Upstream provenance

Native-tool Orchestrate — 2026-10-28

Adapted pstack 0.15.36 from PR #1043, merged at c742dd61706fbacc1b95f1a20b7b988cf096ce6c (head b8fdac4934681f0aa6f7dbda3a583466f0809d4a). The MIT license retains Lauren Tan's attribution.

Orchestrate is rewritten around the coordinator's own tools. Workers spawn through the selected execution method. State lives in the host's native task tracker where available and a short user-owned notes file. PRs are watched through native PR and CI event subscriptions where the host has them. The coordinator calls no bookkeeping scripts, runs no polling loops, and never writes code. The new Topology section gives each PR one owner, the only agent that pushes, rebases, or retargets it. It also covers fresh workers per independent task, sub-coordinators only for wide areas (at most three levels), forks for built state, a cross-family verifier per PR at its current head, and a pilot unit before fan-out. The Supervision section names evidence in every brief and reads diffs over summaries. It checks live PR state before reporting and retries stalled workers by cause.

Compatibility change: the Python orch CLI (skills/z-mode/scripts/orch/orch.py, store.py) and its store and frontier tests are deleted. Integrations that called it must keep their own state. Its validator, mypy, ty, and host-contract entries are removed.

Zstack keeps action authority. An owner pushes and merges only within the user's authorized actions and follows Shipping for merges. Without merge authority or an other-family verifier, the PR stops at merge-ready and the gap is reported. The verifier may run through the native-hosts other-provider route. Cursor's cloud VM default becomes host remote isolation where available and permitted. The Project tracker and store become the host's task tracker plus .agent-work/<slug>/notes.md. control-ui and control-cli become the project's verification harness. Per-agent reasoning and speed choices are dropped in favor of the delegation contract. Upstream's retries by cause are kept but limited to zstack's lifecycle rule. That allows one retry with a fresh agent and a corrected brief, and no unrequested model switch. Upstream's messaging of workers for follow-ups becomes reuse only for agent-held state; other fix rounds go to a fresh agent. Upstream's ten in-flight workers becomes an upper bound on the lifecycle window, which starts at 2-3. Shipping's child retarget and rebase go to the child's own owner. Herdr pane assignments and the pause-safely handoff for hosts without a durable runtime are retained. Generated harness copies, Cursor manifests, and package.json changes are excluded. Native manifests advance together to 0.22.0 because a bundled tool is removed. See validation for actual checks and limits.

Proven interrogation, cross-family judging, and operator-landed queues — 2026-10-27

Selectively adapted pstack 0.15.35 from PR #986, merged at ab1f702ce8e8b6557edb56b09f52cde8bfb83f17 (head 357e07aa616a2211b2cc2c6d775ceff051c2fad4). The MIT license retains Lauren Tan's attribution.

Interrogate is rewritten around executable proof. Adversaries work in their own worktrees and must return a command and its output for each finding. The lead reruns every command in a fresh isolated worktree and keeps only failures that break the stated intent. The reviewer prompt, rubric, code-quality lens, lead-judgment framework, and Act on / Consider / Noted / Dismissed verdict are deleted with the upstream references. Fixes are still never applied. Zstack keeps its target checkout untouched, removes the worktrees it creates, and requires --no-ext-diff patches because an external diff tool otherwise yields no applicable patch.

Arena's judge must come from a model family other than the coordinator's. An inherited or same-family model never judges, and when no other-family judge runs, the coordinator skips judging and says so instead of labeling a parent-only judgment. Rejected model seats are dropped and the run tops up to two candidates. A shared model-rejection rule joins the native delegation contract: no guessed identifiers, the closest same-family model the error lists, or the inherited default.

Autopilot-stack becomes Autopilot-full without the merge. Independent items stay off trunk, an item stacks only on the item it builds on and rebases onto that parent, and the run stops at verified and green for the operator to land. Autopilot-full adds the upstream blocking rule, which blocks only on a demonstrated failure or a doc line the change made wrong, plus parked follow-ups. It also skips extra verification for a wording-only change; zstack limits that to wording that alters no behavior or instruction, since skill text is shipped instruction here. Z-mode restates the user's goal in its first message, delegated briefs, PR ## Why, and final report. Explain the number now applies before a measurement as well.

Excluded: ketchup-video, which zstack never adopted; Cursor model slugs, /setup-pstack pool ordering, and role-rule fallbacks, since zstack has no role-to-model rule; the Cursor agent-store copy rule and store.ts cleanup, since zstack's orchestration uses a local task directory and a Python CLI; and autonomous-run, babysit, and multi-phase-plan wording that zstack already covers or never carried. Generated harness copies, Cursor manifests, and poteto-help are excluded as before. Native manifests advance together to 0.20.0 for the changed workflows. See validation for actual checks and limits.

Statechart compiler — 2026-10-27

Adapted pstack PR #977 (0.15.33, merged at 34b1b7c96f09f81a21a6c78f65b7833f7dcb5b24), PR #982 (0.15.34, merged at 740d2c0786246cafac41906198f656b135194d3b), and PR #981 (test only, merged at 519d36aa98955bf680db01bcde5fdfba686dc901). The scripts, corpus, and Pele theme are copied from the 519d36aa tree. The MIT license retains Lauren Tan's attribution.

The .mjs pipeline is replaced by the TypeScript compiler: scripts/statechart SPEC.tla checks the model, draws the Mermaid chart through a view module, and reports shortest paths, with qualified INSTANCE action names (#982) and the macOS reap wait in the SIGKILL test (#981). The from-chart skeletons, --fixed/--against comparisons, mapping modules, summary.json/stories.md, and Mermaid PNG flag are removed with it, as upstream removed them. Keeping them would mean maintaining a second value parser and checker beside the one upstream now maintains.

Zstack retains explicit invocation, resolved installed paths, owned scratch folders, scoped authority, view-module review, the real-code reproduction requirement, and local Pele installation. It adds a view guard: two function keys that project to one view key, such as model value A and string "A", stop the run instead of silently losing an entry. It also fixes a test-only bug: the relative-symlink launcher test failed under macOS's default /var → /private/var TMPDIR. The earlier safe-integer guard is not carried forward, because TLC rejects integers beyond 32 bits before the parser sees them. Upstream's mise fallback for TLC and Java is kept, but the skill requires approval before the first install. Generated Claude harness duplicates, Cursor manifests, poteto-help, and upstream CI changes are excluded. The fast-check differential suite and the Pele/Mermaid render suite run only on request. All native manifests advance together to 0.19.0 for the removed workflows. See validation for actual checks and limits.

Statechart comparison and checking fixes — 2026-10-22

Adapted pstack PR #776, merged at 7698375f92a729f8bb89d113c633dce9ff9dc6db, and PR #779, merged at 4ec7ffa173eaa70d3d9467a627baf0954645a353. Upstream still declares 0.15.30; these exact merges identify the imported changes. The MIT license retains Lauren Tan's attribution.

The shared statechart skill now selects separate model configs, rejects mapping reads of absent comparison fields, marks unchecked rules explicitly, checks temporal properties separately, uses multicore invariant checks with single-worker counterexample replay, retains diagnostics after checking/publication failures, and falls back to copying when directory rename is unsupported. It also records expected spec failures and uncited action conditions. Helper naming and shared regex cleanup from #779 are included.

Zstack retains explicit invocation, Pele rendering, safe typed TLC values, exclusive scaffold creation, foreign-output protection, and its stronger requirement that fixed models retain original rules. That requirement now covers temporal properties and preserves each rule's category; property names receive the same filename-safety validation as invariants. A focused parser regression also covers TLC 1.8's named temporal-error wording, which upstream leaves unsupported. Generated Cursor/Claude duplicates, upstream packaging, and automatic TLC installation are excluded. All native manifests advance together to 0.18.1 for these fixes. See validation for actual checks and limits.

Statechart and code reproduction — 2026-10-22

Adapted pstack PR #741, merged at eb4db2b55b2b7dec3d9282754cddf0a56bf4ff10 in pstack 0.15.30 (2026-10-21 23:08:56 Arizona time). The shared skill retains TLC state-graph projection, per-invariant runs, counterexample stories, chart-to-TLA+ skeletons, model comparisons, guarded/parallel charts, and the final merged output-folder protection and regression fixtures. The repository's MIT license retains Lauren Tan's copyright.

Zstack adds explicit Codex policy, portable installed-resource paths, optional local Pele SVG rendering alongside Mermaid PNG, renderer regressions in the standard Node test entrypoint, and a stricter distinction between model findings and reproduced software bugs. A reproduced upstream collision bug is also fixed: skeleton generation refuses occupied model/config/mapping paths, including dangling symlinks, and creates files exclusively. Upstream test assumptions about Linux preload support and canonical temporary paths are adapted for macOS. The existing formal-verification generator and maintenance skill receive that same reproduction requirement. Pele is resolved from the caller's directory rather than the installed plugin cache; it is not bundled. Its API, compatibility, and security documentation informed the integration.

Cursor manifests, generated Claude harness duplicates, poteto-help, and upstream CI configuration are excluded. Native Codex, Claude Code, and Hermes continue discovering one shared skill folder. All three native manifests advance together to 0.18.0 for the new skill. See validation for actual execution and remaining limits.

Sharpen prompt rewriting — 2026-10-20

Selectively adapted pstack PR #729, merged at 88d47514b82e7c44f00f2b2a72ff6dd12158038c in pstack 0.15.29. The new shared sharpen skill retains the four asks: expert approach, original source material, specific methods and tools, and an independent check, followed by a rewritten prompt with user choices left open. Zstack makes the no-execution boundary explicit, preserves stated constraints, and keeps unknown inputs conditional. Both native invocation flags remain explicit-only.

There was no existing zstack sharpen skill. Codex packaging, Claude Code, and Hermes discover the shared skill; upstream generated Claude harness copies, Cursor packaging, and poteto-help routing are not imported. The guides and reference catalog describe the separate review-and-send handoff. All three native manifests advance to 0.17.0 for the added skill. The MIT license and Lauren Tan attribution remain intact. Upstream evaluation scores are not local behavior evidence.

Formal-verification maintenance — 2026-10-19

Selectively adapted pstack PR #716, reviewed at 011ef53277caad816e4fcf89929f3b47fdb4a050 (merge 7b353f78f6ec193bdbe48ed8059444eb09b6878e): verification drift review, per-check coverage and last verified commits, counterexample triage, and blocked-check checkpoint retention. Zstack keeps explicit invocation, isolated real-code mutations, manual model-translation limits, and verification-only edits. Dirty snapshots are recorded separately from commit checkpoints, and inferred legacy bases are not treated as prior proof. The MIT license and Lauren Tan attribution remain intact.

Reviewed PR #724 at b2a3fd3ebb23846e1c02daf55357254db984be19 (merge d457567332aa6989e753ea289f30ea2bfbc80bc0). Its TLA+ default and real-code generated-input fallback already match zstack’s generator and guide from #721. Cursor manifests, generated Claude harness copies, poteto-help routing, and upstream version numbers are excluded; the new shared skill advances all three native zstack manifests to 0.16.0 under local versioning policy. Upstream evaluation scores are not local validation.

Ketchup batch coverage and formal verification — 2026-10-18

Selectively adapted pstack #692, merged as 16541c6bc1ed6c8d808b2f78a33364003e82b967, #708, merged as bc1443790cc980d0ea857663260a831299884508, and #709, merged as 4382e687a50b8d2c5ce4d141d001186589cff8d5.

Ketchup already allowed replies beyond 150 words. The port makes coverage of each constituent item in a grouped decision explicit, including evidence, risk, and differing recommendations. It retains zstack's flexible length target, scoped read-only evidence gathering, and existing visual/table guidance instead of importing upstream presentation budgets.

The new formal-verification generator retains property selection, code correspondence, counterexamples, reproducible checks, stated limits, and deliberately broken variants. Local adaptations add native discovery and resource conventions, explicit invocation policy, isolated mutations, baseline comparison, model-versus-implementation limits, and honest draft/blocked outcomes. It remains separate from user-path verification and feature-map maintenance. The three native manifests advance together from 0.13.0 to 0.14.0 for the added skill. The MIT license and Lauren Tan attribution remain intact. This is a selective adaptation; upstream evaluations are not local behavior evidence.

pstack naming, host compatibility, and documentation review — 2026-10-17

Reviewed three merged updates from pstack 0.15.25: #702 at a968159326d1fc6b7d483491ec5c457986a20af2, #705 at 1b800800583ee8ff7f2676caf7a5cb7f0804be85, and #704 at 019a6a20592c10bf92869e5740368f578a9997fd.

Adapted #704's documentation consistency checking into the existing structural validator: current overview totals are compared with the skill, principle, playbook, and agent inventory, and current-source GitHub links in the README, guide, and reference pages must resolve inside the checkout. Existing local-link and rendered-reference-anchor checks remain responsible for their respective contracts. Regression fixtures cover stale counts and missing source targets while preserving historical links, examples, and guide subsets. This is a selective adaptation, not a copy of the upstream checker or a complete upstream update; the MIT license and Lauren Tan attribution remain intact.

No port is needed for #702: the validator already enforces folder/name equality, lowercase hyphenated identifiers, and the 64-character limit, with regression coverage. poteto-mode became z-mode, and make-bot-ui was retired. #704's reduced review frequency after minor follow-up changes already appears in the local autopilot-full playbook; Cursor's /autopilot command correction does not apply to this native-host library.

#705's persistent role-rule file would duplicate zstack's native model configuration and inheritance. Claude loads the checkout in place, and there is no help skill requiring the upstream generated guide bundle; Codex deliberately packages the skills and hooks needed by its resource contract. Native resource resolution and forge fallbacks already exist. Blanket skips with placeholders are not adopted because required acceptance checks must remain visible as gaps. No host behavior or packaged instructions change, so the native plugin version remains 0.13.0 under the repository's versioning policy.

Earlier adaptations

zstack, Zach's updated skillset, adapts Lauren Tan's pstack, copied from the Cursor plugins repository. The copied manifest declared version 0.15.5. The exact copied commit was not recorded and has not been established. The MIT license and original copyright are retained in LICENSE.

Adapted pstack PR #633, merged as e5136d8, on 2026-10-15. The four principles now use observable contracts, legitimate-state handling, direct premise tests, and state reconciliation with reliable staleness evidence. Correct and the principles reference use the same criteria. The port retains zstack's explicit invocation policies, scope limits, required validation, safe failure handling, and separate publication authority. It changes guidance only; upstream's evaluation is not a local behavior result. The existing MIT license and Lauren Tan attribution remain intact.

Selectively adapted pstack PR #634, merged as 99743e9c, on 2026-10-15. Interrogate drops the lead-judgment preamble, Why section, and Verdict Calibration, including its suggested five-item Act On limit; the caller still requires categories, evidence, and rationale. Shorter type-system examples preserve all eight principles: valid states, constructions, branding, external parsing, validation over casts, exhaustiveness, schema derivation, and strengthening only where needed. The API migration principle loses only its redundant introduction and conclusion, retaining caller inventory, coordinated removal, time-boxed adapters, and compatibility exclusions. Swarm stays unchanged because local guidance already combines concise reporting with useful worker-count and host-limit rules. Native manifests advance from 0.12.0 to 0.12.1 under the packaged-wording patch policy. This is a selective port, not a wholesale upstream update; explicit invocation, native scope and permissions, the MIT license, and Lauren Tan attribution remain intact. Upstream static validation is not a local behavior result.

The port keeps the portable engineering principles, reference prompts, decision logger, orchestration bookkeeping, and GitHub PR watcher. Workflow instructions now target Codex and Claude Code with native agents and local skill discovery.

The Cursor plugin manifest, branding assets, make-bot-ui, and dormant automations/benny pack are retired. The latter two require Cursor routines, secret cards, webhooks, or an automation runtime unavailable in this library. Recover their original sources from upstream if a replacement runtime becomes a concrete project. They are not installed or advertised as working here.

The native Codex plugin packages shared skills and a session hook through a local marketplace. The Claude Code plugin loads the same skills, hook, and agent roles in place through its own local marketplace. Public marketplace publication and a replacement agent daemon remain outside this release.

The Herdr integration is zstack-specific coordination guidance based on Herdr's agent guide, concepts, working guide, automation documentation, and upstream operational skill, reviewed 2026-10-04. It references upstream operations rather than vendoring that skill. Explicit Herdr execution applies across existing playbooks; launching a requested CLI does not establish cross-provider consensus or inherited host configuration.

The ketchup skill was initially adapted from pstack 0.15.17, commit 978ca6e9, and selectively updated with 0.15.18's visual guidance from PR #570, commit ea32d956, reviewed 2026-10-12. It retains current-conversation catch-up, read-only inspection, earlier unresolved decisions, and option/recommendation summaries. The port adds explicit-only Codex policy, scoped native history and status readers for the supported hosts, current-head CI checks, and disclosed access gaps. Plain wording is self-contained rather than invoking bro's restatement workflow. Before/after and option comparisons use existing evidence; missing captures, unknown numbers, and rendering gaps are disclosed. The stricter chat-only boundary excludes file writes, new captures, app/job launches, and generated screen images. The retained MIT license and Lauren Tan copyright cover the adaptation; this is not a full upstream update.

Ketchup's bundled decision-visual reference adapts HumanLayer's show-me implementation at commit 653b6411c1f70c275a18e37673b042ff99f67ceb (show-me plugin 1.0.1), inspected 2026-10-12. It adopts the smallest-useful-view approach, adjacent concise explanation, pseudocode, call/component/file trees, Mermaid, and diffs. It omits standalone HTML creation and the macOS open command to preserve ketchup's cross-host read-only boundary. No HumanLayer plugin is installed or required, and no upstream scripts were executed. The full HumanLayer MIT notice, including Copyright (c) 2026 HumanLayer, ships beside the adapted reference in linked and packaged skills and in the documentation site's notices.

The personal naming migration uses zstack, z-mode, setup-zstack, and z-agent. Upstream names and URLs remain in attribution and historical baseline records.

The benchmark-checklist and principle-explain-the-number skills were adapted from pstack 0.15.6, commit 23e4138. Their routing targets z-mode, preserves explicit invocation on both hosts, and uses portable host checks. This is a selective port, not a full update to that release.

The same commit supplies the fresh-agent lifecycle policy, adapted to native host limits and exclusive write ownership. Reuse remains available for costly agent-held state.

Its PR workflow updates are also adapted: built-in tools take precedence for supported operations, descriptions use concise section headings, and stack bases and readiness are read back. Repository templates, requested draft status, and publication authority still govern the workflow.

The create-verification-skill readiness fixes are selectively adapted from PR #573, head commit 8dd98e9, reviewed 2026-10-12. Launch must reach a usable state, with stateful targets completing required migrations and seed data from empty isolated state; missing seed prerequisites fail the journey without weakening it; setup supplied only by a stand-in stays unverified on the real target. The port retains native project paths, explicit invocation, helper portability, and cleanup that preserves evidence. All three native manifests advance to 0.6.1 under the local compatible-fix policy. These rules repair the existing requirement to drive the real user path and report proof honestly; they add no skill, surface, invocation policy, hook, or stored-state contract. Ketchup’s 0.6.0 visual guidance added a new presentation capability, whereas this port corrects false readiness and verification claims within the existing workflow. The retained MIT license and Lauren Tan copyright cover the adaptation; no other upstream changes are included.

The correct skill is adapted from pstack 0.15.7, commit 9511e603. It preserves repeated-mistake classification, the architecture/types/lint/tests/docs ordering, historical failure proofs, and the rule-to-enforcement table. The port adds scoped native history, explicit-only metadata for both hosts, isolated reproductions, and observed-versus-configured CI reporting. Upstream's automatic commits and future-correction trigger are replaced by task-specific authority and invocation scope. This is a selective skill port; the upstream version bump is not applied to zstack. The retained MIT license covers the adaptation.

The prompting recipes and usage guidance are selectively adapted from pstack 0.15.12, commit 807c0310, and 0.15.13, commit 2cbf5850. The port covers prototypes for unresolved questions, planning, benchmark validation, inspectable verification, repeated-mistake prevention, cost controls, and trust before unattended work. Examples use native Claude Code and Codex invocation and retain task-specific authority. The help skill, Cursor Projects, cloud-agent instructions, and Benny automation pack are excluded; native scheduling remains available only when verified and requested. No upstream version bump or invocation-policy change is applied. The retained MIT license covers this documentation adaptation.

Selective pstack 0.15.16 prompt cleanup

Adapted the actual diff at commit 00b14f1d, merged from PR #538. The upstream change touches 35 Markdown files and a Cursor manifest. This port changes 28 corresponding Markdown files; it does not replace local prompts wholesale. Upstream reports words and bytes, not measured token savings.

Upstream area	Local mapping and selection
architect/SKILL.md	Same skill. Ask for the architectural explanation how prescribes, grounded in runtime flow and callers. Retain task-sized design, implementation authority, and checkpoints.
arena, how, interrogate, reflect, why entrypoints	Same skills, unchanged. Local entrypoints are already shorter and use native contracts, scoped research, capability gaps, and task-specific authority.
benchmark-checklist, figure-it-out	Same skills. Shorten retained prose; keep macOS core-count support, counterbalanced run order and cache preparation, runtime budgets, and exclusive delegated writes.
Ten principle-* skills	Same skills: build-the-lever, encode-lessons-in-structure, explain-the-number, minimize-reader-load, model-the-domain, never-block-on-the-human, redesign-from-first-principles, subtract-before-you-add, test-behavior-not-implementation, type-system-discipline. Combine repeated explanations and checks. Keep local scope/memory boundaries, literal test expectations and concrete assertion examples, and language-independent type guidance.
interrogate/references/{reviewer-prompt,rubric}.md	Same references. Reduce repeated review instructions; preserve evidence, severity, and repository-specific tool conventions.
reflect/references/{divergent-reviewer,judgment-reviewer,tooling-reviewer,synthesizer}.md	Same references. Shorten routing and review prose, adapting source reads to native skill paths. Keep transcript injection defenses, lookup scope, invocation-policy checks, and local backlog proposals.
technical-writing, unslop	Same skills. Shorten prose while retaining writing layers, source attribution, examples, and stable rule numbers.
why/references/{epistemics,investigator-prompt,synthesizer-prompt}.md	Same references. Consolidate confidence and investigation rules. Keep all five tiers, citations, contradictions, counterfactual checks, unanswered searches, source ownership, and actual access gaps.
why/references/sources/{code-archaeology,datadog,sentry}.md	Same source playbooks. Shorten explanations and examples; keep query commands, time bounds, sampling/retention cautions, and correlation limits. These are conditional source examples, not claims of connector availability.
poteto-mode/SKILL.md	Maps to z-mode/SKILL.md, unchanged. Keep validated direct links instead of introducing a title-to-path naming convention. Local routing, native contracts, and Herdr selection remain intact.
poteto-mode/playbooks/opening-a-pr.md	Maps to z-mode/playbooks/opening-a-pr.md. Tighten description guidance and avoid commit bodies repeating subjects. Retain publication authority, safe worktree handling, built-in tool/CLI fallbacks, attachment requirements, stack bases, and requested draft state.
poteto-mode/playbooks/trace-forensics.md	Maps to the same playbook under z-mode, unchanged. The local version is already shorter and permits direct queries without mandatory conversion or delegation.
Cursor plugin manifest	Excluded. Native Codex, Claude Code, and Hermes versions advance together from 0.5.0 to 0.5.1 under the local compatible-wording policy.

Skill descriptions, invocation flags and host metadata are unchanged. No Cursor model defaults, automatic tracker writes, forced ready PRs, reset recovery, or mandatory fan-out are imported. The MIT license and Lauren Tan copyright remain intact. The branch starts from current main after ketchup PR #30; ketchup and unrelated Herdr plugin work are outside this diff. See validation for measured size and check coverage.

PR #31 review also corrected two retained example issues: valid empty-result assertions are no longer classified as vacuous, and the completion predicate now names a non-optional nullable date representation. The investigator prompt again requires verbatim quotation only when exact wording matters, consistent with its paraphrase option. Clarity repairs retain the architectural grounding, confidence tiers, source ownership, and treatment of Seer output as inference.

Selective pstack 0.15.23 updates

Reviewed the October 13 pstack changes that reached manifest 0.15.23: PRs #595, #597, #602, #604, #606, and #609. Each adopted change lands as one layer of a stacked PR series and adds its own paragraph below. Cursor manifests, the Cursor README, and poteto-help routing tables are excluded; each layer updates the local docs it affects.

Deep-dig signal. PR #604, merge 36cecbf2, adds one rule to fix-root-causes: when the deciding fact took a deep dig, propose the cheap, always-on signal that would have shown it at first look. The port adopts it verbatim and states that a proposal does not authorize adding the signal outside task scope. The principles reference example gains a matching sentence. Native manifests advance to 0.6.3 as a patch: the rule extends an existing principle's debugging guidance, while the README reserves MINOR bumps for skill, agent, invocation-policy, hook, and stored-state contract changes.

z-agent discovery. PR #597, merge b93e8117, lets poteto-agent find poteto-mode without a supplied path and choose the highest installed version. Both native z-agent roles already fell back to a local search; the port names where to search on each host, including configured host roots and project-scoped installs, follows symlinked installs to their real paths, selects the highest numerically compared manifest version with unversioned copies last, and notes that explicit-only z-mode can be missing from the skill catalog while installed. The parent still supplies the absolute path under the delegation contract. Upstream's removal of ignored mode and reminder frontmatter needs no port: z-mode never carried those keys, and the validator rejects unsupported frontmatter. Upstream's slash-menu note is Cursor-specific; local docs already direct Codex users to the skill picker and Claude Code users to /zstack:z-mode. Native manifests advance to 0.6.4 as a compatible fix.

in-your-own-words. PR #602, merge 7d8d9091, added a skill that restates the user's goal before work and waited for confirmation. PR #606, merge dbf49813, reworded it around the user's goals and problem and removed the pause. The port ships the final upstream body unchanged as an explicit-only skill on Codex, Claude Code, and Hermes. Its description names the native invocation without Cursor's slash. Docs separate it from bro, which simplifies the agent's last message, and add the upstream recipe to the local recipe list. Tests assert its installer links, Hermes registration, and Claude explicit-only flag; the existing implicit-invocation check covers its Codex policy. Native manifests advance to 0.7.0 because this adds a skill.

Shorter ketchup. PR #595, merge d00ba6ee, capped text at 250 words. PR #609, merge c86f349b, tightened that to well under 150 words including visuals and added box and decision-line limits. The local adaptation keeps plain language, checked numbers, and reduced repetition, but treats about 150 words as a target: every unresolved decision, material risk, and evidence gap takes precedence. It drops mandatory visuals and box, cell, label-word, and line-count limits. Visuals are used only when they clarify status or a decision better than prose. bro's voice is adopted without its restatement task. Compact tables, diffs, and trees work in terminal hosts; Mermaid and existing captures apply where supported. Upstream's /visualize charts and generated images are excluded. The chat-only boundary, scoped history, current-head checks, access gaps, and HumanLayer-derived visual reference are retained. Native manifests remain 0.8.0 for this unreleased layer's revised output contract.

Optional formal-verification setup offer

Also adapted the tool-selection paragraph from pstack PR #721, reviewed at 5bd8f4a2ae7c0ca0b9a2bd937839e847dd8ab534: TLA+ is the default for event-order models across implementation languages, with suitable existing tooling retained. When no verifier supports the implementation language, the generator discloses the limitation and tests real code with generated inputs instead of a handwritten function model. The local adaptation explicitly labels the fallback as testing rather than formal proof and records reproducibility details. The guide follows this distinction; the existing model-correspondence and mutation requirements remain. This addition remains within the branch's unreleased 0.15.0 version.

Selectively adapted pstack PR #715, reviewed at abe6307fd058da7f29714ed2fff5e13b926fe0d4. Setup checks existing formal-verification tooling and skills before offering the existing generator once when relevant; acceptance leads to a user-invoked handoff through the host, not automatic generation. The guides add the offer, an invariant example with bounded-proof limits, and the existing helper and evidence-retention requirements. Missing skill files alone do not establish missing formal checks. Cursor model rules, poteto-help, Origin support, generated harness trees, and references to the unshipped sharpen skill are excluded. Native manifests advance to 0.15.0 for the added setup behavior; invocation policies remain unchanged.

Selective pstack orchestration cleanup

Reviewed the October 14 pstack orchestration cleanup from PR #613, merge ce649c03, manifest still 0.15.23. Three changes are adopted, each as one layer of a stacked PR series with its own paragraph below. Excluded: per-spawn effort and fast-mode selection, which conflicts with the local rule that native delegation inherits the host's configured model unless an override is requested; Cursor grok/composer model-family fallbacks; Origin forge support, including the worktree audit's Origin PR column, since gh already reports draft PRs as OPEN and the audit holds them; and wording fixes to text z-mode never carried.

Owner-approval gate. A completed Code Review Gate check run with an ACTION_REQUIRED conclusion now maps to the owner-approval gate on both the rollup and fast-check read paths, so the watcher neither fails on it nor waits on a human. Other checks with that conclusion still fail and keep their reported state. Because GitHub then reports the head rollup as FAILURE and the PR as BLOCKED, that aggregate refusal is also excused while such a gate is present; a pending gate does not excuse it. Ruff cleanup on this layer moves hooks/xray.py to datetime.UTC, raising the session hooks' Python floor from 3.10 to 3.11; README and the setup guide state it. Native manifests advance to 0.9.0 as a MINOR release because that changes the hook runtime contract.

Frontier without Graphite. orch frontier set no longer reads Graphite metadata. It asks gh for the repository's default branch as trunk, walks the PR base chain from the checkout's current branch down to trunk, then up through child PRs, and takes head SHAs from local refs/heads branches. A merged parent that its child still targets stays in the chain, and so does a child merged into its parent's branch; a reused branch resolves to its open PR, else its most recently created one. A head that targeted a branch counts as its child only if that head's current PR, from its full history, still targets it. A closed PR with an open or merged descendant, reached walking up (even through further closed PRs or beside an open sibling) or walking down, fails rather than leaving an orphaned chain; descendants are judged by their current PRs too. Fork PRs that reuse a stack branch name are ignored, GH_REPO is cleared so gh reads the checkout's repository, and each query asks for up to 1000 rows. On trunk the chain is empty. Elsewhere, a branch with no PR, more than one live child, a base loop, or a missing local branch fails instead of guessing. Upstream's --forge origin option is excluded with the rest of Origin support. The orchestrate playbook and playbook reference drop the Graphite prerequisite. Native manifests advance to 0.9.1 as a compatible fix.

Shipping safety. Shipping pins every force-push lease to the SHA last pushed or taken over and stops on an unknown push. Right before a merge it checks whether trunk's changes can affect the patch: untested code that would land forces a rebase, a changed caller or dependency runs the patch's checks on the merge result, and a third rebase after merge prep is held and reported. After a parent merges, any child is retargeted to trunk and only the child's own commits, counted from the parent commit recorded beneath them when the link was created or rebased rather than the parent's possibly rewritten final head, only when the landing authority covers the child, and only when that range has no merge commits, move onto the merge result before the patch re-check. Autopilot-full runs full independent review once at the code-ready head and repeats it only for new behavior, a redesign, or a large new diff; a rebase, conflict resolution, or minor fix gets one verifier. Autopilot-stack hands a link to the coordinator once its owner stops writing; only the coordinator then rebases or retargets it, and pushes need publication authority. Upstream's longer procedural rewrites of these playbooks, Cursor cloud-agent and /loop mechanics, and forge-specific merge commands are excluded; the local playbooks keep their shorter shape and explicit merge authority. Native manifests advance to 0.10.0 as a MINOR release because Shipping and the autopilots gain new procedures, not only wording.