Skip to content

Releases

v3.1.1: Installation and handoff fixes

Choose a tag to compare

@allemaar allemaar released this 06 Nov 10:40

Fixed

  • Selective installer runs check declared sibling dependencies before copying and report missing names without silently expanding the selected installation.
  • Forced replacements detect intervening edits and retain displaced copies outside skill discovery; failed runs preserve recovery evidence instead of blindly restoring or deleting it.
  • Degraded handoff activations retain premise and imported-command authorization checks. Receivers bundle the canonical retro schema locally; acceptance coverage keeps its warning behavior and successor offers follow their declared conditions.
  • Prime skills declare their existing Caller Options prerequisite. Explore summaries target three feasible options without inventing alternatives, and the statistical rewrite example preserves null-model assumptions.
  • Reader flows, installed-copy comparison coverage, first-install setup, pinned parser examples, update-method effects and Handler terminology match their owning procedures.

Updating

  • Select required sibling skills explicitly when installing into an incomplete runtime; the installer now stops before copying and lists missing dependencies rather than adding them automatically.
  • Forced replacements retain recovery copies under open-skills-backups beside the runtime skills directory. Keep these copies until separately reviewed for cleanup; failed activation may require inspected manual recovery.
  • No skills were added, removed or renamed. Existing installed copies change only through their selected update method; publishing this patch does not update local installations.

open-skills 3.1.0

Choose a tag to compare

@allemaar allemaar released this 06 Nov 08:08

Added

  • implement-task delivers one authorized outcome without requiring a plan file; governing approved plans and explicitly selected handoffs retain their own contracts.
  • update-open-skills inspects existing installed skills, selects a stable release and preserves local changes, backups and installation ownership. Already-current links need no replacement; unresolved conflicts remain explicit.
  • A shared glossary, scope statement and installation/update guide explain the library, method-specific prerequisites and source-versus-directory visibility.

Changed

  • Every skill guide now explains its mechanism and role, answers task-specific questions and names observable successful results, while preserving use/non-use, examples, flow, subagent conditions and source links.
  • README navigation and task journeys include the new capabilities and retain the collection's personal, practical voice. Generated catalogs, guides and directory grouping stay aligned with all 92 skills.
  • Installation guidance distinguishes editable copies, source links, the bundled installer and manager-owned updates. It names the separate current external-CLI prerequisites and does not promise unobserved automatic updates or directory reach.

Fixed

  • Reader flows put full-set comparison, decision stops, complete voice handoffs and current-record checks before dependent recommendations, actions or output. Conditional examples, catalog binding and preflight requirements match their owning skills.
  • Installation guidance explains first installation of the new updater and separate selection of newly added skills; guide links lead directly to that shared guidance.
  • Insight-explore’s option-count check accepts an evidenced bounded-search result with one or zero feasible candidates, matching its existing rule and prose without permitting invented alternatives.
  • Guide flows preserve required delegation stops, high-stakes question preparation order, concurrent-edit stops and existing plan formats.
  • Historical Grok validation notes omit private host and session locators while retaining dated observations and limitations.
  • The plan template uses the skill-folder execution command; human-output drops an orphaned footer fragment and its four bundled contracts use unambiguous section numbering.

open-skills 3.0.1

Choose a tag to compare

@allemaar allemaar released this 05 Nov 09:22

Fixed

  • Map-check establishes readability of every governed Markdown entrypoint before emitting usable resolution, including plain links without fragments. Unreadable entrypoints return check-incomplete; readable plain and fragment targets retain their existing resolution.
  • Repository-root helper commands match the bundled fixes for required recap status and unique paths, malformed provenance pairs, point-forecast warnings, and receipt-marked figure checks. Repository-relative schema and example paths remain intact.
  • Verify discovery metadata now agrees with its procedure and guide: a missing plan retains intent/source checks and cannot yield a clean verdict. Affected docs, source fingerprints and discovery projections are synchronized.

open-skills 3.0.0

Choose a tag to compare

@allemaar allemaar released this 05 Nov 08:34

Breaking and migration

  • Mailbox exclusive claims require settlement. Existing claim order is no longer enough to start exclusive work. Obtain a causal grant from the declared task/artifact writer or current Handler naming the exact holder, scope, output and finite lease; reconcile conflicts before acting. FULL retains its implemented session, handshake, claim and rollup mechanics, but frozen membership/recipient snapshots, quorum authority and built-in authentication are unavailable. Update room instructions and callers relying on those promises before continuing dependent work.
  • Unknown topology is absence, not trunk. Consumers of orient-record/1 must treat an omitted optional on_branch as unknown, preserve the uncertainty gate, and omit unsupported dependent layout/fields and forecasts. Known true/false meanings and the record version remain unchanged. Update consumers that assume every topology field is present or coerce absence to false.
  • Rename/copy recaps stop explicitly. diff-recap returns UNSUPPORTED INPUT for rename/copy statuses, scored forms and two-path input; it emits no single-path recap for them. Keep the diagnostic distinct from NO DIFF. Callers needing those inputs must handle the unsupported result; do not strip scores, discard paths or disable detection to bypass it. Ordinary A/M/D rows retain their format.

Added

  • Six skills: ask-me for constructive interviews with optional records, build-prototype for bounded runnable experiments, debug-task for executed diagnosis and authorized repairs, write-spec for conversation-to-spec synthesis, define-terms for domain language and selective decision records, and teach-topic for evidence-based learning progression. All have dedicated guides.
  • Six additional workflow recipes cover expert questionnaires, workflow discovery, interactive coauthoring, TDD where useful, module-boundary enforcement and PR evidence. Design scope records separate reasons and reconsideration conditions for three deferred capabilities.
  • Two worked recipes cover creating a small skill and previewing a portable library, with runtime-specific installation choices, sample input, expected results and explicit limits.
  • skills.sh.json groups all public skills using the existing family taxonomy. The spine generator maintains it with the catalogs; this changes directory presentation, not installation or verified reach.

Changed

  • The README leads with practical usage and a personal account of the collection; the workflow index and generated guide introduction make entry points clearer without changing skill behavior or installation permissions.

  • skills-help can offer short journeys from installed capabilities. Code reviews in cold-review distinguish applicable standards from approved specification while retaining general findings and the existing review budget.

  • New-session handoff is compact without losing acceptance, source or canonical-save safeguards; unsaved drafts remain explicitly incomplete.

  • human-rewrite may explain a missing premise using explicitly supplied established context with separate attribution. Shared human and orient contracts preserve source facts, uncertainty and their existing render boundaries.

  • plan-execute distinguishes task evidence from plan approval and checks the selected current candidate, while retaining every unspecified-decision and drift stop.

  • Skill Smith creation and behavior-changing updates, plus Skill Library Creator setup, clarify only consequential unresolved choices after preflight. Complete briefs skip the interview; deeper questioning is opt-in. Declined or unanswered optional features stay off, and answers do not grant separate action permissions. Assessment, diagnosis, formatting-only updates and existing generated managers retain their behavior.

Fixed

  • Mapping retains dotted Markdown stems alongside eligible attachments, caches exact-path lookup, rejects stray or duplicate M3 declarations and malformed scanner frontmatter, and fails incomplete on unreadable governed entrypoints. Orientation rules consistently omit unsupported topology while retaining independent evidence; ordinary in-session extraction no longer inherits a leaf-only primed-brief prerequisite.

  • Extraction now coordinates gathering and cold review at the root without nested mandatory delegation. Mailbox exclusive claims require causal settlement; unsupported FULL membership/recipient snapshots, quorum and built-in authentication are explicitly unavailable. Existing CORE and implemented FULL mechanics remain.

  • Mapping owns one M3/address contract across its independent engines, preserves invalid/capped outcomes and raw style/alias disagreements, and documents full-scope rechecks separately from mutation bounds. Unknown orientation topology omits the optional boolean rather than asserting trunk; existing known values and record version remain.

  • Domain checks separate registration, parking, verified sale and purchasability, use exact operator matches and retain authoritative lookup for every bulk candidate. YON writing declares versioned field/profile dependencies and repairs extended examples; diff-recap explicitly stops on unsupported scored rename/copy or two-path inputs without changing ordinary rows.

  • Worked examples and discovery now preserve stated precision, coverage, provenance and independence; headless historical login wording and all five source projections agree without refreshing access evidence. Inline routing cards and conditional module/API handoffs are explicit.

  • Retrospectives apply the repository’s adopted documentation conventions and disclose incomplete catalog coverage instead of assuming library-specific formats or an empty index.

  • Handoff activation and execution clauses, audit search coverage, mailbox recovery, map-maintenance early exits and copy-versus-production records now agree with their existing owners; no implicit push-only operation is introduced.

  • Verification, exploration, freshness and priming now preserve unavailable evidence, honest candidate sets and pointer closure; orientation audit proxies follow acceptance, topology uses one worker, and ETA point warnings no longer exempt high confidence.

  • Helper and review routes now preserve subject fit, required lens coverage, canonical update ownership and honest missing-owner outcomes; selective human-family installations can finish without inventing an unavailable improvement pass.

  • Dependency maintenance resolves existing local tools, checks actual exit status and uses bounded conflict repair; existing headless package membership and library target subsets stay exact.

  • Artifact edits preserve current identity, unrelated content and recoverable preimages; objective gates precede writes, selected amendments stay exact, and source-owner and publication-target boundaries are explicit.

  • Existing mode, priming and insight dispatch owners now require bounded observable cancellable workers and honest partial outcomes; prime-sweep capture identities are unique per invocation.

  • Quota-reader shutdown now cancels owned reads within a bounded cleanup window and preserves completed results and partial cleanup evidence. Budget checks use an already-installed reader without package acquisition, reject numeric flags missing values and anchor the wrapper to its skill directory.

  • Recap checks reject duplicate paths and missing status; orient checks reject malformed nonempty provenance pairs. Figure-only checks recognize receipt-marked figures while preserving other code and quotations; shared helper copies stay synchronized.

  • Question defaults cannot supply missing consent, coupled-constant audits distinguish inspection from authorized repair, and cold reviewers remain read-only. Handoff verification inspects command effects within current authorization; legacy source-heading and drift-warning predicates agree.

  • Handoffs, retros and plans preserve existing artifacts on collision and require current identity and recovery for selected revisions. Interview records reuse their destination owner’s protection; synthesize route installation also checks backup and staging paths.

  • Adversarial reviews permit evidenced no-material-finding outcomes, merge convergent findings and rank up to three supported issues without forcing severity; later-pass panel options meet the existing minimum.

  • The Windows Antigravity supervisor accepts observed read_file byte-offset pagination with integer and exact-path checks, while rejecting unknown fields and combined line/byte addressing. Grok guidance records the observed rawInput.target_file read shape. Timeout and incomplete-result checks remain in force.

open-skills 2.0.0

Choose a tag to compare

@allemaar allemaar released this 30 Oct 21:01

Breaking

  • Skill Smith replaces two entrypoints. skills-audit and new-skill-creator are retired without compatibility entrypoints. Use skill-smith for skill authoring and maintenance, including useful rename and migration guidance. Install the complete replacement folder, update explicit references, and preserve local customizations before removing old installations. Library setup and publication retain separate owners.

Added

  • Skill Smith provides task-selected creation, assessment, consolidation, diagnosis and update procedures. Optional type recipes, nested primer navigation, record-keeping and scoped data handling load when relevant; assessment remains read-only unless edits are explicitly in scope.
  • Skill Library Creator creates portable libraries with a self-contained family manager, guided setup, templates and link tools. The manager operates independently of the creator, supports selected local installation scopes, and preserves customizations through reviewed upgrades.
  • Five insight skills extend the family: premortem for future failure paths, root cause for observed problems, reference class for outside-view estimates, test assumption for bounded test design, and value of information for deciding whether further evidence is worth its cost.
  • Seven headless skills cover bounded Claude, Codex, Gemini and Grok CLI work, provider/model selection, model-state refresh and usage checks. The family includes provider-specific execution, supervision, authentication and evidence handling.
  • The write-copy family adds a promotional-writing router and an event-video route for campaign copy, voiceover handoff and reusable event details.
  • A guide for every skill explains purpose, use and non-use cases, an invocation example, workflow, outputs and subagent use. Guides share the catalog inventory; cited-source fingerprints and generated-file checks expose documentation drift.

Changed

  • Insight methods use evidence more selectively. Cross-examination reuses valid findings, distinguishes missing evidence from missing options, selects at most one relevant specialist and keeps deeper loops and fan-out behind their existing gates. Assessment makes evidence, source dependence, optional decision methods and insufficient-evidence outcomes explicit. Adversarial passes keep substantive critics fresh while retaining their required recurring self-view.
  • Worked examples load conditionally in human-draw, human-merge and human-rewrite. Core procedures and safeguards remain available on required routes; moved example content is preserved. This reduces mandatory reading on ordinary routes without claiming measured token savings.
  • Public discovery describes 84 skills and 53 protocols. Version sources, plugin metadata, README, catalog, relationship graph and guides move together. The contributor guide now includes a maintainer closeout checklist.
  • Coordination and improvement guidance is clearer. Mode ownership and worker eligibility are explicit; self-improvement distinguishes concrete useful practices and fixed pitfalls while retaining approval and deduplication boundaries.

Fixed

  • Published DCO baseline recovery: advance the enforcement baseline to d2a6d03 after the already-published commits b8ec680 and d2a6d03 landed without sign-offs. Their missing certifications remain recorded; history is not rewritten or retroactively certified. Every later non-merge commit still requires a DCO sign-off.

  • Planning, cleanup and Git scope: named-plan selection, safety prerequisites, physical cleanup targets, authorization boundaries and exact outgoing commit scope are explicit. Plan examples now provide the concrete actions their consumer requires.

  • Handoff and orientation: confirmed-empty changes are distinguished from absent or unconfirmed evidence; output forms, manifests, freshness and validation agree more closely. These changes preserve recovery and verification requirements.

  • Mapping and evidence: bounded discovery reports missing or failed evidence honestly; checker descriptions and arithmetic examples match what their checks establish. Priming separates excerpt matching from independent source verification and retains unresolved contradictions.

  • Provider tooling and dependencies: valid quota results survive partial failures, runner finalization stays within its budget, stale process PATH handling can find installed CLIs, and optional tool dependencies no longer imply silent installation.

  • Conformance uses the parser's version-2 line in local fallback and CI rather than an unbounded latest-version fallback.

open-skills v1.8.0

Choose a tag to compare

@allemaar allemaar released this 10 Sep 11:50

Added

  • Instruction quality control is public and portable. skills-audit moved from the maintainer-only library into this pack, and directives-audit adds the same audit-first, explicitly gated housekeeping model for agent directives.

Changed

  • plan-phases now gates safety prerequisites before the action. Authorization, no-follow target proof, backup or readback readiness, live-data scope, and destructive-target confirmation get a blocking pre-action /verify gate; outcome gates stay after the step.

  • The MYK kernel moves to v2.4 and its tools stop trusting prose. map-check and map-scan now name map-rules and its routed references as their protocol source instead of a hashed kernel string, detect file-versus-directory case-fold collisions without requiring those leaves on disk, and keep unqualified CLEAN banned. Lint and the spine generator gain first-class requires-skills support so a skill's declared dependencies are validated and projected onto every generated surface.

  • MYK and synthesis now load only the rules needed for the task. map-rules routes from a small core into operation-specific references; synthesize loads one need-specific shape plus its grammar for substantive output. Their opt-in installers now add minimal routes only to selected, supported directive surfaces.

Fixed

  • leak-guard now scans exactly what could ship. It enumerates tracked plus untracked-but-unignored files through git ls-files instead of walking the whole checkout, so a gitignored local ledger no longer fails the guard while an untracked file that could be committed still does. Without git it falls back to the full walk.
  • Six protocol stamps caught up with their last edits. agent-mailbox, domain-check, extract-signal, map-maintain, new-skill-creator, and plan-phases carried a trailing @STAMP older than their last modifying commit; yon-dag now reports 0 stamp-drift warnings.

open-skills v1.7.0

Choose a tag to compare

@allemaar allemaar released this 17 Aug 16:59

Added

  • Six skills extend output, evidence, and delegation work. agent-output defines compact agent-to-agent reports; prose-audit diagnoses repetition and filler without rewriting; synthesize compiles faithful, purpose-matched Handler-facing output; synthesize-init installs that report layer into standing directives through an explicit reversible gate; freshness-contract decides when cached evidence must be refreshed; and route-task records experimental capability-based delegation decisions.
  • Map Your Knowledge (MYK) ships as a complete five-skill family. MapYourKnowledge.md introduces the system, with map-rules, map-this, map-init, map-check, and map-maintain all present in this release. This corrects the earlier Unreleased note that described only three as shipped and two as future work.

Changed

  • The human-output family and report layer now share one public routing model. /synthesize is the front door for purpose-matched delivery, /human-output retains deep writing craft, /human-rewrite owns repair, and the adjacent audit and agent-report skills keep their narrower jobs.
  • Public and generated discovery surfaces now agree on 68 skills. Forty include protocol.yon; 28 are Markdown-only. Catalogs, taxonomy, plugin metadata, README, threat model, conformance report, and social-preview source carry the same live inventory.

open-skills v1.6.8

Choose a tag to compare

@allemaar allemaar released this 28 Jul 20:38

Agent Mailbox reliability

  • Separates message disposition from CTA lifecycle so deferred, audited, and reply-expected work cannot disappear behind a clean cursor.
  • Adds bounded boundary reconciliation, trailing addressed-message audits, and explicit readiness checks.
  • Makes collaboration continuity defaults simpler and fail-visible across Claude, Codex, and other runtimes.
  • Regenerates the public discovery surfaces from the settled contract.

open-skills v1.6.7

Choose a tag to compare

@allemaar allemaar released this 27 Jul 20:00

A patch release that keeps Agent Mailbox reliability in a small mandatory kernel while making collaboration patterns composable and optional.

  • Adds resource-first activation, working agreements, composable arrangements, CTA signal extraction, reusable recipes, and an experimental pattern bank outside normal routing.
  • Preserves whole-inbox reconciliation, durable per-message dispositions, cursor-as-accelerator semantics, exact causal envelopes, and bounded listener behavior.
  • Adds participant-owned workspaces with safe pen transfer, owner-only scratch, and bounded Handler-directed remediation.
  • Adds self-healing guidance for contract drift without making optional resources kernel dependencies.
  • Clarifies that no-reply-required does not disable Work-or-Listen and that useful independent review requires observable challenge behavior.

Validation: strict YON parsing, semantic DAG checks, 57-skill lint, generated-spine consistency, privacy leak guard, staged patch integrity, isolated installed-copy dogfood, and independent SOL/OPU review with bounded final repairs.

open-skills v1.6.6

Choose a tag to compare

@allemaar allemaar released this 27 Jul 10:32

A patch release that makes Agent Mailbox collaboration quieter, safer, and more recoverable across agent runtimes.

  • Native task wake is now two-phase: finite candidate sets remain wake-pending until every member has a durable outcome and the same listener run completes a clean recheck.
  • Listener runs require current per-run authority even when reusing a task; lease recovery compares full identity, retained stale leases degrade honestly, and task archival remains informational.
  • Proven listener engines are inspected before rebuilding, while capability tier stays advisory and mechanical duties plus failure reporting remain mandatory.
  • Participant workspaces now use portable callsigns, create-once artifacts, owner-only scratch, immutable successor continuations, causal path/hash delivery, privacy constraints, and evidence-based legacy migration.

Validation: strict parser and semantic-DAG checks for the base and all three lazy protocols, pack lint, generated catalog consistency, privacy leak guard, strict tag consistency, isolated real-installer dogfood, and independent SOL/OPU cold review.