not introduce Neon Auth, Object Storage,
+Functions, AI Gateway, an ORM, or new infrastructure unless the user asks.
+
+## Start with the repository
+
+Read the relevant local documentation before changing code:
+
+- `docs/index.md`
+- `docs/architecture.md` for the API and cache boundaries
+- `docs/data-pipeline.md` for ingestion and schema flow
+- `docs/collectivex.md` for the separate CollectiveX database
+
+Existing database code lives in `packages/db/`. Reuse its connection helpers,
+tagged SQL patterns, migrations, and scripts instead of creating a parallel
+client or configuration layer.
+
+## Connections and credentials
+
+- Main read path: `DATABASE_READONLY_URL`
+- Main administrative writes: `DATABASE_WRITE_URL`
+- CollectiveX read path: `DATABASE_COLLECTIVEX_READONLY_URL`
+- CollectiveX writes and migrations: `DATABASE_COLLECTIVEX_WRITE_URL`
+
+Keep credentials in environment variables. Never print, commit, copy into
+source, or expose connection strings in logs or responses. Use the read-only
+connection for diagnostics and normal reads; use a write connection only when
+the requested task authorizes mutation.
+
+The application uses `@neondatabase/serverless` for serverless reads and
+`postgres` for administrative or transaction-heavy scripts. Preserve that
+split unless runtime requirements clearly demand a change.
+
+## Schema and query changes
+
+1. Inspect the current migration and query code before proposing a schema
+   change.
+2. Add an append-only migration; never rewrite a migration that may have
+   already run.
+3. Keep raw database rows in API responses unless a documented route is an
+   explicit exception.
+4. Make multi-step writes atomic and safe under concurrent Vercel requests.
+5. Verify indexes and bounded query behavior for new filters or ordering.
+6. Add focused query/migration tests and run the repository checks.
+
+Useful commands:
+
+```bash
+bun run admin:db:migrate
+bun run admin:db:migrate:collectivex
+bun run admin:db:verify
+bun run typecheck
+bun run test:unit
+```
+
+Do not run migrations, destructive SQL, or production writes during a review
+or diagnostic request. For authorized schema work, prefer an isolated Neon
+branch and confirm the target before applying changes.
+
+## Current Neon documentation
+
+Neon changes over time. For platform-specific behavior, verify against the
+official documentation index rather than relying on this compact skill:
+
+- https://neon.com/docs/llms.txt
+- https://neon.com/docs/connect/choose-connection
+- https://neon.com/docs/serverless/serverless-driver
+- https://neon.com/docs/introduction/branching

.claude/agents/ingest.md

@@ -23,7 +23,7 @@ You ingest benchmark runs from `SemiAnalysisAI/InferenceX` GitHub Actions into t
 cd /Users/quilicic/InferenceX-app/packages/db
 DATABASE_WRITE_URL='<provided direct non-pooled write URL>' \
 GITHUB_TOKEN=$(gh auth token) \
-pnpm exec tsx src/ingest-ci-run.ts --download <RUN_ID> SemiAnalysisAI/InferenceX
+bun src/ingest-ci-run.ts --download <RUN_ID> SemiAnalysisAI/InferenceX
 ```
 
 Then refresh the materialized view (the script's auto-refresh sometimes races):
@@ -118,7 +118,7 @@ cd /Users/quilicic/InferenceX-app/packages/db
 INGEST_RUN_ID=$RID INGEST_RUN_ATTEMPT=1 INGEST_ARTIFACTS_PATH=$TMPDIR INGEST_REPO=SemiAnalysisAI/InferenceX \
 DATABASE_WRITE_URL='<provided direct non-pooled write URL>' \
 GITHUB_TOKEN=$(gh auth token) \
-pnpm exec tsx src/ingest-ci-run.ts
+bun src/ingest-ci-run.ts
 rm -rf $TMPDIR
 ```
 
@@ -162,7 +162,7 @@ If the user doesn't specify a description, DO NOT skip the entry and DO NOT bloc
 - **Multi-attempt artifacts**: a single GitHub run can spill across runners (`h200-cw_00` + `h200-dgxc-slurm_1`); the logical-name dedup strips the `_<runner>_<attempt>` suffix.
 - **Materialized view dedup tiebreaker**: `latest_benchmarks` picks rows by `date DESC, wr.run_started_at DESC`. Backfilling old data may not surface unless dates align with the user's date picker selection.
 - **Date alignment for partial runs**: when a re-run only covers a subset of concs (`replace ONLY the points this run produces`), align dates with prior full sweep via `UPDATE benchmark_results.date = '<full-sweep-date>'` so the frontend's max-date-per-group dedup doesn't drop the older sweep.
-- **Agentic interactivity normalization (`*_intvty`)**: for `agentic_traces` runs, interactivity MUST be the slow-tail reciprocal of the ITL percentile — `*_intvty = 1/*_itl` (so `p90_intvty = 1/p90_itl`). Some harness versions emit `*_intvty` as `p(1/ITL)` instead (fast-tail — inverts percentile order, e.g. p90 shows ~`1/p10(ITL)`), which silently contaminates cross-run Pareto comparisons. The ingest mapper (`benchmark-mapper.ts`) now **derives `*_intvty` from `*_itl` and discards the artifact's value** for agentic rows, so a normal ingest is self-correcting — no manual step needed. The frontend `agenticAliases` does the same for overlay / `?unofficialrun=` rows. If you ever load agentic data through a path that bypasses the mapper, run `pnpm --filter @semianalysisai/inferencex-db db:backfill-agentic-intvty --yes` (idempotent; rewrites `mean/p75/p90/p95 _intvty = 1/_itl`) then refresh the MV + purge cache. `std_intvty` is intentionally left alone (the reciprocal of a std is meaningless; the API strips it anyway).
+- **Agentic interactivity normalization (`*_intvty`)**: for `agentic_traces` runs, interactivity MUST be the slow-tail reciprocal of the ITL percentile, so `p90_intvty = 1/p90_itl`. Some harness versions emit `p(1/ITL)` instead, which inverts percentile order and contaminates cross-run Pareto comparisons. The ingest mapper derives `*_intvty` from `*_itl` and discards the artifact value for agentic rows. The frontend `agenticAliases` does the same for overlay and `?unofficialrun=` rows. Do not ingest through a path that bypasses these normalizers; the retired one-shot backfill is no longer available. `std_intvty` stays unchanged because the reciprocal of a standard deviation is meaningless, and the API strips it.
 
 ## Process
 
@@ -180,7 +180,7 @@ This agent ingests **benchmark runs**. The HF agentic trace **datasets** (`semia
 
 ```bash
 cd packages/db && DATABASE_WRITE_URL='<direct write url>' \
-  pnpm exec tsx src/ingest-weka-dataset.ts <hf-dataset-id> \
+  bun src/ingest-weka-dataset.ts <hf-dataset-id> \
   [--label "…"] [--variant full|256k] [--description "…"] [--limit N]
 ```
 

.claude/commands/debug.md

@@ -1,5 +1,5 @@
 ---
-allowed-tools: Bash(git log:*), Bash(git diff:*), Bash(git blame:*), Bash(pnpm test:*), Bash(pnpm typecheck*), Bash(pnpm lint*), Bash(pnpm dev*), Bash(curl:*), Read, Glob, Grep
+allowed-tools: Bash(git log:*), Bash(git diff:*), Bash(git blame:*), Bash(bun run test:*), Bash(bun run typecheck*), Bash(bun run lint*), Bash(bun run dev*), Bash(curl:*), Read, Glob, Grep
 description: Systematic debugging — root cause before fixes
 ---
 
@@ -57,7 +57,7 @@ BEFORE attempting ANY fix:
 
 1. **Create Failing Test** — regression test reproducing the bug with exact triggering input (per CLAUDE.md testing requirements)
 2. **Implement Single Fix** — address root cause, ONE change, no "while I'm here" improvements
-3. **Verify Fix** — run `pnpm test:unit` and `pnpm typecheck`, confirm no other tests broken
+3. **Verify Fix** — run `bun run test:unit` and `bun run typecheck`, confirm no other tests broken
 
 ## Red Flags — STOP and Return to Phase 1
 

.claude/commands/fix.md

@@ -1,5 +1,5 @@
 ---
-allowed-tools: Bash(pnpm build*), Bash(pnpm typecheck*), Bash(pnpm lint*), Bash(pnpm fmt*), Read, Edit, Glob, Grep
+allowed-tools: Bash(bun run build*), Bash(bun run typecheck*), Bash(bun run lint*), Bash(bun run fmt*), Read, Edit, Glob, Grep
 description: Incrementally fix build, type, and lint errors with minimal safe changes
 ---
 
@@ -12,14 +12,14 @@ Incrementally fix build and type errors with minimal, safe changes.
 Run all checks and capture errors:
 
 ```bash
-pnpm typecheck 2>&1
-pnpm lint 2>&1
+bun run typecheck 2>&1
+bun run lint 2>&1
 ```
 
 If both pass, run the full build:
 
 ```bash
-pnpm build 2>&1
+bun run build 2>&1
 ```
 
 If everything passes, announce "All checks pass — nothing to fix." and stop.
@@ -47,7 +47,7 @@ For each error:
 - A fix introduces **more errors than it resolves**
 - The **same error persists after 3 attempts**
 - The fix requires **architectural changes** or touching >3 files
-- Errors stem from **missing dependencies** (need `pnpm install`)
+- Errors stem from **missing dependencies** (need `bun install`)
 
 ## Step 5: Summary
 

.claude/commands/verify.md

@@ -1,5 +1,5 @@
 ---
-allowed-tools: Bash(pnpm test:*), Bash(pnpm typecheck*), Bash(pnpm lint*), Bash(pnpm build*), Bash(pnpm dev*), Bash(pnpm fmt*), Bash(curl:*), Bash(git diff:*), Bash(git status*), Read, Glob, Grep
+allowed-tools: Bash(bun run test:*), Bash(bun run typecheck*), Bash(bun run lint*), Bash(bun run build*), Bash(bun run dev*), Bash(bun run fmt*), Bash(curl:*), Bash(git diff:*), Bash(git status*), Read, Glob, Grep
 description: Verify work is complete before committing — evidence before claims
 ---
 
@@ -34,39 +34,39 @@ Run each step. Report the actual output — not what you expect.
 ### 1. Type checking
 
 ```bash
-pnpm typecheck
+bun run typecheck
 ```
 
 Required: exit 0, no errors
 
 ### 2. Linting
 
 ```bash
-pnpm lint
+bun run lint
 ```
 
 Required: exit 0, no errors
 
 ### 3. Formatting
 
 ```bash
-pnpm fmt
+bun run fmt
 ```
 
 Required: exit 0, no formatting issues
 
 ### 4. Unit tests
 
 ```bash
-pnpm test:unit
+bun run test:unit
 ```
 
 Required: all tests pass, 0 failures
 
 ### 5. Dev server starts
 
 ```bash
-pnpm dev --hostname 0.0.0.0 --port 3000 &
+bun run dev -- --hostname 0.0.0.0 --port 3000 &
 curl --retry 10 --retry-delay 2 --retry-connrefused -sSf http://localhost:3000 >/dev/null
 ```
 
@@ -75,7 +75,7 @@ Required: server responds successfully
 ### 6. E2E tests
 
 ```bash
-pnpm test:e2e
+bun run test:e2e
 ```
 
 Required: all tests pass

.claude/commands/write-plan.md

@@ -58,14 +58,14 @@ Use this format:
       [Exact test code]
 
 - [ ] **Step 2: Run test to verify it fails**
-      Run: `pnpm test:unit -- path/to/test`
+      Run: `bun run test:unit -- path/to/test`
       Expected: FAIL
 
 - [ ] **Step 3: Write minimal implementation**
       [Exact implementation code]
 
 - [ ] **Step 4: Run test to verify it passes**
-      Run: `pnpm test:unit -- path/to/test`
+      Run: `bun run test:unit -- path/to/test`
       Expected: PASS
 
 - [ ] **Step 5: Commit**

.claude/settings.json

@@ -12,7 +12,7 @@
         "hooks": [
           {
             "type": "command",
-            "command": "pnpm fmt:fix; pnpm lint:fix; true",
+            "command": "bun run fmt:fix; bun run lint:fix; true",
             "timeout": 10,
             "async": true
           }

.claude/skills/write-inferencex-blog/SKILL.md

@@ -321,7 +321,7 @@ The editor runs as a background Node process on `127.0.0.1:4747`, reads from and
 
 While the user reviews in the browser, you can:
 
-- Run `pnpm lint && pnpm typecheck` against the working tree to catch any MDX errors that would block the pre-commit hook later.
+- Run `bun run lint && bun run typecheck` against the working tree to catch any MDX errors that would block the pre-commit hook later.
 - Save the chart image into `packages/app/public/images/{slug}/benchmark-light.png` (and `benchmark-dark.png` if the user provided both) so the `<Figure>` placeholder in the preview shows a real path.
 
 **Concurrent-edit collision warning.** The browser editor auto-saves the user's textarea ~800 ms after their last keystroke. If you re-edit a paragraph the user has open in CodeMirror, your `Edit` call writes to disk first, then the editor's debounced save overwrites your change with the user's stale buffer the next time they type or the timer fires. Failure mode: user asks you to expand a paragraph, you expand it on disk, user types one more character in the browser, the one-liner comes back. When you need to edit a section the user is actively working on, **tell the user explicitly to either close the browser tab or hit the "↻ Reload from disk" button before resuming editing**. Don't rely on them noticing the collision — it looks like nothing happened from their side.
@@ -342,7 +342,7 @@ git push -u origin blog/{slug}
 gh pr create --title "feat(blog): ..." --body "..."
 ```
 
-The pre-commit hook runs `oxlint`, `oxfmt`, and `tsc --noEmit`. All three must pass. If lint/format fails, run `pnpm lint:fix && pnpm fmt:fix` and re-commit (don't `--no-verify`).
+The pre-commit hook runs `oxlint`, `oxfmt`, and `tsc --noEmit`. All three must pass. If lint/format fails, run `bun run lint:fix && bun run fmt:fix` and re-commit (don't `--no-verify`).
 
 After the PR opens, expect Cursor Bugbot to flag correctness issues in the prose (numeric overstatement, claims contradicted by tables, wrong attribution). Treat its findings as real review comments — fix them in a follow-up commit, then resolve the threads. Branch protection on master requires resolved review threads before auto-merge fires.