release: sync Cursor pstack 0.15.0 as Open Pstack 1.4.0

[ericlitman]

Closes #61. Tracks PSTACK-2.

What changed

Open Pstack 1.4.0 syncs Cursor pstack 0.15.0 (71ed0d1), four upstream commits after the recorded 0.14.7 point. How's critique mode and the how critics role are retired, and the two critic references are deleted. Two principle leaves are added, principle-attack-the-premise and principle-test-behavior-not-implementation, each with a narrow local correction recorded in CHANGES.md. The rest is upstream's density and punctuation passes applied on top of the port's Claude Code and Codex adaptations, plus the exact 361,140-byte logo and the verbatim upstream README.

Out of scope: make-bot-ui, the invocation-blocking flags, the Cursor-only model defaults, the Claude manifest logo field, and upstream's new "about 40 lines" squash-body cutoff. All existing exclusions in UPSTREAM.md stand.

Plan: docs/plans/upstream-0.15.0.md. Lever: scripts/upstream-audit.py inventories the range (96 paths), and scripts/upstream-merge.py applies it mechanically (24 verbatim, 32 clean three-way merges, 32 files with 65 hand-resolved hunks, 2 deletions). A reviewer can rerun both against 56bfd14 and 71ed0d1.

Blast Radius

Every pstack skill body changes prose. The rules the port depends on (PR 44 shipping guards, Sol defaults, provider dispatch, user-invocable: false) are covered by tests/skill-collision-repro.sh; removing one protected shipping instruction makes it fail, restoring it passes. A user with an old 16-row model sheet keeps working; the how critics row has no consumer, and setup reports it as an unknown role before any write.

Verification

• Bun tests (158 pass), strict typecheck, static invariants, and plugin validation pass on the host.
• The exact candidate is installed in every affected harness.
• The changed behavior passes from each real user surface.
• The installed version, action, and observed result appear below.

Host gates on dfdbd60: bun install, test, typecheck, manifest parse, static invariants, diff --check, no conflict markers, logo bytes and size, README verbatim, critic refs gone, no how critics role, 54 skills / 23 principles, no Cursor-only terms in skill bodies, claude plugin validate. All pass.

Live evidence:

Claude Code. Candidate dfdbd60 loaded from the worktree with claude -p --plugin-dir plugins/pstack in a fresh session on a one-file fixture repo. The init event reports plugin pstack version 1.4.0 from that path. Action: "Invoke the pstack:how skill and explain how upstream-audit.py classifies a changed path. Simple question." Observed: the session invoked Skill pstack:how, read how/references/explainer-prompt.md and provider-dispatch.md, took the simple path, and started exactly one pstack:pstack-fable-max explainer lane. No critic lane, no critique rubric read. It returned a grounded explanation of the classification table (Overview, How It Works, Gotchas). Transcript archived locally.

Codex. Same candidate skill tree exposed to a fresh codex exec session as the fixture repo's .agents/skills (workspace-write sandbox with network). Action: same How question. Observed: Codex read how/SKILL.md, the explainer prompt, provider-dispatch.md, and codex-tools.md, chose the simple path, and dispatched one read-only external claude:fable@max explainer through pstack-runner. Receipt: status complete, exit code 0, provider reported claude-fable-5-1, model verification passed. No explorers, no critics, no parent-model substitution. Two earlier attempts under read-only and no-network sandboxes failed at dispatch for sandbox reasons (runner cannot write receipts; child claude got ENOTFOUND), and in both the session refused to substitute itself, which is the contract.

Retained port rules. PSTACK_STATIC_ONLY=1 bash tests/skill-collision-repro.sh passes on the candidate. Deleting the "Stop before any rebase, force-push, retarget, arm, or merge" instruction from shipping.md makes it fail with "Shipping can mutate the frontier before every merge request is confirmed off"; restoring it passes.

🤖 Generated with Claude Code

[greptile-apps]

Too many files changed for review (103 files, 100 file limit).

Bypass the limit by tagging @greptile-apps to review.

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request

[openswebot]

Open SWE Review found 3 potential issues.

Open in Web

[ericlitman]

Open SWE round 1, adjudicated against the code. All three findings held and are fixed in 4cb58f6.

• user-invocable: false on how/why/unslop/typescript-best-practices. Real regression. origin/main never carried the flag on those four, and UPSTREAM.md excludes every invocation-blocking flag for them. It entered through commit 0044ae5 via a mislabeled rule in the implementation brief. The four lines are removed; the principle leaves keep the flag as before. CHANGES.md now says so.
• Tracking. Issue Sync Cursor pstack 0.15.0 #61 predates this PR and is the repository's canonical tracker per AGENTS.md. The plan now names it, and the PR closes it. Linear PSTACK-2 is the operator's batch queue and mirrors the same work.
• Lever auto-applies diverged add/delete rows. Real, reproduced: a synthetic audit with all four add/delete rows relabeled port-diverged-review mutated the tree. The script now reports those rows as conflicts and mutates nothing (rerun: conflicted 4, removed 0, zero files changed). Behavior for this range is unchanged: 24 verbatim, 32 clean, 32 conflicted, 2 removed.

Host gates rerun on 4cb58f6: all pass. One Bun run showed a single transient failure that did not reproduce on the immediate rerun (158/158); recorded in the batch ledger.

[ericlitman]

Open SWE round 2: two findings confirmed resolved, one still open on the plan doc's Linear mention. Fixed in the next commit: the plan now names only GitHub issue #61, and its stale "prepare an upstream proposal" paragraph now records that the proposal was dropped at execution. The repository's own docs follow AGENTS.md; the operator's queue lives outside the repository.

[ericlitman]

/gavel rerun