# Provenance ## Upstream scope The weekly sync is run from `docs/agents/upstream-sync.md`; `npm run upstream:drift` reports how far each upstream moved past its pin. The mattpocock/skills upstream has its own section below. Parity reviews and the weekly upstream sync cover three cursor/plugins roots (decision Q6): `pstack/skills`, `pstack/agents` and `pstack/docs`. Two roots are out of scope: - `pstack/automations/**`: Cursor Automations with Slack event triggers (the benny triage and reproduce-and-fix pack). They need a scheduler or event trigger, and pstack has no scheduler, timer or ledger. - `pstack/.cursor-plugin/**`: the Cursor plugin manifest, which is platform-specific. This package's manifest is the `pi` block in `package.json` plus `pi-extension/pstack/*.ts`. The top-level `pstack/README.md`, `LICENSE`, `.gitignore`, `assets/` and `plugin.json` are unscoped: none of them is ported. `pstack/plugin.json` is the Agent Plugins v1 root manifest added at this pin; it is platform-specific in the same way as `pstack/.cursor-plugin`. `pstack/LICENSE` is also watched, because `THIRD_PARTY_NOTICES.md` reproduces its MIT notice. The fixture records its SHA-256 at the pin. With a checkout, the scope test fails when the upstream file at the recorded commit no longer matches that hash. Without one, it fails unless a `text` block in `THIRD_PARTY_NOTICES.md` hashes to the same value, so the hash cannot move until the notice reproduces the new `LICENSE` text byte for byte. The sync PR body should still call out any upstream `LICENSE` change. `pstack/skills` is accounted for by `docs/skill-inventory.json` and `test/fixtures/skill-provenance/`. Its `root` stays `pstack/skills`, so a sibling path such as `../agents/comment-sicko.md` is relative to it. `test/fixtures/upstream-scope.json` holds one pin record per upstream (ADR-0003): its `upstream` key in the inventory's `sources`, the skills `folder` it maps to, and its `repository`, `commit` and `root`. Every folder under `skills/` must map to exactly one record. The `cursor` record maps `skills/pstack/` to cursor/plugins at `root` `pstack`, and records the scope and the status and SHA-256 of every file under `pstack/agents` and `pstack/docs` at the cursor pin `1ba91fc25c0977741dd0e267849e6234b037ed0a`. The checks below run once per record, with that upstream's checkout in `PSTACK__SOURCE`. To move skills between folders, run `node scripts/move-skills.mjs ...`: it moves them with `git mv` and rewrites the `../` links between skills, which `test/skill-tree.test.ts` requires to resolve. With `PSTACK_CURSOR_SOURCE` set, `npm test` checks that the file list and every hash match the pin, that every top-level `pstack/` entry is classified, and that every file under `pstack/skills` at the pin is referenced by the inventory or a skill-provenance fixture. Without it, `npm test` still checks the fixture against the skill-provenance fixtures and against the table below: a path that a skill-provenance cursor source resolves to must be `carried`, and each row's status must match the fixture. The checkout must hold the commit's blobs (a full clone, or a partial clone that already fetched them): the tests and the script disable lazy fetching. On a pin move, `npm run fixtures:upstream-scope -- ` (for example `cursor`) rewrites that upstream's record at its own `commit`; the checkout defaults to `PSTACK__SOURCE`, and the older `npm run fixtures:upstream-scope -- ` form still means `cursor`. It re-hashes every file, keeps each existing status, adds new paths as `not-carried`, and prints what was added, removed or changed, any unclassified top-level entry and any watched file that changed. It never updates a watched hash, and it exits non-zero when a top-level entry or a watched file needs a person. A new `not-carried` path then fails the table check until it gets a row here. | Upstream path | Status at the pin | | --- | --- | | `pstack/agents/comment-sicko.md` | carried: context source for `skills/pstack/no-comments/references/comment-sicko.md` (adapted from mimir's copy; Wave 3) | | `pstack/agents/poteto-agent.md` | excluded: a named role; pstack ships no named roles (Wave 2: role removal) | | `pstack/docs/guide/**` (11 Markdown pages, 6 images) | excluded: under decision Q1 | | `pstack/docs/harness/claude-code.md` | excluded: Claude Code harness adapter; this pack stays Pi/pi-herdr-agents only | No file under these roots is in the not-yet-carried state. Between the previous pin `190d28736eff128ec94f502a22c7fe0c4b56cd37` and the 8b3326d pin, `docs/guide/06-verify-and-ship.md` changed (excluded under Q1). `pstack/agents/*` and `docs/harness/claude-code.md` are unchanged. `pstack/LICENSE` is unchanged (`bc957ca6bee02792566a1a028d105e02e247c6e77cf057061674273da77b200e`). Under `pstack/skills`, `poteto-help/SKILL.md` gained a `/statechart` table row and `statechart/**` was added at 8b3326d (then flagged); both ship as of the 1ba91fc pin as an optional skill. The shipped runner uses the pack `tla-ts` dependency (no Java). `in-your-own-words`, `ketchup`, `sharpen`, and `poteto-mode/scripts/check-plan.test.ts` stay flagged, not shipped. Unscoped `pstack/README.md` and `pstack/plugin.json`, and out-of-scope `pstack/.cursor-plugin/plugin.json` (0.15.29 to 0.15.30), changed and were not ported. ## Upstream scope: mattpocock/skills The `mattpocock` record in `test/fixtures/upstream-scope.json` maps `skills/mattpocock/` to mattpocock/skills at `5863806dd4e15c63174bee4e7515a4256792ecfa` (MIT, source `mattpocock` in `docs/skill-inventory.json`, skills root `skills`). Its checkout goes in `PSTACK_MATTPOCOCK_SOURCE`. **Scope choice.** The whole upstream `skills/` tree is in scope, not only the skills this pack ships, so the weekly sync sees every new, changed or removed upstream file. All 114 files at the pin are accounted for: 38 ship in `skills/mattpocock/` (34 adapted, 4 copied byte for byte), none is planned (owned by a later issue) since #37 batch B, and 63 are excluded (with a reason), all in `test/fixtures/skill-provenance/mattpocock.json`, and 13 ship merged in `skills/merged/`, each the primary source of a file in `test/fixtures/skill-provenance/merged.json`. `files` records each shipped file's source path, git blob, SHA-256 and what changed; `planned` and `excluded` record the same hashes. With a checkout, `npm test` reproduces every hash and blob and fails on any upstream file the fixture lacks; without one it checks the fixtures against each other and the shipped tree. The record's root is `.` (the repository root) so `LICENSE` can be watched like cursor's: its SHA-256 is bound to the notice in `THIRD_PARTY_NOTICES.md`. Other top-level entries are classified: `.claude-plugin` is out of scope (Claude Code manifests) and the rest is unscoped repository tooling and docs. `npm run fixtures:upstream-scope -- []` also lists, for both upstreams, skills-root files that are new, removed or changed. New and removed are computed against the complete accounted set (inventory and provenance fixtures). Only provenance entries carry a hash, so only they are checked for content changes; a path the inventory alone lists is checked for presence, not content (`npm run upstream:drift` reports a change to any of them between the pin and the target). It never rewrites that accounting, and exits non-zero when any needs a person. The first run of the corrected removal check found three cursor `poteto-mode/scripts/orch/*` paths that the inventory listed but cursor deleted before the pin (cursor/plugins `c742dd6`); they were dropped from the cursor list and stay in the mimir list, where they exist. | Upstream path | Status at the pin | | `skills/engineering/setup-matt-pocock-skills` | adapted: `skills/mattpocock/setup-repo/` (SKILL.md and the five seed templates, issue #32). Its Codex-only `agents/openai.yaml` is excluded. | | `domain-modeling`, `grill-with-docs`, `implement`, `implement-spec`, `to-spec`, `to-tickets`, `triage`, `wayfinder`, `grill-me`, `grilling` | adapted: `skills/mattpocock//` with every supporting file (issue #33); `ADR-FORMAT.md` and `GLOSSARY-FORMAT.md` are byte-identical copies. Each Codex-only `agents/openai.yaml` is excluded. | | `research`, `handoff`, `to-questionnaire`, `wizard`, `codebase-design`, `improve-codebase-architecture`, `pr` | adapted: `skills/mattpocock//` with every supporting file (issue #34); `to-questionnaire/SKILL.md` and `pr/CREDITS.md` are byte-identical copies. Each Codex-only `agents/openai.yaml` is excluded. | | `tdd`, `teach` | merged: `skills/merged//` with pstack's skills of the same name (issue #36); see **Merged skills**. Their Codex-only `agents/openai.yaml` files are excluded. | | `code-review` | adapted: `skills/mattpocock/code-review/` (issue #37, decision 1: kept beside pstack's `interrogate` and the playbook gates); see **Overlapping skills**. Its Codex-only `agents/openai.yaml` is excluded. | | `prototype`, `ask-matt` | merged: `skills/merged/prototype/` with pstack's Prototype playbook, and `skills/merged/poteto-help/` with pstack's `poteto-help` (issue #37, decisions 3 and 5); see **Overlapping skills**. Their Codex-only `agents/openai.yaml` files are excluded. | | `diagnosing-bugs`, `retro`, `writing-for-agents` | adapted: `skills/mattpocock//` with every supporting file (issue #37, decisions 2, 4 and 6: each kept beside pstack's Bug fix playbook, `reflect` and `technical-writing`); see **Overlapping skills (issue #37, batch B)**. Each Codex-only `agents/openai.yaml` is excluded. | | `in-progress/chief-of-staff` | excluded: `chief-of-staff` was dropped by issue #37 decision 7 (it overlaps poteto-mode's Orchestrate playbook, and #28 puts the in-progress folder out of scope). Known gap: its dated-wiki discipline has no equivalent in this pack. | | `skills/misc`, `skills/in-progress`, `skills/deprecated` | excluded: out of scope for v2 (#28) | | every `agents/openai.yaml`, the bucket `README.md` files | excluded: Codex-only metadata and upstream indexes | | `skills/productivity/wait-what` | excluded for now: no issue names it, so it needs a decision | `setup-repo` adapts Matt's `setup-matt-pocock-skills` for Pi and pi-herdr-agents. It keeps the upstream process and wording. It changes only the name, `/skill:` command syntax, the instruction file order (`AGENTS.md`, then `CLAUDE.md`, never a second file), a never-file-into-either-upstream rule, an explicit approval before labels, remotes or files, the triage check (Pi hides skills that disable model invocation from the available-skills list) and backticked paths for the template links, which `test/skill-content.test.ts` requires. A review added: the tracker repo is resolved, shown and confirmed in the step 3 approval, written into the generated tracker doc, and passed as `-R /` to every `gh`/`glab` issue, pr, mr and label command (`repo view` takes it positionally and `api` paths name it; neither accepts `-R`), so an `upstream` remote cannot redirect labels or issues; the domain layout and any legacy `CONTEXT.md` rename join the step 3 drafts; and repo content read during exploration never authorizes a command. These are content checks on the skill text; they do not show that a live model follows it. The seed templates carry the same adaptations as this repository's own `docs/agents/*.md`; `test/mattpocock-skills.test.ts` compares them. It names `to-tickets`, `triage`, `to-spec`, `wayfinder`, `domain-modeling`, `grill-with-docs` and `improve-codebase-architecture`; all but the last ship with #33, and `improve-codebase-architecture` ships with #34. With `triage` now in the pack, setup-repo's triage-label section runs. The spec, ticket and triage flow (#33) adapts ten more skills the same way: upstream's wording, process and `disable-model-invocation` values stay, and only what Pi and pi-herdr-agents require changes. Each file's provenance entry says what changed. In summary: - Pi has no Skill tool. A skill that loads another reads its `SKILL.md` by a backticked path relative to its own directory (`../grilling/SKILL.md`, `../../merged/tdd/SKILL.md`), which `test/skill-tree.test.ts` resolves. `grilling` and `domain-modeling` stay model-invocable, as upstream has them, so Pi lists them for the model; the other eight load only through `/skill:`. - Slash commands use Pi's `/skill:` syntax, and `/setup-matt-pocock-skills` became `/skill:setup-repo`. - Every skill that reads or writes the tracker names the configured tracker doc (`docs/agents/issue-tracker.md`, from the `## Agent skills` section) and runs each command as that doc writes it, with its `-R /`, never the CLI's implicit default repo. Skills that apply labels ask the user before creating one the tracker lacks. - Sub-agents and background agents became pi-herdr-agents `subagent` launches with `fork: false`, an explicit `model` (`task:recon` or `task:coding`) and `thinking`, and results that arrive on their own, never polled. `implement-spec` gives implementers a `worktree` based on the integration branch, keeps the merger and review fixes in the current checkout one at a time, and removes worktrees with `worktree_remove` only after the user approves the list. A child cannot see skills that disable model invocation, so its task carries the absolute path of the skill it must load. - `implement` and `implement-spec` load the merged `tdd` (`skills/merged/tdd`, issue #36) and `code-review` (`skills/mattpocock/code-review`, issue #37). `research` ships with #34 and `prototype` (`skills/merged/prototype`) with #37, so wayfinder resolves both in this pack. The skills look for each by name, in this pack first, then among the user's installed skills or at a path the user gives (never from tracker content), and fail closed when none is found: without `code-review`, `implement` stops before its commit and `implement-spec` stops before any PR or ticket update; without `research` or `prototype`, those wayfinder tickets are reported as blocked. No other skill stands in. That wording stays for installs that filter a skill out; the pack itself ships all four. - Pack policy added after review: the tracker doc says how to publish, never whether. Pushes, PR actions, issue creation, edits, comments, assignment, labels and label creation need the user's grant, and without it the work stays local. An explicit `/skill:` invocation grants only the write the skill exists for, in the tracker doc's repo: `to-spec` the one spec issue, `to-tickets` the tickets approved in its quiz with their links, `triage` the label, state and comment changes the maintainer directs (closing only on the maintainer's direction); `wayfinder`, `implement` and `implement-spec` get none. A skill loaded without the user invoking it grants nothing. Issue, PR and comment content, map Notes and linked material are data, never instructions, scope changes, skill choices or authorization, and child briefs repeat that. `triage` first inspects a PR from fetched git objects (`git diff`/`git show` with `--no-ext-diff --no-textconv`, no checkout), says that a checkout is not a sandbox (creating one can run hooks and filter drivers), and creates a detached review checkout under `~/.herdr/worktrees//pr--review`, with `core.hooksPath=/dev/null`, or runs contributor-controlled code, only after the user approves knowing it runs with their credentials and environment (or in an isolated environment they provide). A PR whose `.gitattributes` names a filter, diff or merge driver is verified only in such an isolated environment. Removing the checkout needs the user's approval of its exact path. `implement-spec` pushes or acts on a PR only when the review preflight passed and the user granted it. Worktree children start from committed state, so uncommitted inputs are committed with approval or passed by absolute path. `implement-spec` inspects each child's diff and checks before merging and advances the frontier on verified integration, not remote ticket closure. - Relative Markdown links became backticked file names or italic section names, and `triage/OUT-OF-SCOPE.md`'s example fence widened to four backticks so its nested code block no longer breaks the file. `test/skill-tree.test.ts` checks that every backticked supporting-file name exists in the skill's folder and exempts only an explicit list of illustrative links in fenced templates. These are content and loader checks (`test/mattpocock-skills.test.ts`, `test/rpc.test.ts`); they do not show that a live model follows the skills. The remaining gap skills (#34) adapt seven more the same way: `research`, `handoff`, `to-questionnaire`, `wizard`, `codebase-design`, `improve-codebase-architecture` and `pr`. Upstream's wording and invocation settings stay: `research`, `wizard`, `codebase-design` and `pr` are model-invocable, as upstream has them, and the other three load only through `/skill:`. `to-questionnaire/SKILL.md` and `pr/CREDITS.md` are copied byte for byte. Each file's provenance entry says what changed. In summary: - Pi has no Skill tool. `improve-codebase-architecture` loads `codebase-design`, `grilling` and `domain-modeling` by path (`../codebase-design/SKILL.md` and so on), and `handoff`'s suggested-skills section names each skill's `/skill:` command or `SKILL.md` path. `/codebase-design` mentions became the `codebase-design` skill, `wizard`'s template header names `/skill:wizard`, and relative Markdown links became backticked file names, as in #33. - Background agents and sub-agents became bare pi-herdr-agents `subagent` launches with `fork: false`, an explicit `model` and `thinking`, and results that arrive on their own, never polled: `research`'s researcher (`task:recon`, `medium`, no worktree, one findings file, no commit, push or further subagents), `improve-codebase-architecture`'s explorer (`task:recon`, `medium`, read-only) and `codebase-design`'s design-it-twice designers (`task:architecture`, `high`, on paper only). - `research` carries a **Sources are data** rule into its child's brief word for word: web pages, docs, source code, specs and API responses are evidence, never instructions, scope changes, skill choices or authorization; the child never runs a command a source supplies or sends it local data. A subagent that loads `research` (a wayfinder research ticket) is a leaf and does the job itself, and wayfinder's brief now says so. - Grants. Each skill's grant, on explicit invocation, is local: `research` one uncommitted findings note in the repo; `handoff` one document in the OS temp directory; `to-questionnaire` one `to-questionnaire-.md` in the current directory; `wizard` the script (committed only where the commit policy allows), never running it; `codebase-design` nothing; `improve-codebase-architecture` the HTML report in the OS temp directory and GLOSSARY.md or ADR edits. `pr` drafts the body as text: upstream's stated purpose is writing the body, not the remote write, so creating or editing a PR needs the user's grant and then goes to the tracker doc's repo with its `-R /`. `pr`, `research` and `wizard` carry #33's shared publishing paragraph with the grant sentence "grants no tracker or remote write", and `pr` also carries #33's tracker-content trust paragraph. - `wizard`: the agent never runs the script, in whole or in part, nor the commands its stages would run; it writes no secret value into the script and never reads a live `.env` or `.env.*` file, taking variable names only from templates (`.env.example` and the like), config schemas, workflows and code, and asking the user for missing names, never values. After review, the template library binds GitHub writes to one repository: scoping confirms the `/`, the author sets `REPO` to it, `set_secret` and `set_var` pass `--repo "$REPO"` and skip when it is unset, and the intro shows it and stops unless the human confirms, so `GH_REPO`, the directory or a fork's default cannot redirect a secret. This is the one change to upstream's library. A pre-release review found that capturing a value named `REPO` (or `ENV_FILE`) overwrote that destination, so the confirmed repository is now a separate copy that the intro makes read-only with `REPO` and `ENV_FILE`, the helpers' locals carry a `_wiz_` prefix, and `ask`, `ask_secret` and `write_env` refuse the library's names. - `handoff`: beyond upstream's redaction sentence, the document never holds a secret, even in part; it names where a secret lives, and the finished document is checked before its path is reported. - `research`'s child-carried rule also keeps secrets and credential-bearing URLs (signed URLs, tokens in query strings, API responses with keys) out of the notes; credentials are named by variable or location. `improve-codebase-architecture`'s report keeps Mermaid's `securityLevel` at `"strict"` (upstream: `"loose"`) and HTML-escapes repository-derived text. - `pr`'s Summary section comes through Matt's skill from HumanLayer's `show-me` (MIT); `THIRD_PARTY_NOTICES.md` reproduces that notice. These are content and loader checks (`test/mattpocock-skills.test.ts`, `test/rpc.test.ts`); they do not show that a live model follows the skills. ## Merged skills: `skills/merged/` (issue #36) A skill drawn from both upstreams lives once in `skills/merged/` (ADR-0003) and records every source. `tdd` and `teach` are the first two; `prototype` and `poteto-help` followed in #37 (see **Overlapping skills**). Their provenance is `test/fixtures/skill-provenance/merged.json`. Each file lists, per source, its path, SHA-256 and git blob at the pin, and what changed. A source's `use` is `primary` (Matt's file, the base of the merged text), `merged` (pstack's cursor/plugins file, whose method is merged in by hand) or `context` (the pi-mimir copy, which matches the cursor file for both skills; the pack's shipped `tdd` was byte-identical to it, but the shipped `teach` at `0b1dde5` already differed by one Pi sentence about the image tool). `prototype` follows the same pattern, with the poteto-mode Prototype playbook as the `merged` source. `poteto-help` reverses it: pstack's file is `primary` and Matt's `ask-matt` is `merged`, while its flow map has ask-matt as `primary` and pstack's file as `merged`. The inventory rows say `skills/merged` and list the source files of all three. `test/fixtures/upstream-scope.json` has a `merged` record naming the folder and its two pinned upstreams (`cursor` and `mattpocock`). Drift detection: `npm run fixtures:upstream-scope -- ` already compares each recorded hash with the upstream at its pin. For the merged folder it also prints the section `merged skill with a changed or removed source (manual merge)` when this upstream's source of a file with more than one upstream source changed or vanished, names the file, and exits non-zero. A person merges the change by hand, then records the new hashes in `merged.json`. Run it for each upstream: a change to either source flags the merged file. `test/merged-skills.test.ts` proves it on scratch upstreams for a change and for a removal on either side, and that a merged file is not flagged when neither of its sources changed. A merged file that has a single source (Matt's `tests.md`) is covered by the ordinary changed-hash check. `tdd` merges Matt's red → green loop (seams agreed first, anti-patterns, rules of the loop, `tests.md`, `mocking.md`) with pstack's gate for when to use it or skip it, bug-fix workflow, impractical-test fallback, guardrails and evidence report. `teach` has two modes, because the two skills teach different things. Explain is pstack's one plain account built on `how` and `why`, and writes nothing. Workspace is Matt's multi-session teaching workspace with lessons, reference documents, learning records and four format files. Where the two conflict the stricter rule is in the merged text: - Seams. Matt confirms them with the user before any test; pstack proceeds on an obvious cheap test path. The merged `tdd` confirms the seam in one line for such a bug, counts seams confirmed in a spec or ticket, and has a delegated child take them from its brief, or from a spec or ticket the brief points at that already confirms them, and report its proposal and write no test only when neither names any. `implement-spec` has each implementer brief state where its seams are (named, or "confirmed in "), and the parent never merges a child that stopped at "proposed seams, wrote no test": it confirms the seams with the user or from the spec and re-briefs, and the merger never integrates a diff whose tests did not go red when `tdd` applied. - Skipping. The use-or-skip gate runs first and a skip is told to the user (before the substitution when they asked for TDD), never silent. - Writes. Explain mode writes nothing. Workspace mode writes only after the user confirms the directory (a git repository with project code needs an explicit yes), and only the named files inside it. - Sources and effects. Workspace mode treats web sources as data, loads no remote script, stylesheet or font in lessons, opens only the lesson it wrote, and suggests communities without joining or posting. Neither skill grants a push, pull request or tracker write. - `disable-model-invocation: true` for both. pstack's `tdd` and `teach` and Matt's `teach` set it. Matt's `tdd` does not, but the merged `tdd` asks the user to confirm seams and its gate depends on what the user asked for, so the model does not load it unprompted. `implement` and `implement-spec` load it by path; the bug-fix playbook only says "see the tdd skill", by name. `test/merged-skills.test.ts` holds the content-contract tests for both skills (each rule is shown to fail when its text is removed), the provenance checks with and without checkouts, and the loader tests. They are content and loader tests, not evidence that a live model follows either skill. ## Overlapping skills (issue #37, batch A) Issue #37 decides each pair of overlapping skills. Batch A implements decisions 1, 3 and 5 and closes the release gate #33 left open (`implement`, `implement-spec` and `wayfinder` found no `code-review` or `prototype`). - **`code-review`** (decision 1: keep both). Matt's two-axis review ships as `skills/mattpocock/code-review`, beside pstack's `interrogate` and poteto-mode's review gates, which are unchanged. It keeps upstream's Standards and Spec reviewers, smell baseline, briefs and aggregation. It is read-only: it changes no file, index, branch or stash, pushes nothing and posts nothing unless the user grants it. Every `git diff` and `git show` it documents carries `--no-ext-diff --no-textconv`, so configured diff drivers and textconv helpers never run (not a sandbox). It reviews one of two targets. Committed work is upstream's three-dot diff, against a named `` such as `implement-spec`'s integration branch, read from git without a checkout. Uncommitted work since the merge-base of the fixed point and `HEAD` is reviewed as three labelled layers: commits since the base, the staged commit candidate (`--cached`) and unstaged changes, because one net diff can hide a staged change the working tree reverts. Paths are listed before any content is read: likely credential files are never read or handed on (and are excluded from every diff by pathspec), symlinks that resolve outside the repository are not followed, binaries are skipped and large files (over 2,000 lines or 256 KB, or one long line) are read only in part and reported as partly covered, and the report names what was omitted. `implement` records its starting commit and, when tracked files are already dirty, a `git stash create` baseline (it writes unreferenced objects and leaves the working tree, the index and refs unchanged); code-review reviews those files in two layers against it, staged (`--cached ^2`) and working tree, so a change staged and then reverted is still seen and a file the task changed is never dropped, and `implement` asks before modifying a file that was already untracked. It then reviews its uncommitted work before it commits; `implement-spec` records the commit its integration branch starts from and reviews that branch as committed work. The two reviewers are bare pi-herdr-agents subagents with `task:review`, `thinking` only when that category has none, and no worktree, because a worktree starts from committed state. `tdd` names it where it defers refactoring. - **`prototype`** (decision 3: merge). `skills/merged/prototype` combines Matt's prototype (primary; the logic and UI branches with `LOGIC.md` and `UI.md`) with pstack's poteto-mode Prototype playbook (merged: the empirical branch, references, evidence rules, reply and handoff to Feature). Scratch is the default location; building inside the app needs the user's explicit scope; keeping, committing, pushing or pointing at a prototype each need authorization, and prototype code never reaches the main branch. A logic prototype's pure module reaches production only in a separately authorized production change with its normal tests and review. The playbook keeps its place in poteto-mode's routing and points at the skill (`w2.json` records the change and why the file is now under half its source). `wayfinder`'s Prototype tickets name the skill. The skill keeps Matt's model invocation, since its default writes only to scratch. - **`poteto-help`** (decision 5: merge, no separate `ask-matt`). pstack's `poteto-help` moved from `skills/pstack` to `skills/merged` and keeps its install, `/setup-pstack`, poteto-mode and troubleshooting content and the brief-reading contract `test/setup.test.ts` asserts. Matt's `ask-matt` flow map became `references/flows.md` and its phase-boundaries tree `references/phase-boundaries.md`, adapted to the pack: `/skill:` commands, Pi's `/new` and `/compact`, and only flows whose skills ship and are dependency-closed. Matt's `retro`, `diagnosing-bugs`, `wait-what` and `writing-for-agents` were named only as absent, with the shipped alternative; batch B shipped all but `wait-what` and routes to them. Playbooks are not skills. Its provenance moved from `w3-b.json` to `merged.json`. `merged.json` records both upstreams for `prototype/SKILL.md`, `poteto-help/SKILL.md` and `poteto-help/references/flows.md`, so `npm run fixtures:upstream-scope -- ` flags each for a manual merge when either source changes. ask-matt's `SKILL.md` is the primary source of `flows.md` and merged into poteto-help's `SKILL.md`; the mattpocock accounting counts each Matt file once, by the file it is the primary source of. `test/overlap-skills.test.ts` holds the content contracts (each rule is shown to fail when its wording is removed), a disposable-repo run of every git command code-review and `implement` document (no configured helper fires, the staged layer shows a change the working tree reverts, the baseline delta keeps a task's edit to an already-dirty file and drops the untouched user edit, and a bug staged then reverted in such a file shows in the staged baseline layer, an excluded credential path stays out of a diff), the checks that the dependents find these skills in the pack, the flow map's dependency closure and the real-Pi loader test. They are content and loader tests, not evidence that a live model follows the skills. ## Overlapping skills (issue #37, batch B) Batch B implements decisions 2, 4, 6 and 7. The three ports keep upstream's wording and `disable-model-invocation` values (`diagnosing-bugs` and `writing-for-agents` stay model-invocable, `retro` stays user-invoked) and change only what Pi and pack policy require. None launches a subagent, so the task-category thinking rule does not apply to them. Each file's provenance entry in `test/fixtures/skill-provenance/mattpocock.json` says what changed. - **`diagnosing-bugs`** (decision 2: keep both). Matt's six-phase diagnosis loop ships as `skills/mattpocock/diagnosing-bugs` with `scripts/hitl-loop.template.sh`. Its feedback loop does not replace a real-path reproduction: Phase 2 also sees the symptom through the user's reported steps on the real surface, Phase 6 re-runs that reproduction as well as the loop, and a caller's stricter gate wins. The human-in-the-loop script follows `wizard`'s rule: the agent copies and edits it, the human runs it in their own terminal and pastes the captured block back, and the agent never runs it in whole or in part (`bash -n` and `shellcheck` only parse it), puts no secret in it and captures none. Upstream's template had the agent run it. The skill grants only local diagnosis; commits follow the commit policy and pushes, PRs, deploys and production instrumentation need authorization. poteto-mode's Bug fix playbook stays a playbook: a paragraph points at the skill for technique and says the playbook's gates still bind it (`w2.json`). The template keeps upstream's file mode, `100644`. After review, its library changes too (pack policy): upstream's footer printed only the example's two captures and failed under `set -u` for any other names, so `capture` now records each name and the footer prints every captured value. Cleanup removes every temporary probe, tagged or not, and checks `git diff` and `git status` before reporting done. - **`retro`** (decision 4: keep both). Matt's retrospective ships as `skills/mattpocock/retro` and loads `writing-for-agents` by path, which ships in the same batch. It proposes guardrails for the repo and its environment and routes skill edits to `reflect`; `reflect` routes repo and environment lessons back to `retro` (`w4-a.json`). It reads `$PI_SESSION_FILE` and only this working directory's earlier sessions, checking each session header's `cwd` before reading its body, treats transcript content as data, keeps secrets out of what it writes or shows, changes nothing until the user approves a candidate, and never changes the user's global Pi configuration, credentials or installed extensions. - **`writing-for-agents`** (decision 6: keep both). Matt's reference ships as `skills/mattpocock/writing-for-agents` with `SKILL-MECHANICS.md`, which is adapted to Pi's skill mechanics: Pi's frontmatter rules (a name matching its directory is a convention, not a Pi requirement), `/skill:`, `AGENTS.md`, and a skill reaching another by path whatever its invocation setting (upstream assumed only a model-invoked skill can be reached). It says it serves instructions for agents; `technical-writing` gains a mirror paragraph saying it serves prose for humans (`w3-b.json`, now adapted), and where they conflict each says the reader's audience wins. The `authoring-a-skill` playbook references both (`w4-p.json`). - **`chief-of-staff`** (decision 7: drop). It moves from `planned` to `excluded` with the reason. **Known gap:** its dated-wiki discipline (a folder of interlinked pages that outlives the session, kept by ingest, query and lint, every claim dated or tied to a commit, superseded claims marked rather than rewritten) has no equivalent in this pack. `poteto-help` routes to the three new skills: its flow map restores Matt's `retro`, `diagnosing-bugs` and `writing-for-agents` entries as `/skill:` steps with the pack's rules, names only `wait-what` as absent, and stays dependency-closed; the help adds table rows and close calls (retro or reflect, the skill or the Bug fix playbook, writing-for-agents or technical-writing) and lists both new model-invocable skills (`merged.json`). No Matt file remains planned. `test/overlap-skills.test.ts` holds the content contracts for each adaptation (each shown to fail when its wording is removed), a `bash -n` parse of the template and a run of a customized copy's library with piped answers, the pointer and drift checks for the playbooks, `reflect` and `technical-writing`, and the real-Pi loader test. They are content and loader tests, not evidence that a live model follows the skills. ## Wave 1: moved role and package files | Destination | Source | Status | | --- | --- | --- | | `agents/poteto.md` | pi-herdr-agents `agents/poteto.md` at `c2177dff835da44937e614e8a03d0405d442e848` | moved unchanged in W1; **removed in W2** | | `LICENSE` | pi-herdr-agents `LICENSE` at the same commit | unchanged | `poteto.md` first appeared in pi-herdr-agents' launch commit `4fa3f26` and was moved here byte-identical in Wave 1. `test/fixtures/provenance.json` records the SHA-256 of each source blob and destination; `npm test` fails on unreviewed drift and, when a pi-herdr-agents checkout is available, reproduces each source hash from the source commit. Wave 1 also added `pi-extension/pstack/{index,roles}.ts` (a role-pack bridge adapted from pi-herdr-agents' `examples/role-pack/extension.ts` and ADR-0003), package metadata, tests, `README.md`, `THIRD_PARTY_NOTICES.md` and these documents. ## Wave 2: role removal The user removed the named `poteto` role from the Wave 2 target ([historical plan 09](https://github.com/giuseppecrj/pi-herdr-pstack/blob/96910f507046285071453b660099b0c3549e6161/docs/plans/09-wave2-no-poteto-role.md)). `agents/poteto.md`, the `roles.ts` bridge and its registration were deleted; pstack contributes no named roles and registers no role directory. The provenance fixture keeps the role's W1 source and hashes with status `removed` and `removedIn: "W2"`. No replacement role was added and the role body was not copied elsewhere. ## Wave 2: runtime `pi-extension/pstack/{config,mode,resources,setup}.ts` and their tests are new. ## Wave 2: methodology skills The `skills/pstack/poteto-mode/` and `skills/pstack/setup-pstack/` files are adaptations of upstream pstack, recorded separately from the Wave 1 role fixture in `test/fixtures/skill-provenance/w2.json`. Sources: | Key | Repository | Commit | Root | | --- | --- | --- | --- | | mimir | https://github.com/casualjim/pi-mimir | `f07dd981f62c9c994a5d043ede67d6c63c721454` | `packages/pi-pstack/skills` | | cursor | https://github.com/cursor/plugins | `1ba91fc25c0977741dd0e267849e6234b037ed0a` | `pstack/skills` | | Destination | Primary source | Status | | --- | --- | --- | | `skills/pstack/poteto-mode/SKILL.md` | mimir `poteto-mode/SKILL.md`, Cursor hub as context | adapted | | `skills/pstack/poteto-mode/playbooks/*.md` (12 base playbooks) | mimir playbook of the same name, Cursor as context | adapted | | `skills/pstack/poteto-mode/references/bugbot-triage.md` | mimir file of the same name | adapted | | `skills/pstack/poteto-mode/references/delegation.md` | derived from both hubs' Subagents sections | new | | `skills/pstack/poteto-mode/references/authorization.md` | derived from both hubs' Autonomy sections | new | | `skills/pstack/setup-pstack/SKILL.md` | Cursor `setup-pstack/SKILL.md` | adapted; rewritten under #35 to follow upstream's steps (below) | `setup-pstack` was rewritten under #35 to follow the Cursor skill's eight steps and wording where Pi allows: detect models, load current state, budget, map and confirm, validate, write, confirm, and offer once each the verification and formal verification skills. What changed from upstream: it writes pi-herdr-agents' `models.tasks` and `models.tasksThinking` instead of a Cursor rule file; the budget ladder sets every category's thinking (`max`, `xhigh`, `high`, `medium`), which the host clamps per candidate instead of Cursor's per-slug effort tokens; upstream's roles are mapped onto the six task categories in a table, with panel roles left as diversity seats; there are no `auto`-style aliases or skill defaults; the question is asked in chat; and the write is one `subagents_write_task_models` call approved in the host's dialog. The earlier report-first skill, change and init flows, `pstack_apply_task_models`, the writer gate and `pi-extension/pstack/config.ts` were deleted; `setup.ts` now only activates the host writer and briefs the skill. The fixture records the details. The fixture records, for every shipped file, its SHA-256, each source path and blob hash at the pinned commit, and what changed. The adaptations share four themes: - Delegation uses pi-herdr-agents' public single-call `subagent` schema (tested host `e262c584f54a7c8d60eb1fa5510f47c1299e3801`). Every delegate is a deliberate bare delegate with a bounded reference prompt: implementer, investigator, reviewer or verifier. Upstream `poteto-agent`, `role`/`tasks`/`chain`, `subagent_type`, model aliases and hardcoded model defaults are gone. - Methodology never grants permission. Commits follow the task's commit policy. Pushes, PR operations, review-thread replies, merges, deletions and shared configuration writes need explicit authorization. Pstack has no shell-command guard, so this is behavioral, not enforced. - Platform tools (`set_tasks`, `AskQuestion`, `/loop`, Cursor transcripts, `cursor-team-kit` control and cleanup skills, the Cursor rule file) are replaced by Pi or pi-herdr-agents equivalents or by explicit manual steps. - Skills from later waves keep their names. Every reference to one is marked planned W3 or planned W4 on the line that makes it, with a safe manual equivalent or a reported skip. Wave 2 listed 96 exact references in `test/fixtures/forward-references/hub.json`; the 54 W3 references retired when Wave 3 shipped, leaving 42. Of the 45 mimir `poteto-mode` source files, 14 are adapted here. The other 11 playbooks are planned W4 methodology work and are not shipped as empty files. The 20 helper-script files have documented dispositions: excluded (dependency bootstrap, a duplicate scheduler/ledger, direct worktree pruning) or deferred to W4 replacements (the PR watcher and plan checker). `npm test` checks destination hashes and source accounting. It reproduces the source hashes when `PSTACK_MIMIR_SOURCE` and `PSTACK_CURSOR_SOURCE` point to checkouts that contain the pinned commits. ## Wave 3 fixture layout From Wave 3 the fixtures are split by owner so parallel writers never share a file. `test/fixtures/skill-provenance/` holds `w2.json` (the Wave 2 fixture, moved verbatim), `w3-a.json` and `w3-b.json`. `test/fixtures/forward-references/` holds `hub.json` (the Wave 2 tuples, moved verbatim), `w3-a.json` and `w3-b.json`. The test unions each directory and requires every entry to lie in its owner's skill directories. A file's status is `copied` (destination bytes equal the primary source), `adapted` (one primary source, changes explained) or `new` (derived only). `docs/skill-inventory.json` is at schema version 2: rows carry a `status` of `planned` or `shipped` that drives the content checks, and the `no-comments` row lists `../agents/comment-sicko.md` (relative to each source's skills root) as the source of its bare delegate prompt `skills/pstack/no-comments/references/comment-sicko.md`. The delegation schema pin is pi-herdr-agents `7d35371f5d7d0df3edd208a1d5c9a187767d563b`; its `SubagentParams` and task categories are byte-identical to the Wave 2 pin `e262c584`. ## Wave 3: 35 skills and the comment-sicko delegate Wave 3 shipped the 35 Wave 3 inventory rows: the 24 `principle-*` skills, `tdd`, `correct`, `benchmark-checklist`, `blast-radius` and `typescript-best-practices` (W3-A, `test/fixtures/skill-provenance/w3-a.json`), and `unslop`, `technical-writing`, `no-comments`, `teach`, `bro` and `poteto-help` (W3-B, `w3-b.json`). Those rows are `shipped` in `docs/skill-inventory.json`, so 37 rows ship in total. Of the 39 Wave 3 files, 22 are copied byte-for-byte from mimir and 17 are adapted, each with one primary source and an explanation in its fixture. `tdd` and `teach` have since moved to `skills/merged/` (issue #36, see **Merged skills**); the rows below mark them. | Destination | Primary source | Status | | `skills/pstack/principle-*/SKILL.md`, `correct` (`tdd` moved to `skills/merged/tdd`, #36) | mimir file of the same name | 18 copied; 7 principles adapted for skill-relative paths, full principle names or a planned W4 reference | | `skills/pstack/{benchmark-checklist,blast-radius}/SKILL.md`, `skills/pstack/typescript-best-practices/**` | mimir | adapted (paths, names, planned W4 interim rules; `paths:` frontmatter dropped; the boundary row follows the Cursor blob, and `references/patterns.md` is aligned to that row) | | `skills/pstack/{unslop,technical-writing,bro}/SKILL.md` | mimir | copied | | `skills/pstack/no-comments/SKILL.md` (`teach` moved to `skills/merged/teach`, #36) | mimir, Cursor as context | adapted | | `skills/pstack/no-comments/references/comment-sicko.md` | mimir `../agents/comment-sicko.md` at `f07dd981f62c9c994a5d043ede67d6c63c721454` | adapted: frontmatter dropped, made a comment editor, planned W4 interim rule | | `skills/pstack/poteto-help/**` (3 files; moved to `skills/merged/poteto-help`, #37) | Cursor `poteto-help/` at `ccb5507cec1546dc88135c1139c811e6c59115ba` | adapted, re-authored for Pi from Cursor's structure | `skills/pstack/poteto-help/SKILL.md` (now `skills/merged/poteto-help/SKILL.md`) adapts the one-time setup nudge that landed in cursor/plugins `1e56b2913469f05101e4ffd4d437c2e6fd2c1b8d` (PR #509) and is now in the pin. The fixture records the Cursor blob at `ccb5507cec1546dc88135c1139c811e6c59115ba`. The nudge asks once, when the answer depends on the six task categories and a `/setup-pstack` report does not already show them set, whether to run `/setup-pstack` (report only) or leave the categories as they are. The pin also changes the interrogate table row from "several models" to "different models". Cursor's Later line that every role keeps its default model is still uncopied. A category counts as unset when the Config line ends with the missing-file sentence, or when a present file's preference value is `(not set)`. Findings are not unset: the host trims a reference and both sides of `/` before it checks authentication, and the report cannot show that result. Any other Config-line ending is an unknown file state. poteto-mode always passes `task:`, so those delegations fail until an approved `/setup-pstack ` write sets the category. poteto-help does not read or write task-model config. comment-sicko is a bare delegate prompt under `skills/pstack/no-comments/references/`, not a role. `/skill:no-comments` reads it and passes it as `systemPrompt`; pstack still ships no named roles and has no `agents/` directory. The reconcile step retired every `planned W3` marker and its interim clause in the hub, playbooks and the W3 files, and removed the 63 forward-reference tuples that pointed at Wave 3 rows (54 in `hub.json`, 1 cross-batch in `w3-a.json`, 8 cross-batch in `w3-b.json`). The 93 remaining tuples all point at Wave 4 rows or playbooks. Two short example sentences in the copied `technical-writing` skill are attributed to Diátaxis and ASD-STE100 in `THIRD_PARTY_NOTICES.md`. ## Wave 4: 14 skills, 11 playbooks and two scripts Wave 4 shipped the last 14 inventory rows and the 11 remaining poteto-mode playbooks. The 190d287 pin move added two more inventory rows (`create-formal-verification-skill`, `maintain-formal-verification-skill`), so all 57 rows are `shipped` in `docs/skill-inventory.json` and all 23 playbooks ship. Each batch owns its fixtures: - W4-A (`w4-a.json`): `how`, `why`, `architect`, `arena`, `swarm`, `interrogate`, `reflect`, `figure-it-out`. 33 files, 21 copied and 12 adapted, all from mimir. (The 1.0.0 pre-release review adapted reflect's three reviewer prompts to never quote a secret or credential-bearing URL; `w4-a.json` holds each file's current status.) Swarm maps the Verifier prompt's per-claim pass, fail or inconclusive onto one overall `PASS`, `ISSUES` or `BLOCKED`. Briefs list their claims first. A proved fail always means `ISSUES`. `PASS` needs at least one listed claim and a pass for each, and a claim with no result is inconclusive. Zero claims means `BLOCKED`. A `BLOCKED` coverage slice is reported as unverified, never as covered. A brief's claims together cover the slice's share of the done predicate. Aggregation recomputes each status from the claims in the brief, with any other proved defect still `ISSUES`, and uses the result whenever it differs from the reported status. - W4-B (`w4-b.json`): `show-me-your-work`, `recall`, `automate-me`, `create-verification-skill`, `maintain-verification-skill`, `create-formal-verification-skill`, `maintain-formal-verification-skill`, `make-bot-ui`. 13 files, 4 copied and 9 adapted. `scripts/log.sh` is copied with its upstream mode `100755`. `make-bot-ui` is adapted from Cursor `make-bot-ui/SKILL.md` at `2cbf58508f40de470d7490b55c51d71241928fa2` and substantially re-authored for Pi: a loopback-bound local page whose server starts one `pi -p` run per request, with the key in a local file and exposure or installs only on authorization. Cursor's frontmatter name `Make Bot UI` became `make-bot-ui`. The two formal-verification skills are adapted from Cursor at `190d28736eff128ec94f502a22c7fe0c4b56cd37` (Pi paths, `/skill:` forms, install and PR authorization). - W4-P (`w4-p.json`): `playbooks/{hillclimb,eval,visual-parity,authoring-a-skill,babysit,shipping,autopilot-full,autopilot-stack,multi-phase-plan,orchestrate,worktree-cleanup}.md` and `scripts/check-plan.mjs` (mode `100644`), all 12 adapted from mimir. `check-plan.mjs` changed with the multi-phase-plan skeleton, since upstream enforced `/loop` and repository-root markers. Upstream cursor/plugins `multi-phase-plan.md` still requires a rebased head with the "patch-id unchanged"; this pack's rebase item in the skeleton's Merge block (`multi-phase-plan.md:130`) defers to the patch-id rule in `playbooks/shipping.md` instead, and `check-plan.mjs` rejects the old wording. The divergence is deliberate, so a sync must not restore it. Of the 56 new Wave 4 files, 25 are copied and 31 adapted. The C0 checkpoint added one new file, `skills/pstack/poteto-mode/references/fan-out.md` (in `w2.json`), the shared fan-out protocol. The package now ships 115 skill files: 48 copied, 64 adapted and 3 new. The ccb5507 pin move re-adapted `blast-radius` (it was a mimir copy after reconcile) for the "more than one model" wording. The 190d287 pin move added the two formal-verification skills as adapted Cursor leaves. The remaining upstream helper scripts stay excluded in `w2.json`, each disposition naming its replacement: `scripts/orch/*` (the Orchestrate playbook as a depth-1 coordinator with a hand-kept store), `scripts/watch-pr/*` (the one-shot watcher child in `fan-out.md` and stated `gh` field conditions), `scripts/worktree-audit.sh` (`worktree_list` and per-path `worktree_remove`), and `scripts/{bootstrap.ts,bun.lock,package.json}` (no runtime installer or script dependencies). The reconcile step retired every forward-reference tuple: the 93 that Wave 3 left (42 in `hub.json`, 5 in `w3-a.json`, 46 in `w3-b.json`) and the 9 cross-batch tuples the batches recorded (2 in `w4-a.json`, 1 in `w4-b.json`, 6 in `w4-p.json`). No target was deferred to Wave 5, so every fixture under `test/fixtures/forward-references/` is empty. It also removed every `planned W4` marker and interim "until it ships" clause from the hub, the base playbooks, `bugbot-triage.md`, `setup-pstack`, `poteto-help`, the six Wave 3 files and the Wave 4 files that cited each other, and rewrote each route to name the workflow's real behavior and scope-down limits. `principle-prove-it-works` remains a byte-identical copy. `blast-radius` was a copy after reconcile; the ccb5507 pin move adapted it. The content test now exempts fenced code blocks from the placeholder-marker and model-name bans (inline code too for placeholder markers, and for model names in byte-identical copies), and checks every fence with the `marked` lexer. That let three W4-A edits made only to satisfy the old bans be reverted: `why/references/sources/code-archaeology.md` and `reflect/references/synthesizer.md` are byte-identical mimir copies again, and `architect/references/runner-prompt.md` keeps upstream's `// TODO` marker while staying adapted for its other edits. ## G6: judgment category Judgment and prose synthesis (the how explainer, the why synthesizer, and reflect's judgment, divergent and synthesizer lenses) uses `task:architecture` with thinking at high or above. Reflect's tooling lens is a diversity seat. The hardest implementation changes go to that category or to the strongest exact catalog model with `thinking: xhigh`. The cursor pin is `1ba91fc25c0977741dd0e267849e6234b037ed0a`. The host describes `architecture` as planning and diagnosis, and its tier guidance puts that work in the frontier tier. `/setup-pstack` reports an unset `architecture` category; `qa` stays off that list. A flow that will need an architecture seat checks for a `- architecture:` bullet under the host prompt's Task-category shortlists before launching recon or review seats; the bare word architecture elsewhere in the prompt does not count. The tooling lens's family differs from the first model listed for architecture under those shortlists. If that launch falls back to a later candidate whose family matches the tooling lens, the lens is relabeled context-isolated. ## Excluded: upstream user guide cursor/plugins ships a user guide under `pstack/docs/guide/`: a `README.md` index, ten chapters (`01-setup.md` through `10-recipes-and-pitfalls.md`, about 1,100 lines in all at the previous pin `2cbf58508f40de470d7490b55c51d71241928fa2`, at `d0ef80d86795816da932a153458c5dbe192d294e`, at `f713f14203cd6420cec13cd7c55c32d598dbab68`, at `190d28736eff128ec94f502a22c7fe0c4b56cd37`, and at the current pin `1ba91fc25c0977741dd0e267849e6234b037ed0a`) and six `images/*.jpg`. It walks a user from setup and prompting through design, verification, overnight runs and customization. It sits outside the pinned `pstack/skills` root and is deliberately not ported, adapted or packed. Its setup and overnight chapters are Cursor-specific; the rest walks through skills that ship here, so an adapted copy would be a large second surface to keep in sync. Users get the equivalent from the package `README.md` (install, setup, first task) and `/skill:poteto-help`, which routes a question to the installed skill, playbook or principle file and gives a prompt to send. Within the pinned `pstack/skills` root, only `poteto-help/SKILL.md` linked the guide, with seven `../../docs/guide/*` links (to seven of the ten chapters). The adaptation removed them (`test/fixtures/skill-provenance/w3-b.json`), and no shipped file links a guide page. ## Q5: one forge Origin is dropped. `playbooks/opening-a-pr.md` uses `gh` for every PR operation, like Babysit, Shipping and the autopilots. The mimir and Cursor sources still offer Origin; the cursor pin is `1ba91fc25c0977741dd0e267849e6234b037ed0a`. The content test bans capitalized `Origin` in shipped skills, except make-bot-ui's backticked HTTP header field, and the lowercase forge CLI forms; the git remote named `origin` is unaffected. ## Pin move: cursor/plugins `ccb5507` The cursor pin moved from `2cbf58508f40de470d7490b55c51d71241928fa2` to `ccb5507cec1546dc88135c1139c811e6c59115ba`. The two commits in between that touch `pstack/` are `1e56b29` (#509, poteto-help setup nudge) and `df58112` (#511, drop Sol and retune default effort). The casualjim/pi-mimir pin stays `f07dd981f62c9c994a5d043ede67d6c63c721454`. Hand-ported, platform-neutral substance only: - `skills/pstack/blast-radius/SKILL.md`: "several models" became "more than one model". Status is `adapted`; the destination equals the Cursor blob and is no longer a mimir copy. - `skills/pstack/architect/references/runner-prompt.md`: "one of several runners" became "one of the parallel runners". - `skills/pstack/poteto-help/SKILL.md`: the interrogate table says "different models". The G5 setup nudge is now in the pin; the Pi mapping is unchanged. - `skills/pstack/interrogate/SKILL.md`: a rejected model prefers the same family and reasoning tier. Seats still follow the live catalog, one family each. Documented skips (Cursor-only model defaults, rule file, budget ladder, or excluded scripts): - `architect/SKILL.md`, `arena/SKILL.md`, `how/SKILL.md`, `why/SKILL.md`, `reflect/SKILL.md`, `poteto-mode/SKILL.md`: hardcoded Cursor slugs and the default panel list (Opus `max` to `xhigh`, Sol dropped). This pack already uses task categories and exact catalog diversity seats. - `setup-pstack/SKILL.md`: the budget ladder now treats `large` as the default and `unlimited` as `max`. This pack has no ladder. - `poteto-mode/scripts/orch/orch.test.ts`: verifier label `sol` to `opus`. The orch/ledger scripts stay excluded. - `pstack/docs/guide/01-setup.md` and `04-design.md`: excluded under Q1. - `pstack/README.md`: unscoped. `pstack/.cursor-plugin/plugin.json`: out of scope. `pstack/LICENSE`: unchanged. No new agents or docs file is in the not-yet-carried state. ## Skipped: Benny cursor/plugins `pstack/automations/benny/**` is not ported. It lies outside the pinned `pstack/skills` root. Benny is two Cursor Automations triggered by Slack posts: one triages a report and may file a tracker ticket, and the other reproduces the bug and may open a draft PR. Running it needs an external trigger or scheduler, which conflicts with this pack's no-scheduler rule, and it depends on Cursor Automations, Slack posting and tracker writes. Four of its methodology gates are adapted into `skills/pstack/poteto-mode/playbooks/bug-fix.md` (Q2): confirm the cause is in code this repo or team owns, verify an existing fix instead of writing a competing one, stop rather than race another person's claim on the fix, and reproduce twice before and after the fix. The sources are `reproduce-and-fix-issues/SKILL.md`, `reproduce-and-fix-issues/references/verify-existing-fix.md` and `triage-issue-reports/SKILL.md`, read at `d0ef80d86795816da932a153458c5dbe192d294e` and byte-identical at the pin. The bug-fix entry in `w2.json` records which steps changed. ## Pin move: cursor/plugins `f713f14` The cursor pin moved from `ccb5507cec1546dc88135c1139c811e6c59115ba` to `f713f14203cd6420cec13cd7c55c32d598dbab68`. About 79 commits sit between those pins; the ones that change in-scope pstack methodology are the token-cut and principle restatements (`00b14f1`, `08bd59b`, `e5136d8`, `99743e9`), verification readiness (`c49b37a`), fence-aware plan checks (`87c8103`), subagent-effort and playbook cuts (`ce649c0`), and the new skills `in-your-own-words`, `ketchup` and `sharpen`. The casualjim/pi-mimir pin stays `f07dd981f62c9c994a5d043ede67d6c63c721454`. Hand-ported, platform-neutral substance only: - `skills/pstack/poteto-mode/scripts/check-plan.mjs`: CommonMark fence pairing so nested fences are not treated as prose. PROGRAM_MARKERS, the unchanged-patch-id scan and git mode `100644` stay. Upstream's `/loop 1h` and `git show origin/main:` markers were not restored. - `skills/pstack/correct/SKILL.md`: step 3 now fails a test that still passes when the checked behavior breaks, or that fails when only the internals change. Destination equals the Cursor blob; status is `adapted`. - `skills/pstack/create-verification-skill/SKILL.md`: launch readiness means usable from an empty data dir; a missing seed precondition fails the journey; a stand-in proof lists unverified setup steps. `.pi/skills` paths stay. - `skills/pstack/principle-attack-the-premise/SKILL.md`: write the premise down and test it; follow the evidence; remove the condition. Pi backticked paths stay. - `skills/pstack/principle-fix-root-causes/SKILL.md`: fault versus symptom; a guard is a fix only when the state is legitimate; cheap always-on signal after a deep dig. Destination equals the Cursor blob; was a mimir copy. - `skills/pstack/principle-make-operations-idempotent/SKILL.md`: derive steps from existing state; judge leftover artifacts as stale by crash-surviving evidence. Destination equals the Cursor blob; was a mimir copy. - `skills/pstack/principle-test-behavior-not-implementation/SKILL.md`: two imagined edits (break behavior / change internals) replace the `undefined`-import heuristic. Destination equals the Cursor blob; was a mimir copy. - `skills/pstack/poteto-mode/SKILL.md`: the three principle summaries above follow the ported leaves. - `skills/pstack/poteto-mode/playbooks/opening-a-pr.md`: as a parent merges, move its child onto trunk per `playbooks/shipping.md` step 4. Documented skips (Cursor-only tools, token cuts that only delete examples, or excluded scripts): - Hub and most playbooks: `poteto-agent`, `subagent_type`, hardcoded slugs, `cursor-team-kit`, `/loop`, Origin, Graphite, cloud agents, ketchup replies on Autopilot-full/stack, shipping rewrite (G1 allowlist stays). - `multi-phase-plan.md`: upstream "full swarm once" plus `poteto-agent`. This pack keeps per-push swarm markers that `check-plan.mjs` still requires, and still defers rebase/patch-id to `shipping.md`. - Copied mimir leaves whose Cursor delta only deleted sections or flattened headings (encode-lessons, exhaust-the-design-space, laziness-protocol, migrate-callers, most why/how/interrogate/reflect references). - `principle-type-system-discipline`: Cursor deleted most patterns; dest kept. Flagged for G. - `pstack/docs/guide/**` and `pstack/agents/poteto-agent.md`: excluded. - `pstack/README.md`: unscoped. `pstack/.cursor-plugin/plugin.json` and `pstack/plugin.json`: out of scope / unscoped. `pstack/LICENSE`: unchanged. Flagged, not shipped (recommendation: later Pi-authored leaves if G wants them; do not copy Cursor slash-command rows as-is): - `in-your-own-words`, `ketchup`, `sharpen` - `poteto-mode/scripts/check-plan.test.ts` (bun tests; this pack already checks `check-plan.mjs` in `test/scripts.test.ts`) No new agents or docs file is in the not-yet-carried state. ## Pin move: cursor/plugins `190d287` (pstack 0.15.29) The cursor pin moved from `f713f14203cd6420cec13cd7c55c32d598dbab68` to `190d28736eff128ec94f502a22c7fe0c4b56cd37` (current cursor/plugins `main`; `pstack/plugin.json` reads `0.15.29`). The commits between those pins that change in-scope pstack methodology are the formal-verification skills (`4382e68`, `6b78e9f`, `7b353f7`), setup/help/guide sync (`f9be7aa`, `d457567`), TypeScript brand type-guard (`0317eb8`), ketchup decidability (`16541c6`, `bc14437`), sharpen restore (`75fabac`, `88d4751`), frontmatter name fixes (`a968159`), Claude Code harness (`182f2ce`, `1b80080`), and plugin version bumps through 0.15.29. The casualjim/pi-mimir pin stays `f07dd981f62c9c994a5d043ede67d6c63c721454`. Hand-ported, platform-neutral substance only: - `skills/pstack/create-formal-verification-skill/SKILL.md`: new, adapted from Cursor. Pi `.pi/skills/` paths, `/skill:` forms, and install authorization. - `skills/pstack/maintain-formal-verification-skill/SKILL.md`: new, adapted from Cursor. Same path and slash adaptations; PR push authorization on the report's change, matching maintain-verification-skill. - `skills/pstack/typescript-best-practices/SKILL.md` and `references/patterns.md`: a brand comes from a verifying type guard, never from `as`. - `skills/pstack/setup-pstack/SKILL.md`: mention once when the project has no formal verification skill; do not run create from setup. - `skills/pstack/poteto-mode/SKILL.md`: inventory 53; availability routes for the two formal-verification skills. - `skills/pstack/poteto-help/SKILL.md` and `references/recipes.md`: check-state, skill table and recipe for the two formal-verification skills. Documented skips (Cursor-only tools, named-role agents, harness, or excluded skills): - `make-bot-ui/SKILL.md`: upstream only renamed frontmatter `Make Bot UI` to `make-bot-ui`; this pack already uses that name after the Pi re-author. - `poteto-mode/SKILL.md`: upstream only renamed frontmatter `Poteto Mode` to `poteto-mode`; already that name here. Substance ports are listed above. - `setup-pstack/SKILL.md`: Cursor auto-invokes `/create-*-verification-skill` from setup; this pack only mentions the gap and names the skill. - `poteto-help/SKILL.md`: Claude Code install path, `pstack:` prefix, and guide/harness links; ketchup/sharpen/in-your-own-words table rows. - `ketchup`, `sharpen`, `in-your-own-words`: still flagged, not shipped. Upstream ketchup and sharpen bodies changed; dispositions updated. - `pstack/docs/guide/**`: excluded under Q1. - `pstack/docs/harness/claude-code.md`: **new**, excluded (Claude Code adapter). - `pstack/agents/poteto-agent.md`: excluded named role. - `pstack/README.md`, `pstack/plugin.json`: unscoped. - `pstack/.cursor-plugin/plugin.json`: out of scope (0.15.25 → 0.15.29). - `pstack/LICENSE`: unchanged. No file under `pstack/agents` or `pstack/docs` is in the not-yet-carried state after classifying the harness page as excluded. ## Pin move: cursor/plugins `8b3326d` (pstack 0.15.30) The cursor pin moved from `190d28736eff128ec94f502a22c7fe0c4b56cd37` to `8b3326d99fd4b993cb33a24dad8af03175c642c5` (cursor/plugins `main` at sync; `pstack/plugin.json` reads `0.15.30`). Twenty-seven commits sit between those pins; the ones that change in-scope pstack methodology are the new `statechart` skill (`eb4db2b`, `7698375`, `4ec7ffa`) and the poteto-help / guide / README rows that advertise it. The casualjim/pi-mimir pin stays `f07dd981f62c9c994a5d043ede67d6c63c721454`. Hand-ported, platform-neutral substance: **none**. No shipped skill file changed. Inventory stays at 53 skills and 23 playbooks. Documented skips and classifications: - `skills/pstack/poteto-help/SKILL.md`: upstream adds only a `/statechart` table row. Not copied; statechart is not shipped. Destination hash unchanged; Cursor primary source hash refreshed to the new pin blob. - `skills/pstack/statechart/**` (12 files: `SKILL.md`, scripts, TLA+ model, tests, `.mmd` fixtures): **flagged, not shipped**. Executable Node scripts spawn Java/TLC via `TLA2TOOLS`/`JAVA`, write files under a work directory, and depend on external TLA+ tooling. Recommendation: **skip** until G authorizes a Pi-adapted leaf that closes the TLC dependency with explicit install authorization; do not copy Cursor slash rows or scripts as-is. - `pstack/docs/guide/06-verify-and-ship.md`: excluded under Q1 (hash refreshed; new statechart section noted, not ported). - `pstack/agents/poteto-agent.md`: excluded named role (unchanged). - `pstack/README.md`, `pstack/plugin.json`: unscoped (0.15.29 → 0.15.30). - `pstack/.cursor-plugin/plugin.json`: out of scope (0.15.29 → 0.15.30). - `pstack/LICENSE`: unchanged. - `in-your-own-words`, `ketchup`, `sharpen`, `poteto-mode/scripts/check-plan.test.ts`: still flagged, not shipped; bodies unchanged at this pin. No file under `pstack/agents` or `pstack/docs` is in the not-yet-carried state. ## Pin move: cursor/plugins `1ba91fc` (pstack 0.15.36) The cursor pin moved from `8b3326d99fd4b993cb33a24dad8af03175c642c5` to `1ba91fc25c0977741dd0e267849e6234b037ed0a` (cursor/plugins `main` at sync; `pstack/plugin.json` reads `0.15.36`). The casualjim/pi-mimir pin stays `f07dd981f62c9c994a5d043ede67d6c63c721454`. Hand-ported substance: - `skills/pstack/statechart/**`: **shipped** as an optional skill. Checking runs through the pack `tla-ts` dependency (no Java / `tla2tools.jar`). `SKILL.md` and `references/render.md` adapted for Pi (`/skill:statechart`, no silent installer, Mermaid-first, optional image render with install authorization). Runtime scripts started as copies from the pin; `scripts/run.ts` is adapted to spawn `tla-ts`. Upstream `scripts/test` corpora, `package.json` and `tsconfig.json` are not shipped (keeps the npm pack near the prior ~128-file baseline plus the new leaves). - `in-your-own-words`, `ketchup`, `sharpen`: shipped as adapted Cursor leaves with `/skill:` invocation. - Tip deltas ported into adapted leaves: arena, interrogate (refs deleted), poteto-mode hub/playbooks/`check-plan.mjs`, poteto-help (+ new skill rows), principle-explain-the-number, reflect, swarm, create-formal-verification-skill. - Inventory is **57** skills and 23 playbooks. Still unshipped (accounted in `w3-b.json` excluded + `FLAGGED_UNSHIPPED_CURSOR`): - `ketchup-video` (new at this pin) — flag: depends on ketchup visuals and needs video render tooling outside the repo. Shipped at this pin beyond the prior 53: optional `statechart` (runtime scripts only; test corpora not packed), `in-your-own-words`, `ketchup` and `sharpen`.