Research cluster: pstack, noodle and the 2,500-PR talk

Checked 2026-10-03. Each claim below starts as a tier-3 hypothesis from the video and is checked
against primary sources fetched during this session.

Author claim (Tier 2/3)

Research execution

Repo discovery (GitHub API through gh, read-only):

• gh api users/poteto/repos?sort=pushed lists no pstack repo. gh api repos/poteto/pstack returns
  404.
• gh search repos pstack returns about 20 community ports (Claude Code, Codex, Pi, OMP, OpenCode,
  ZCode, Hermes, ChatGPT). Several name their upstream as cursor/plugins/pstack.
• gh api repos/cursor/plugins/contents/pstack is the canonical tree. I read README.md,
  .cursor-plugin/plugin.json, LICENSE, the skills/ listing (47 entries), agents/, and
  automations/. I also read these files in full or in part: skills/recall/SKILL.md,
  skills/poteto-mode/playbooks/autopilot-full.md, skills/swarm/SKILL.md,
  skills/create-verification-skill/SKILL.md, and the frontmatter of automate-me, unslop, how
  and why. From the full recursive tree (158 files) I listed every file that is not Markdown.
  From docs/guide/ I grepped README.md, 06-verify-and-ship.md and 07-overnight.md.
• gh api repos/cursor/plugins/contents/cursor-team-kit/skills lists the skills. I read
  control-ui/SKILL.md.
• gh api repos/poteto/verification-skill-example: read the repo metadata, the file tree,
  README.md and .cursor/skills/verify-atlas/SKILL.md.
• gh api repos/poteto/noodle: read the metadata, the root listing, brain/, .agents/skills/,
  skills/, README.md, the latest commits and the releases.
• gh api repos/mattpocock/skills: read the metadata and searched the tree for grill and
  wayfind SKILL.md paths.

Talk and X post:

• WebSearch poteto Lauren Tan "2,500" PRs talk found nothing relevant. An extended WebSearch
  found the X status x.com/poteto/status/2102050467505430555, an AICoder news item, and a
  third-party YouTube re-upload.
• WebFetch of the X status returned HTTP 402.
• Applied the x:read skill. The URL passed its gate and was rebuilt as
  https://x.com/poteto/status/2102050467505430555. Step 1 POSTed to xtomd.com/api/fetch
  (curl exit 0, HTTP 200) and returned the post text, createdAt, and engagement counts. The
  spool file was deleted after reading.
• WebFetch of the YouTube oEmbed for NjoZoUm85x0 returned the uploader's name.
• WebFetch of the AICoder article returned the headline. Its slug date is 20260921.
• WebFetch of cursor.com/marketplace/cursor/pstack returned only page chrome. The page renders
  with JavaScript, so it is unreached.

Recency gate (as of 2026-10-03):

Consensus

C1. Confirmed, with one nuance.

• The X post by @poteto reads "here's how i shipped 2,500 PRs last month to production". It says the
  talk "was originally supposed to be for Cursor Compile in London ... so i'm making it available
  for free here on X". It carries a video.
• createdAt is Mon Sep 21 15:01:00 +0000 2026. That is 11 days before the stream, which went live 2026-10-02 (release_timestamp
  2026-10-02T15:59Z; 2026-09-30 is only when the scheduled stream was created), and it still
  matches "about 10 days ago".
• Views were 2,939,751 when read on 2026-10-03. The figure comes from X's counter, relayed by the
  xtomd converter. The other counts were likes 14,315, reposts 1,172, replies 520 and bookmarks
  27,140.
• The count at stream time could not have been higher than this figure, so "about 3 million" is
  a round-up but consistent.
• A third party (uploader "Raner") posted a YouTube re-upload titled "How I Shipped 2000 PRs Last
  Month". That title is the re-uploader's and does not contradict the author's own figure of 2,500.
• Settled by: https://x.com/poteto/status/2102050467505430555, through xtomd.

C3. Confirmed.

• poteto/noodle exists. It is MIT licensed, described as "Orchestrate agents using skills", and
  written in Go.
• Its root has a brain/ directory containing architecture.md, principles.md, principles/,
  vision.md, todos.md, plans/ and codebase/.
• It also has .agents/skills/, with 29 entries including unslop, reflect, brain,
  adversarial-review and execute. A top-level skills/ holds only skill-creator.
• The claim that noodle "became the basis of" pstack is the author's own statement. The timeline is
  consistent with it: noodle dates from February to March 2026, and pstack's first commit is
  2026-05-23. Some skill names recur in both, such as unslop and reflect.
• Settled by: https://github.com/poteto/noodle.

C11. Partly confirmed. The bundled Playwright/CDP CLI is not public, so that part is unverified.

• What public pstack contains.
  • pstack ships no Playwright or CDP CLI.
  • create-verification-skill is a generator. It writes a project-local
    .cursor/skills/verify-<app>/ with Launch, Doctor, Drive, Evidence and Cleanup sections plus a
    features/ map. It picks "browser/CDP for web and Electron, a tmux/PTY harness for CLI/TUI,
    plain HTTP for services".
  • maintain-verification-skill keeps that feature map honest.
  • The browser-driving skill control-ui ships in cursor-team-kit, not pstack. It holds only
    inline Playwright chromium.launch() / connectOverCDP() snippets and no packaged CLI.
• The example repo shows the pattern. poteto/verification-skill-example has a skill that
  documents a CDP driver CLI, control-atlas.mjs, with doctor, snapshot, screenshot,
  click, trace, profile, network-log and similar commands.
  • Its README says "CLI omitted on purpose; usage only in SKILL.md" and "Private reference.
    Heavily fictionalized".
  • The repo has no license.
  • So the pattern of a skill wrapping a CDP CLI is confirmed. The CLI itself, and its Playwright
    glue, are unverified.
• The "thin wrapper over deterministic tools" principle is confirmed in public pstack.
  • skills/poteto-mode/scripts/ bundles a tested TypeScript watch-pr CLI (cli.ts,
    github.ts, policy.ts and *.test.ts), orch/ with orch.test.ts, check-plan.mjs and
    worktree-audit.sh.
  • show-me-your-work/scripts/log.sh is another bundled script.
  • The principle-build-the-lever skill states the rule as "Build the tool that does it or
    proves it (codemod, script, generator...)".
• Settled by: https://github.com/cursor/plugins/tree/main/pstack/skills/create-verification-skill,
  https://github.com/cursor/plugins/tree/main/cursor-team-kit/skills/control-ui and
  https://github.com/poteto/verification-skill-example.

C27. Mechanism confirmed, with corrections to the detail.

• playbooks/autopilot-full.md exists, and the README describes it as "run independent prs to
  merged with one owner per pr and a root swarm verdict on each round". The guide's example
  prompt is "/poteto-mode full autopilot on this queue".
• The playbook says to "fan out parallel independent verifiers per the swarm skill and aggregate
  to one verdict". It requires a lane that will "Prove the load-bearing behavior live on the real
  surface the change touches (with ... control-ui...)" and a "Regression lane against trunk".
• Correction: the playbook does not use the word "fuzz". Its lanes are re-running the gates,
  a live proof on the real surface, two or more diff-audit lanes, and a regression run on trunk.
• Correction: the verifiers do not fix. Proven findings go "to the owner in one fix-forward",
  and the per-PR owner agent squash-merges after a clean verdict.
• Tunable: confirmed. The swarm skill says "Set N from the user or derive it from the shape".
• "Token intensive": unverified. No source I read states it.
• Settled by:
  https://github.com/cursor/plugins/blob/main/pstack/skills/poteto-mode/playbooks/autopilot-full.md
  and https://github.com/cursor/plugins/blob/main/pstack/skills/swarm/SKILL.md.

C30. Confirmed for recall. The mining-for-corrections part belongs to a different skill.

• skills/recall/SKILL.md describes itself as "Reconstruct your recent working context from your
  own chat history, live state, and the shared record". It reads Cursor transcripts at
  ~/.cursor/projects/<slug>/agent-transcripts/<uuid>/<uuid>.jsonl, fans out cheap subagents over
  them, and returns a brief.
• Turning habits and corrections into a skill is automate-me: "optionally pulling fresh
  evidence from recent transcripts" and drafting a <name>-mode skill. Recall's own step 1 routes
  that work away: "Turning habits into a durable skill is automate-me".
• The advice to mine transcripts for corrections and turn them into lint rules is opinion. No
  source settles it.
• Settled by: https://github.com/cursor/plugins/blob/main/pstack/skills/recall/SKILL.md.

C32. The named skills are confirmed to exist. Whether they compose well is opinion.

• mattpocock/skills (MIT) contains skills/productivity/grill-me/,
  skills/engineering/grill-with-docs/ (the ASR "grill me with docs" at [62:59]),
  skills/engineering/wayfinder/ and skills/productivity/grilling/.
• pstack ships poteto-mode with 23 playbooks. "Execution skills" is not a category pstack uses.
• Settled by: https://github.com/mattpocock/skills and
  https://github.com/cursor/plugins/tree/main/pstack/skills/poteto-mode.

pstack inventory as it exists now (v0.15.5):

• Canonical location. github.com/cursor/plugins/tree/main/pstack, inside Cursor's official
  plugins monorepo. poteto/pstack does not exist. The X bio links
  cursor.com/marketplace/cursor/pstack.

• Install and license. /add-plugin pstack, a Cursor command. MIT licensed, "Copyright (c) 2026
  Lauren Tan".

• Harness. Cursor only:

  • the manifest is .cursor-plugin/plugin.json;
  • the skills write to .cursor/skills/ and read ~/.cursor/rules/pstack-models.mdc;
  • recall reads ~/.cursor/projects/... transcripts;
  • swarm spawns Cursor cloud workers.

  Claude Code, Codex and other versions are community ports. GitHub reported these star counts
  on 2026-10-03:

  • michael-denyer/pstack-claude: 779 stars;
  • ericlitman/open-pstack: 349 stars;
  • adjohn/pstack: 3 stars.

• Structure.

  • skills/<name>/SKILL.md, with optional references/ and scripts/ directories.
  • agents/poteto-agent.md and agents/comment-sicko.md.
  • automations/benny/, a dormant pack for triaging Slack bug reports.
  • docs/guide/, with 10 chapters.
  • Every skill read sets disable-model-invocation: true. poteto-mode is the router that
    invokes the others.

• Skills (47). 24 working skills and 23 principle skills:

  • Working skills: poteto-mode (23 playbooks, including autopilot-full, autopilot-stack,
    orchestrate and hillclimb), how, why, recall, blast-radius, architect, arena,
    swarm, interrogate, automate-me, make-bot-ui, setup-pstack, reflect, teach,
    tdd, no-comments, typescript-best-practices, figure-it-out, show-me-your-work,
    create-verification-skill, maintain-verification-skill, unslop, bro and
    technical-writing.
  • Principle skills: one per principle, each named principle-<name>.

• Not bundled. deslop, control-cli and control-ui ship in cursor-team-kit.

Staleness

• pstack is fast-moving.
  • 90 commits in about 4 months. The README carries an upgrade note for rules "written before
    0.15.3".
  • The default models changed on 2026-09-23 (re-anchor: tighten-your-output routes to compress/simplify without presence gate or fallback #414: "default to Opus 5.5 and Grok 4.7"). The same
    day, claude-memory: orphan-rule-check hardcodes .work/ exclusion instead of resolving memory_dir from topic-docs seam #419 cut "19 more instructions Opus 5.5 does not need".
  • Skill counts and playbook names can change within weeks. Pin reads to 12d587dfb / v0.15.5.
• noodle has been dormant since 2026-03-19. It is historical evidence of lineage, not a living
  reference.
• View counts change continuously. The 2,939,751 figure is a 2026-10-03 snapshot.
• The verification-skill-example has not changed since it was created on 2026-07-30. Its README
  marks it a fictional reference.
• The video's description of the verification CLI and of autopilot fuzzing describes Cursor-internal
  tooling. The public playbook may diverge from it, and already does: verifiers report, owners fix.

Promoted tier

Repo implication

These are pstack structures a Claude Code plugin marketplace could borrow. pstack is MIT licensed,
so code or text copied from it needs attribution.

• One router skill. poteto-mode keeps 23 playbooks as separate files under playbooks/.
  Every other skill sets disable-model-invocation: true and runs only when the router or a user
  invokes it.
• Single-rule principle skills. Each of the 23 principles is its own skill, indexed inline in
  the router, so other skills can cite a principle by name.
• Tested helper scripts inside the skill. For example, poteto-mode/scripts/watch-pr/*.ts
  ships with its own *.test.ts, which moves the deterministic work out of the prompt.
• A generator and maintainer pair for verification. create-verification-skill writes a
  project-local verify skill with a per-feature map and proves it once end to end before handing
  it over. maintain-verification-skill audits drift and ends in one of three outcomes: clean,
  changed or blocked.
• Verification by a separate agent before merge. autopilot-full gives each PR one owner and
  requires a swarm verdict on the exact SHA before merge. The guide's description of the Shipping
  playbook adds that "the agent that judges a change is never the one that wrote it".
• Not checked. I have not verified that disable-model-invocation, Cursor cloud subagents or
  .cursor/rules model routing behave the same way in Claude Code. The community Claude Code ports
  exist because these Cursor primitives had to be translated.
• Shape only for the example repo. verification-skill-example is unlicensed, so borrow only
  its layout idea: a SKILL.md documenting a CLI's command surface, plus a features/*.md map.

Sources

Repo touchpoints: poteto / pstack practices vs claude-code-plugins

Target: worktree melodic-software-claude-code-plugins-video-digest-MN9dGgmLyso at c8fa858c9.
Every path below is repo-relative and was read or grepped this session. Excerpts are fragments of
one line. Classifications: covered, partial, absent, conflicts.

Cross-cutting fact: the repo already audited pstack at main@60c641e4
(docs/upstream/cursor-pstack.md:9). Its attribution table has rows for why, recall, teach,
tdd, reflect, arena, technical-writing, blast-radius, unslop and show-me-your-work
(docs/upstream/cursor-pstack.md:31-40). That record has no row or section for a verification
skill or for "full autopilot" (grep of cursor-pstack.md for verif|autopilot: no autopilot
hit, and every verif* hit is audit-method prose or the recall text), so the record never
evaluated those two pstack pieces.

1. Verification as the core skill (C8, C11)

Classification: partial

Evidence:

• plugins/playwright/skills/playwright/SKILL.md:20: "Wraps Microsoft's @playwright/cli for token-efficient browser automation."
• plugins/playwright/skills/playwright/SKILL.md:59: "Trace recording for debugging, video recording" (reference row).
• plugins/testing/skills/run-e2e/SKILL.md:92: "Autonomous live verification of a running application: start it, navigate, interact, screenshot, assert."
• plugins/testing/skills/run-e2e/SKILL.md:128: "Delegate the drive loop to a subagent ... returning only the evidence paths."
• plugins/testing/skills/run-e2e/SKILL.md:23: "the bundled skill itself defers to a project skill of that name"
• plugins/testing/skills/run-e2e/SKILL.md:153-154: names the native /verify and /run-skill-generator as the per-app surfaces.
• plugins/verification/skills/confirm/SKILL.md:38: outcome verification is "intent match + evidence + (runtime) /testing:run-e2e".
• Grep of run-e2e/context/e2e.md and e2e-config.md for reus|saved|persist|regression baseline|codegen found no persisted flow mechanism.

Gap: The agent can drive the app and capture traces through a wrapped third-party CLI. The repo has no per-app verification skill or glue CLI of its own, and grep found no mechanism that keeps verification flows between runs; it leaves per-app skills to the native run//run-skill-generator.

2. Skills as thin wrappers over deterministic scripts (C11, C13)

Classification: covered

Evidence:

• plugins/discipline/skills/script-the-deterministic-work/SKILL.md:42-47: "Deterministic ... Script it, run it" / "Detect-then-judge ... Script only the detect half".
• plugins/playbooks/skills/skill-authoring/SKILL.md:155: "Give it code | Helper scripts > prose instructions"
• plugins/source-control/skills/babysit-prs/SKILL.md:25: "deterministic gate scripts; judgment stays with the agent."
• plugins/testing/skills/audit/SKILL.md:14: "skill runs a deterministic script detector over the suite."
• plugins/autonomy/reference/routines/tech-debt-sweep.md:36-39: "codemod-style transformations that need no judgment, runs with no agent session".
• Count (loop over plugins/*/skills/*/SKILL.md): 113 of 331 skills have a sibling scripts/ dir. This shows that scripts exist. It does not show that those skills are thin wrappers.

Gap: The discipline and the authoring rule both exist. Codemods appear only as the deterministic half of one routine contract, and no migration or codemod skill exists.

3. Observed agent mistakes become lint rules, types or hooks (C14, C16, C25)

Classification: partial

Evidence:

• plugins/review/skills/audit-enforceability/SKILL.md:2: "map it through the crosswalk to the cheapest enforcement rung ... It proposes a rung and never implements one."
• plugins/review/skills/audit-enforceability/SKILL.md:28: "It reads one findings file, classifies each row"
• plugins/harness-config/skills/audit-automation-gaps/SKILL.md:34: "The enforcement hierarchy (strongest first): compiler settings → static analyzers/linters →"
• plugins/session-flow/skills/retro/reference/ecosystem-improvement-catalog.md:113: "A quality gate was missed that could be enforced by a hook"
• docs/upstream/cursor-pstack.md:35 (reflect row): "Rejected — the structural-enforcement check: harness-config:audit-automation-gaps owns it with a default-REJECT posture."
• docs/plugin-philosophy.md:923: "A new standing instruction requires observed, repeated stumble evidence"
• docs/plugin-philosophy.md:940-942: "a hook that corrects model behavior is an ablation candidate like any prose instruction."

Gap: The rung crosswalk reads only a review-findings file and never implements a rung. Mistakes seen in sessions (retro, running-retro ledgers) have no route into it. plugin-philosophy.md:940-942 is in tension with the practice, but codebase lint rules are policy and fall in the durable tier, so it is not a conflict.

4. Outer loop triggers and coordinator agents (C20, C21, C22)

Classification: partial

Evidence:

• plugins/autonomy/reference/trigger-dispatch.md:3-4: "adapters normalize signals from four surface classes into the governed work-item queue"
• plugins/autonomy/reference/trigger-dispatch.md:6: "every concrete instance (which surfaces exist," is an org-binding outcome.
• plugins/autonomy/reference/trigger-dispatch.md:22-23: channel-feed carries open gaps: "whether a channel-monitor may ambiently initiate work".
• plugins/autonomy/skills/setup/templates/trigger-adapters.md:44: example "on: issues types labeled/assigned" enqueue adapter.
• plugins/work-items/skills/work-loop/SKILL.md:2: "Drain the work-item backlog as a self-paced autonomous loop".
• plugins/implementation/skills/implement-dispatch/SKILL.md:15: "the main window orchestrates instead of editing."
• plugins/work-items/skills/work/SKILL.md:189: "The orchestrator never edits source itself."
• plugins/work-items/skills/triage/SKILL.md:109: "when this item shares one underlying decision with other open items".
• plugins/autonomy/reference/routines/duplicate-detection-sweep.md:34: "Advisory report. One run report: link candidates, clusters".

Gap: The outer loop is covered at the contract level, though the chat-channel class is unverified. Coordinators that never edit source are built. Triage clusters items by shared decision and the duplicate sweep finds copies, but nothing sends a burst of related reports to one coordinator to find a shared root cause.

5. Routines log findings to a buffer for later batch review (C24)

Classification: partial

Evidence:

• plugins/session-flow/skills/running-retro/SKILL.md:2: "files findings from the transcript so far into a running ledger, routed, never applied"
• plugins/autonomy/reference/routines/tech-debt-sweep.md:11: "periodically characterizes debt hotspots and files them as work items"
• plugins/autonomy/reference/routines/duplicate-detection-sweep.md:36: "No direct change. Nothing lands in the repository"
• plugins/review/skills/fanout/SKILL.md:46: "Review modes report; a separate fix action applies."
• plugins/review/skills/fanout/SKILL.md:69: fix consumes "the merged set of unconsumed persisted findings".

Gap: Several buffers exist (ledgers, advisory reports, persisted findings, queued work items). fanout fix groups one branch's unconsumed findings by finding class, and audit-enforceability reads one file. Nothing reads the buffer across branches or runs to name a shared cause and fix the environment once.

6. Review by sampling and self-merging agents behind a verification gate (C25, C26, C27)

Classification: partial

Evidence:

• plugins/source-control/skills/babysit-prs/SKILL.md:43-44: worker "merges PRs the merge gate proves 100% ready"; autopilot "merge everything the gate proves ready".
• .claude/source-control.md:39,43: this repo runs worker with merge rung c3-autonomous.
• plugins/autonomy/reference/guardrails/work-classes.md:146: "0 human-reverted merges is a correctness signal, not an acceptance rate."
• plugins/work-items/reference/pipeline-shape.md:46: Verify lane uses "a fresh-context verifier, distinct from the implementer."
• plugins/work-items/reference/pipeline-shape.md:79: the tier-reading dispatcher "is a later #513 stage" (verifier depth is placeholder-only).
• Grep of babysit-prs/scripts/babysit_merge.py for e2e|verif|playwright|run-|smoke|runtime: the only "verification" is commit-signature verification (:374), and there is no app-driving step.
• Grep for review-by-sampling (\bsampl(e|ing)\b) over source-control, review, autonomy and loop-lane: no practice hit.
• AGENTS.md:28 asks before merging, but the same paragraph says a loop prompt that explicitly authorizes it covers it, so it does not contradict c3-autonomous.

Gap: Agents already merge their own PRs here up to C3, behind a gate on CI, review threads and rulesets. The gate never runs fuzzing verifiers against the live app, verifier depth (one verifier or ten) is not built, and there is no post-merge sampling review. Convention boundary: docs/conventions/loop-lane/README.md:50-52, "C4 (structural) and C5 (untrusted-provenance) stay human-gated unconditionally", rules out a full dark factory.

7. Mining past transcripts for corrections; carrying context forward (C30)

Classification: partial

Evidence:

• plugins/session-flow/skills/retro/context/codify.md:11: "A correction happened that should prevent the same mistake next time"
• plugins/session-flow/skills/retro/SKILL.md:107: "aggregates metrics across every chained transcript"
• plugins/session-flow/skills/retro/scripts/parse_transcript.py:182: collects "tool_rejections" per transcript.
• docs/upstream/cursor-pstack.md:35 (reflect row): "an accepted new-skill candidate now goes to /playbooks:skill-authoring" from both retro skills.
• plugins/session-flow/skills/find-handoff/SKILL.md:2: "a bounded scan of recent transcripts" (recovers a save-point only).
• docs/upstream/cursor-pstack.md:32: pstack recall row, "Omitted".
• docs/upstream/cursor-pstack.md:157: "the stumble evidence the scope-lock rule and any transcript-mining step both currently lack."

Gap: Per-session corrections already route to skills and memory. No skill mines many past transcripts for repeated corrections. Carrying context forward from past chats (recall) was declined by a recorded, evidence-gated decision, which is not a prohibition.

8. Skills shrink toward pure workflow as models improve (C31)

Classification: covered

Evidence:

• docs/plugin-philosophy.md:928-932: "Generation-triggered ablation. Instructions are disposable per model generation."
• plugins/harness-config/skills/unhobble/SKILL.md:2: "strip a repo's standing instructions, log stumbles ... restore only those with repeated same-cause evidence"
• plugins/harness-config/skills/unhobble/SKILL.md:44: strip scope includes ".claude/rules/, .claude/skills/, .claude/agents/".
• plugins/harness-config/skills/audit-instructions/SKILL.md:16: "prescriptive step lists that now constrain more than they help"
• .claude/rules/skill-bodies-state-current-rules.md:2: "point at the live upstream source for any volatile specific instead of restating it"
• docs/plugin-philosophy.md:933: "Evals outlive instructions."

Gap: Ablation works on whole surfaces (a file, a skills dir, an enabled plugin). Trimming implementation detail inside a skill while keeping its workflow is a static pass in audit-instructions. Per-skill evals act as a safety net for deletions, not as a strip-and-measure loop.

9. Words as compressed intent: grilling and named concepts (C6)

Classification: covered

Evidence:

• plugins/planning/skills/interview/SKILL.md:2: triggers include "'grill me'".
• plugins/planning/skills/interview/SKILL.md:100: for a domain term, "invoke /domain-driven-design:curate-language".
• plugins/domain-driven-design/skills/curate-language/SKILL.md:2: "maintain a consuming project's ubiquitous-language glossary".
• plugins/testing/skills/test-value/SKILL.md:2: "'is this test tautological'"
• plugins/testing/skills/audit/SKILL.md:2: finds "tautological, vacuous, or assertion-free tests".
• docs/upstream/mattpocock-skills.md:23: "batch-grill-me / grilling rounds | planning:interview ... Derived (behavior)".

Gap: No material gap. Grilling and named-concept vocabulary both ship, and the glossary is kept per project.

10. Composing third-party skill libraries with this marketplace (C32)

Classification: covered

Evidence:

• docs/upstream/cursor-pstack.md:16: "Every entry here is a reauthor, not a fork and not a vendored baseline."
• docs/upstream/cursor-pstack.md:18-19: unlike "playbooks's boris pack, which vendors a verbatim upstream copy".
• docs/upstream/mattpocock-skills.md:3-4: "Single source of truth for everything in this marketplace derived from mattpocock/skills".
• docs/upstream/mattpocock-skills.md:22: "wayfinder | planning:wayfind | Partial".
• plugins/playwright/skills/playwright/vendor/SKILL.md: a vendored upstream skill (file present).
• docs/conventions/native-references/README.md:5-6: "a read-time presence gate that routes, never an assertion that the native thing is there."
• docs/conventions/seam-phrasing/README.md:49: carve-out for installing "a plugin from ANOTHER marketplace".
• plugins/harness-ops/skills/audit-native-overlap/SKILL.md:18-19: overlap audit targets native surfaces ("a native surface never shadows ours").

Gap: Upstream skills are absorbed with recorded provenance, which matches the video's "own your knives". Nothing covers a consumer who installs pstack or mattpocock/skills next to this marketplace: no overlap audit and no routing rule for colliding triggers such as grilling or TDD.

Summary