agent-ready-repo — Now

[core][3.1.14] — 2026-10-21

2026-10-21T00:00:00Z

The finding-adjudication commands classify, raw-classify, inspect, and record now live in review-artifact.py. They print exactly what the loop-cohort.py review verbs print, and those verbs keep working.
Without cohort state, review-artifact.py record --run-id <uuid> keeps one record file per review round under .context/reviews/<run-id>/ and counts the review budget from those files. Lifting the cap needs an owner authority reference.
The reviewer roster, each role's boundary and fallback, and the high-risk definition now also live in review-rules.json. A check fails if they drift from the work-loop skill text.
The supervisor, still dark-launched, now picks reviewers from that rules file and a recorded selection. It closes review only when structured review readiness says so, and it records repair, fixed-by, and accept-risk decisions.
Policy families can now be selected by supervisor stage, or by the new PRE-APPROVAL constant. Each key selects what its engine states select today.

[core][3.1.15] — 2026-10-21

2026-10-21T00:00:00Z

The offline intent view can now show what is ready to work on: pipe the workspace-status status output into the export with --readiness - and the view lists active, ready and blocked items, marks the one work-loop would pick, and shows each blocked item's findings in What is broken. If the check fails, the page says why and the export still succeeds.

[core][3.1.13] — 2026-10-21

2026-10-21T00:00:00Z

The work-loop traceability check now reads an intent's Kind:, Level: and Slug: only from its header block above the first ## heading, and ignores commented-out lines. It places each intent on the outcome, opportunity and capability ladder exactly as the intent navigator shows it.
A retired (tombstoned) intent no longer appears as a node on the ladder.
An intent whose Kind: sits only in its body, in a comment, or under a lowercase kind: label no longer counts as a ladder rung. Move the Kind: line into the header block to keep it on the ladder.

[core][3.1.12] — 2026-10-20

2026-10-20T00:00:00Z

A spec's Discovery: line, which names the intent or research the spec came from, can now name its intent as outcome:<slug>, opportunity:<slug>, capability:<slug>, or intent:<slug>. Close-out still finds the same parent intent.
A spec you add or change is refused before the work loop completes the change when its Discovery: points at an intent by file path or link, or names no existing intent of that kind. Research and notes paths stay plain paths and are not checked.
Older specs you do not touch keep working with their legacy paths.

[product-engineering][2.0.2] — 2026-10-20

2026-10-20T00:00:00Z

The longer shaping route now continues from the capability map into intent decompose, and align-value-stream proposes a hard-limits README section.

[converters][0.9.9] — 2026-10-20

2026-10-20T00:00:00Z

The render-proof skill's locked markdown-it moves to 14.3.3, which fixes two high-severity quadratic-time rendering advisories.

[architect][0.15.19] — 2026-10-20

2026-10-20T00:00:00Z

Architect-design can write a solution repo's shared system reference and offers a physical architecture with a bill of materials after a design.

[product-engineering][2.0.1] — 2026-10-20

2026-10-20T00:00:00Z

When the architect pack is installed, intent de-risk and intent decompose now offer architect-design before the work is cut, and you can decline once; teams without it see no change.

[core][3.1.11] — 2026-10-20

2026-10-20T00:00:00Z

Authors and agents get a check that catches untyped or broken Parent intent: and Brief: pointers in the artifacts a change adds or modifies, so a bad link to an intent or brief is fixed before it merges.
Artifacts a change leaves untouched are never checked, so older documents with legacy pointers do not block new work.
Skills now write typed pointers by default, so new specs and briefs start out well formed.
The work-loop finish step and the repository gate both run the check, and an incomplete check counts as a failure, not a pass.

[core][3.1.10] — 2026-10-19

2026-10-19T00:00:00Z

Picking an item in the offline intent view now shows a ready-to-paste prompt that starts the named skill on that file, copied with one click or selected for you to copy by hand.

[code-intelligence][0.1.10] — 2026-10-19

2026-10-19T00:00:00Z

Runs on Wicked Estate 0.30.0 — re-run the install command to upgrade. The documented capabilities are unchanged; this release tracks the new upstream release and raises the floor and pin to 0.30 / 0.30.0.

[core][3.1.8] — 2026-10-17

2026-10-17T00:00:00Z

The offline intent view now answers "How far along is each capability?": every capability under its product strategy, with how much of the work beneath it is delivered, being built or not yet started, and a Browse tree you can narrow to open or done work and regroup by capability.

[core][3.1.7] — 2026-10-17

2026-10-17T00:00:00Z

The offline intent view has a calmer, more polished look: soft panels, tinted state labels that still read as a word and a shape, a clearer detail card, and short fades that stop when your system asks for reduced motion.

[architect][0.15.18] — 2026-10-15

2026-10-15T00:00:00Z

Architect-diagram can export a saved Mermaid diagram to an editable draw.io file on request.

[core][3.1.6] — 2026-10-15

2026-10-15T00:00:00Z

Pack page rewritten to the catalogue's pack-page template.

[catalogue-curation][0.4.7] — 2026-10-14

2026-10-14T00:00:00Z

The propose-catalogue-pack shell check now names agentbundle catalogue verify instead of a removed lint script.

[experience-design][4.2.0] — 2026-10-14

2026-10-14T00:00:00Z

Experience Design now writes the words your screens show: ux-writing drafts error, empty-state, button, and label copy keyed to each screen and state.

[core][3.1.5] — 2026-10-14

2026-10-14T00:00:00Z

See what an intent waits for and what waits for it, in the query, its text output and the offline view.
Pick an intent in the offline view and its Dependencies part lists "A waits for B" lines, with Show and Steps controls; dependency problems also appear in "What is broken?".

[core][3.1.9] — 2026-10-14

2026-10-14T00:00:00Z

Plans can now declare the external effects a task performs, and stub fences carry the command and exit codes that prove a test fails first.

[atlassian][0.10.2] — 2026-10-14

2026-10-14T00:00:00Z

The pack's activation-eval comment now names the live agentbundle pack evals run command instead of a removed script.

[product-engineering][2.0.0] — 2026-10-14

2026-10-14T00:00:00Z

ux-writing now ships in the experience-design pack, beside the screen flows and brand register it writes from. After upgrading, reinstall product-engineering once so it lets go of the old copy; the product-engineering 2.0.0 release note has the steps.

[product-engineering][1.0.3] — 2026-10-14

2026-10-14T00:00:00Z

Requirements ingest now says when it applies: a multi-slice or cross-repository document goes to author-delivery-brief create, a single-feature one to new-spec; shipped skills state the content-not-runtime rule directly instead of citing internal records.

[product-engineering][1.0.5] — 2026-10-14

2026-10-14T00:00:00Z

Every product-engineering skill now ships activation evals, so you can check that each shaping step fires on the right requests and stays quiet on its neighbours'.

[desk-research][1.1.10] — 2026-10-14

2026-10-14T00:00:00Z

The pack's activation-eval comment now names the live agentbundle pack evals run command instead of a removed script.

[figma][0.3.4] — 2026-10-14

2026-10-14T00:00:00Z

The pack's activation-eval comment now names the live agentbundle pack evals run command instead of a removed script.

[core][3.1.4] — 2026-10-14

2026-10-14T00:00:00Z

The pack's activation-eval comment now names the live agentbundle pack evals run command instead of a removed script.

[product-engineering][1.0.4] — 2026-10-14

2026-10-14T00:00:00Z

The nine longer-route shaping skills now open with what they produce, in the words people use, and take 45% less room in the skill listing every session loads.

[converters][0.9.8] — 2026-10-14

2026-10-14T00:00:00Z

The pack's activation-eval comment now names the live agentbundle pack evals run command instead of a removed script.

[experience-design][4.1.6] — 2026-10-14

2026-10-14T00:00:00Z

The pack's activation-eval comment now names the live agentbundle pack evals run command instead of a removed script.

[contracts][0.3.8] — 2026-10-14

2026-10-14T00:00:00Z

The pack's activation-eval comment now names the live agentbundle pack evals run command instead of a removed script.

[architect][0.15.17] — 2026-10-14

2026-10-14T00:00:00Z

The pack's activation-eval comment now names the live agentbundle pack evals run command instead of a removed script.

[core][3.1.2] — 2026-10-13

2026-10-13T00:00:00Z

The offline intent view opens on four session-start questions: what to work on next, what is open under a capability, what has no parent, and what is broken.
Each item shows its name first, its state as a word plus a shape, and a plain next step that names the skill to use and why.

[core][3.1.3] — 2026-10-13

2026-10-13T00:00:00Z

Review failures recorded by a shadow-enabled run now surface as read-only knowledge candidates through a pull scan, and one invalid merged knowledge record is quarantined on enquiry and the distillation pending page instead of breaking every read.

[product-engineering][1.0.2] — 2026-10-13

2026-10-13T00:00:00Z

Discovery guidance now hands off through work-intake at G3 and ingests an outside requirements document as a new brief.

[core][3.1.1] — 2026-10-13

2026-10-13T00:00:00Z

Open briefs and specs are now named by their title in navigate-intents query results and in the offline export, so you can tell them apart without opening each file.

[core][3.1.0] — 2026-10-11

2026-10-11T00:00:00Z

See every outstanding intent, brief, and spec placed under its parent intent, without opening files one by one and without workspace.toml. The new navigate-intents skill derives the intent graph from preamble headers each time it is asked and writes nothing back.
Walk the intent tree by level, parent, children, and recorded status. One broken parent link no longer hides the rest of the tree — it shows as a flagged line naming what is wrong. An incomplete outstanding-work list is refused rather than returned silently.
close-work now checks every intent ancestor that an artifact's parent pointers record, including capability, outcome, and opportunity parents. close-work and navigate-intents now agree on every parent link.
When an intent and a brief (or spec) share a slug, both now count, so a parent can no longer look ready to close while one of them is still open.
A broken parent pointer or a fault while building the intent tree now refuses closure with a named reason: parent-edge-refused, artifact-not-in-graph, or intent-graph-unavailable.
Relate two intents with one Related intents: line. navigate-intents shows the relation from both ends, and the intent lint refuses a value that is malformed, names the intent itself, or matches no live intent of that kind.
Export the whole intent graph as one HTML file that opens offline, to share with a reviewer or browse without a terminal. It lists every intent with its parents, outstanding work, refused pointers, and parentless intents. It links each source file. It works from the keyboard. Record text is shown as plain text. The file is always written outside the repository.

[product-engineering][1.0.1] — 2026-10-11

2026-10-11T00:00:00Z

Frame an intent that sits beside another one with an optional Related intents: line. The template offers it as comment-only guidance, so it stays absent until you fill it.

[core][3.0.5] — 2026-10-11

2026-10-11T00:00:00Z

Add a conforming reviewer to any repository without changing work-loop code. The structured review boundary ships a generic reviewer port: select an obligation, submit a versioned structured report, and the boundary assesses findings and derives review readiness from it. No reviewer-role condition or reviewer-specific hook exists in the runtime modules.
Integrate the compatibility adapter that reconciles legacy review events with structured boundary facts. The legacy review authority and reviewer-specific procedure remain active; moving authority belongs to the review-authority cutover follow-on.

[core][3.0.4] — 2026-10-11

2026-10-11T00:00:00Z

workspace-status now suggests intent for a shape entry. Install product-engineering 1.0.0 or later so that suggestion resolves to a skill you have.

[product-engineering][1.0.0] — 2026-10-11

2026-10-11T00:00:00Z

Shape product work with one skill. Use intent frame to turn an idea into an outcome and the opportunity behind it. Use intent de-risk to test the assumption that would sink it. Use intent decompose to break what survives into child intents or a delivery contract.
Not sure how much shaping a piece of work needs? Ask intent which way to go. Route mode names the short, longer, or supervised route and where it starts, then offers the first step without taking it.
Every intent mode shows where you are on the intent lifecycle, offers the next step, and recommends the shaping review at the point it applies.

[agentbundle][0.52.1] — 2026-10-11

2026-10-11T00:00:00Z

The workspace-status MCP tool now dispatches a shape entry to intent. Install product-engineering 1.0.0 or later so the dispatch resolves.

[code-intelligence][0.1.9] — 2026-10-11

2026-10-11T00:00:00Z

Runs on Wicked Estate 0.29.2. A large blast radius or lineage answer now arrives in pages instead of being cut off at the first one. The agent passes each page's token (next_cursor) back with --cursor until it is null, so every row the search found is retrievable.
If the answer changes between pages, the agent starts again from the first page instead of mixing two answers.
One wicked-estate lineage call now follows dependencies 24 levels deep by default, up from 8.

[core][3.0.3] — 2026-10-10

2026-10-10T00:00:00Z

After a design you saved before building ships, close-work can now offer to reconcile it into your current architecture documentation, so a future-state design is not mistaken for a description of the running system.

[code-intelligence][0.1.8] — 2026-10-10

2026-10-10T00:00:00Z

Runs on Wicked Estate 0.26.0. See what a function depends on with wicked-estate lineage <name>; a plain name or a symbol id both work.
Impact analysis reads each dependent's own depth from blast-radius --json, so one call separates direct callers from indirect ones.
Trace where a value flows through TypeScript, TSX and JavaScript code with wicked-estate lineage <name> --relation flows_to.
wicked-estate stats --json tells you which commit the graph was built from and how many commits behind it is.
An exit 1 with nothing on stdout is treated as a command error, not as "nothing found".

[architect][0.15.16] — 2026-10-10

2026-10-10T00:00:00Z

When a repository has no reference architecture, architect-design now says so, lowers its confidence in stack assumptions, and offers to hand off to adapt-to-project (existing code) or init-project (a new project) to establish one. It never drafts one itself.
architect-design now treats a saved design as future-state and points to reconciling it into current architecture once the change ships.

[product-engineering][0.13.25] — 2026-10-10

2026-10-10T00:00:00Z

With the architect pack installed, shaping now reaches architecture at the right moment. frame-intent offers architect-design when framing hits a system-shape question. frame-domain and de-risk-intent offer architect-assess for a current-state model. decompose-intent and map-capabilities offer a design before cutting work. Without architect, no architect offer, mention, note, or error appears.
When a saved design covers a feature, decompose-intent now carries it into the delivery contract's design context, so the spec stage reads it.
explore-options, diverge-solutions, de-risk-intent, and decompose-intent now use a current-architecture document, when one exists, to judge feasibility and check slices against subsystem boundaries.

[core][3.0.2] — 2026-10-10

2026-10-10T00:00:00Z

When work-loop plans a rename, removal, or refactor, checks that a fix reaches live code, or when bug-fix traces a bad value through its callers, it can now hand that question to repository-exploration. That skill uses a code-navigation tool your agent already has when one fits, and otherwise searches the repository.
repository-exploration now activates when a workflow or a user needs evidence to settle a pending decision, and leaves planning, building, and fixing to the workflow that asked.

[code-intelligence][0.1.7] — 2026-10-10

2026-10-10T00:00:00Z

Install the code-intelligence pack on its own, at repo or user scope. It no longer requires the core pack.
The skill and both agents now carry their own evidence rules. Provider output is treated as data, and a file location the code graph returns is never opened; call sites are confirmed with the agent's own search.
Answers end with an evidence note: the question, the commands and sources used or passed over, and why.

[product-documentation][0.1.2] — 2026-10-10

2026-10-10T00:00:00Z

author-product-docs documents any repository — libraries, CLIs, APIs, apps, services, frameworks, plugins, and agent packs — by first discovering the product surface and checking the docs against its own sources.
Documentation audits start with a journey gap report that covers nine reader stages, from discovery to contributing.
It now writes release notes, changelogs, migration guides, contributing guides, troubleshooting pages, quickstarts, installation pages, and docs landing pages.

[core][3.0.1] — 2026-10-09

2026-10-09T00:00:00Z

A work-loop transition that your arguments cannot satisfy now stops without touching state.json. Correct the command and run it again; previously the refusal left a marker behind that no later command could clear, and the run could only be rescued by hand-editing state.

[core][3.0.0] — 2026-10-08

2026-10-08T00:00:00Z

Send all intake requests to work-intake. The capture-work skill is removed and has no replacement alias.
A former spec/<slug> or other legacy workspace entry surfaces as an unsupported_legacy finding and cannot dispatch. Rewrite it in canonical form by hand and delete the old alias in the same edit — a canonical entry is refused dispatch while its old alias survives anywhere in the workspace.
Use workspace-status repair-plan --migration-selection, repair-apply, and repair-rollback only to recover or roll back an operation already recorded in the migration ledger.

[agentbundle][0.52.0] — 2026-10-08

2026-10-08T00:00:00Z

The workspace-status MCP tool now surfaces a former legacy entry (a bare spec/<slug> string or other accepted legacy shape) as an unsupported_legacy finding. The entry is absent from ready and blocked; ordinary reconciliation no longer accepts legacy shapes as lifecycle members.
A canonical entry with a surviving historical alias appears in blocked with a duplicate_membership finding and does not appear in ready. Delete the alias and the canonical entry dispatches normally.

[product-engineering][0.13.23] — 2026-10-08

2026-10-08T00:00:00Z

Skills and examples in the product-engineering pack that previously directed users to capture-work now direct them to work-intake, matching the Core 3.0.0 surface.
frame-situation and diverge-solutions note that old shaping entries still appear in the information-only shaping lists until rewritten in canonical form.

[core][2.30.1] — 2026-10-08

2026-10-08T00:00:00Z

Close work whose descendants have terminal statuses with dates or notes, such as Shipped (2026-09-11) or Archived — note. The closure check now reads the leading status word and preserves the full value in its evidence.

[code-intelligence][0.1.5] — 2026-10-08

2026-10-08T00:00:00Z

Ask what a function depends on, all the way down, and get an answer from the wicked-estate command line. You no longer need the optional MCP server for forward dependencies, graph walks, or rules lookups.
Ask which of a change's dependents matter most, and get an importance ordering from the command line, with any dependent it could not rank named rather than dropped.

[core][2.30.0] — 2026-10-07

2026-10-07T00:00:00Z

Installed Core now answers "which specs and briefs deliver this feature intent?" one way. A typed resolver ships inside the close-work and work-loop skills, so every Core install carries it and it runs without agentbundle importable. Both close-work and the traceability lint use its answer. A missing, contradictory, or unsafe mapping is reported by a stable code instead of being guessed.
close-work runs the resolver whenever it closes a spec or checks a feature's spec or brief delivery. The traceability lint runs it only when the repository has a delivery anchor (briefs, a rollup, or populated discovery layers), and fails with delivery-resolver-unavailable when the spec or intent folder has moved.

[code-intelligence][0.1.4] — 2026-10-07

2026-10-07T00:00:00Z

The pack now includes a worked example showing how to answer a change-impact question with a code index and how to answer the same question without one, with the limits of each approach named rather than rounded away.
The example marks which steps come from the core pack's shared rules for answering code questions and which are specific to Wicked Estate, so another code-intelligence tool does not need to copy Wicked Estate's commands.

[core][2.29.0] — 2026-10-06

2026-10-06T00:00:00Z

Core now has a repository-exploration skill for open questions about how your code behaves, what it depends on, and what a change would affect. It uses a code-intelligence tool your agent already has only when that tool fits the question, keeps the tool's limits, checks key claims against the source, and stops once the question is answered. With no tool installed it still answers from the repository.
Tool output stays data. A tool's description, its results, or a file it points to cannot add a folder the agent may read, start a refresh or install, or run a change on its own. A refused file stays unread.

[core][2.28.0] — 2026-10-05

2026-10-05T00:00:00Z

Core now has a repository-grounding skill that answers "what already governs these paths?" from the repository alone, and can add evidence from a code-intelligence tool your agent already has. Provider evidence is labelled, its limits are kept, a claim that conflicts with source never decides the answer, and with no tool installed the answer is the same.
A path or file: URI a provider returns is read only if the file sits inside the repository or a folder that you or the calling workflow approved. Symlinks, hard links, non-regular and oversized files are refused with a stated reason, and the locator text never reaches a command line raw, so shell metacharacters in it cannot run.

[governance-extras][1.0.0] — 2026-10-04

2026-10-04T00:00:00Z

navigate-decisions replaces the retired rfc-status skill. It queries both ADRs and RFCs for exact lifecycle, checked supersession lineage, contextual references, guidance context, and record detail. It also exports a self-contained offline HTML explorer with list, lifecycle-graph, guidance-context, and record-detail views.

[frontend-engineering][0.4.4] — 2026-10-04

2026-10-04T00:00:00Z

A token taxonomy that says nothing about a visual domain the surface needs — no value, and no unresolved record — is now an upstream gap. The build holds that axis and routes it to whoever produced the taxonomy, instead of inventing a type scale or stroke widths that read as resolved.

[core][2.27.14] — 2026-10-03

2026-10-03T00:00:00Z

Installed Core can now rename an intent as a recoverable retire-and-issue operation. The old path becomes a tombstone, the successor receives a fresh target-token ordinal, citations and workspace.toml move with it, and an interrupted run can be driven forward or back.

[agentbundle][0.51.0] — 2026-10-03

2026-10-03T00:00:00Z

Protected catalogue installation without manual token injection. AgentBundle now picks credentials automatically from a priority-ordered set: bearer token, JFrog CLI 2.105.0+ profile, exact-machine .netrc record, and anonymous. Public catalogues need no setup. A broken configured provider stops resolution; no fallback across types. Set AGENTBUNDLE_CA_BUNDLE for the direct path; use SSL_CERT_FILE or SSL_CERT_DIR to reach jf api on Linux.

[code-intelligence][0.1.3] — 2026-10-03

2026-10-03T00:00:00Z

The agent can now ask for the route from one symbol to another using wicked-estate path. A result of found: false with depth_bounded: true means the route was not proven absent — the search hit its depth limit — and you can raise --max-depth (up to 16) to look deeper.
A blast radius now tells you when the depth limit cut it short. depth_horizon_reached: true in the JSON output means the result is a floor, not the full set; --depth N (maximum 24) lets you search deeper, and --depth 1 isolates the direct dependents from the transitive ones.

[frontend-engineering][0.4.1] — 2026-10-02

2026-10-02T00:00:00Z

The approved-visual-target rung now resolves only when the direction artifact carries visual_target: confirmed. An approved target whose confirmation was never recorded no longer satisfies the rung; the distinction is in the artifact, not in a reviewer's memory.

[experience-design][4.1.3] — 2026-10-02

2026-10-02T00:00:00Z

A visual target now binds composition only once a person confirms it and converge records visual_target: confirmed in the direction artifact. An unconfirmed target is still captured — converge records it as visual_target: unconfirmed — it just does not bind, and the rung that reads the field is downstream in frontend-engineering.

[core][2.27.11] — 2026-10-01

2026-10-01T00:00:00Z

A work-loop run whose state.json records no retry cap now shows the default cap of 5 on its event line, instead of no cap at all. The run was always held to 5, so a dashboard or alert watching how close it is to being refused was reading nothing where it should have read 5. A malformed cap, which the guard refuses, still reports null, and so does a retry count nothing has recorded; the key is always present, which is how a consumer tells "not recorded" from zero.

[core][2.27.9] — 2026-10-01

2026-10-01T00:00:00Z

Work planning now records only declined architectural additions that materially affect scope or design, and sizes review units by whether they can be independently understood, verified, or reviewed.

[core][2.27.7] — 2026-10-01

2026-10-01T00:00:00Z

Renaming a live intent whose title mentions tombstones now works. The rename validator reads a document title as a title rather than as a damaged retirement marker, so an intent called something like "Tombstone migration plan" is no longer refused. Nothing that should be refused now passes.

[core][2.27.6] — 2026-09-30

2026-09-30T00:00:00Z

Work-loop can now explain a scheduled wave before dispatch without authorizing concurrent writes. loop-cohort wave-decision --json reports unfinished tasks as parallel-capable or sequential, includes pairwise Touches: relations, and returns fixed public-safe JSON refusal envelopes while leaving ADR-0005's post-write gate and ADR-0061 D5 unchanged.

[governance-extras][0.11.3] — 2026-09-30

2026-09-30T00:00:00Z

The governance-index template and its how-to guide no longer promise that generate-iac will fill in missing domain rows for you. It never did: it offers to create an index when you have none, and it reads the one you already have. Both places now say so, and the guide tells you to add the row yourself.

[architect][0.15.15] — 2026-09-30

2026-09-30T00:00:00Z

Well-architected reviews now route data/ML, GenAI/agentic, and serverless workloads only through their supported reference material.

[experience-design][4.1.2] — 2026-09-30

2026-09-30T00:00:00Z

You can now record whether a direction's visual target was ever confirmed, and tell that apart from never having asked. A visual_target key on the direction artifact holds one of none, unconfirmed, or confirmed, and converge writes it every time it runs, so the answer is there without you remembering to add it. Nothing reads the field and nothing is gated on it: it is a record you can act on when you want to, and ignore when you do not.

[core][2.27.5] — 2026-09-29

2026-09-29T00:00:00Z

Work-loop can now prove a current branch is fresh before it tries to write Git metadata. If the branch is stale, it still stops: environments that can prepare the update ask before updating, and locked-down environments tell you to refresh the branch separately.

[core][2.27.4] — 2026-09-29

2026-09-29T00:00:00Z

Renaming an intent no longer risks silently handing its old number to a different intent later. A retired filename now leaves a tombstone behind, and the number allocator counts tombstones alongside live intents, so a retired number stays out of circulation for as long as its tombstone stands.

[experience-design][4.1.1] — 2026-09-29

2026-09-29T00:00:00Z

Adding a component or section to a surface that already has an aesthetic direction no longer restarts the vibe conversation. The inherit route now scopes the new element against the direction's existing goals and axis commitments, which is what its route table always promised.

[code-intelligence][0.1.0] — 2026-09-29

2026-09-29T00:00:00Z

Ask what is actually true about a codebase before changing it. Index the repository once, then ask what calls a symbol, what a change would break, or how the system is organised — answered from a resolved call graph rather than a text search.
Answers carry the index's own limits. The unresolved-reference count, the truncated-row count, and the traversal horizon are reported rather than rounded away, so a blast radius reads as a floor instead of a total.

[frontend-engineering][0.4.0] — 2026-09-29

2026-09-29T00:00:00Z

Frontend builds now stop on upstream design gaps instead of filling them from a fallback, local premise, or category habit. When a resolved direction is missing the needed taxonomy, or a taxonomy records a domain as unresolved, the run holds only that axis and routes it back to the recorded owner or operation.
Standalone visual work now has a clear admission rule: it is allowed only when there is no applicable design artifact, no incumbent system, and no upstream authority left to complete. That keeps greenfield work moving without turning unfinished design decisions into implementation choices.

[experience-design][4.1.0] — 2026-09-29

2026-09-29T00:00:00Z

Design artifacts now tell downstream builders who owns every unresolved domain and which operation supplies it, so “unresolved” routes work back upstream instead of inviting a consumer to invent a value.
The design thread now makes design-system conditional, not merely optional. Run it when a direction exists and neither a completed design system nor a coherent incumbent system supplies every concrete value the surface needs; skip it when an incumbent system already covers the work.

[product-engineering][0.13.22] — 2026-09-29

2026-09-29T00:00:00Z

frame-intent now refuses a configured destination before writing when resolving it follows a symlink out of its anchoring root. An ordinary absolute destination outside the repository still asks first, and a personal-scope destination remains legitimate.

[core][2.27.3] — 2026-09-28

2026-09-28T00:00:00Z

`loop-cohort status` now shows what each wave's dispatch records actually say. A wave whose every task was declined used to read exactly like one whose every task was implemented; both reported the same single flag. The new per-wave counts separate receipts, declines, superseded records and unaccounted tasks, so someone reading a finished run can tell which waves were implemented and which were not.

[frontend-engineering][0.3.5] — 2026-09-28

2026-09-28T00:00:00Z

The guides now describe the pre-flight the pack actually ships. The design-handoff how-to states the four-rung precedence in order and says what happens when a slot is empty; the tutorial walks the rung an adopter without a design tree actually uses, working from a stated premise rather than copying a named product.
Where your token taxonomy already resolved a value, the build now uses that value as given instead of re-deriving its own. Re-deriving a number the design step already decided is how a surface drifts from the direction it was supposed to inherit; what the taxonomy left as a relationship, or recorded as unresolved, is still the build's work to settle.
You can now see what the frontend reviewer actually checks. Its reference page documents all seven lenses, including the visual-authority lens, and says what evidence confirms each one — the diff for five, a rendered capture for one, the evidence manifest for the last.

[experience-design][4.0.3] — 2026-09-28

2026-09-28T00:00:00Z

`design-system` now decides your product's visual system instead of handing you an empty table of token names. It resolves typography, color, spacing, shape, depth and motion. Each decision records the authority behind it, so you can tell an inherited value from a derived one at a glance.
A system you already have is inherited, not replaced. The skill finds the source of visual truth your interface actually reads from and judges whether it is coherent. It fills only the gaps your current work needs, keeps your naming conventions, and will not build a second system beside your first.
Nothing in the pack carries a palette, typeface, scale, radius, shadow or easing value. Your product's values come from your direction and your existing system, so two unrelated products no longer end up looking the same.

[core][2.27.2] — 2026-09-27

2026-09-27T00:00:00Z

Ask the agent to explain a diff, branch, commit, or pull request and you get back one self-contained offline HTML lesson — background, intuition, a code walkthrough, and a five-question quiz — written for that specific change rather than poured into a shared template. The agent reports the exact file path it wrote.

[core][2.27.1] — 2026-09-27

2026-09-27T00:00:00Z

Interrupted work-loop transitions now replay without separate cohort repair commands. The five registered cohort-effect transitions apply their effects through loop-cohort.py, record one durable transition identity, and safely resume without double-advancing a wave, double-counting a retry, or duplicating a review round.

[frontend-engineering][0.3.4] — 2026-09-27

2026-09-27T00:00:00Z

When your team has already agreed a visual direction, the frontend build now inherits it instead of quietly picking a different look. The skill resolves visual authority from the design artifacts your repository already points at, falls back to the visual system you already have, and only invents a premise when a surface is genuinely greenfield. It no longer names a product as an aesthetic anchor.
Significant visual work is now looked at while it is being built. The agent renders a representative state, compares it against the direction it inherited, corrects once where the difference is material, and renders again to check — then goes on to the usual gates. The loop is bounded: a further pass happens only if you ask.

[experience-design][4.0.2] — 2026-09-27

2026-09-27T00:00:00Z

creative-direction now turns product context into a buildable visual contract, not just a mood statement: it records the engagement mode, product-specific visual thesis, first-viewport thesis, honest evidence and asset status, and optional visual-target boundaries without requiring Product Engineering, Frontend Engineering, a comp, browser control, or image analysis.

[experience-design][4.0.1] — 2026-09-27

2026-09-27T00:00:00Z

The first approval now happens where you can actually make it. The journey approval covers the journey *and* the screens derived from it, so it comes after the screen flow exists rather than before. The check that earns it — does every screen trace to a moment in the journey — needs both halves in front of you.
The adopter guide opens with the work, not the roster. The experience-design guide front door now answers "I have an experience to design — how do I get from the user's outcome to reviewed design intent?" It leads with one ordinary-language request, the stage sequence, the three decisions that are yours, what you end up holding, and which downstream pack consumes each artifact. The skill reference is still there, one level down.

[product-engineering][0.13.21] — 2026-09-27

2026-09-27T00:00:00Z

Feature and capability intents can now carry product facts for downstream experience work only when a surface is materially affected. The optional handoff names the affected journey or surface, first-success behavior, product proof, claim evidence, constraints, prohibited claims, and unknowns while backend-only work and product-level framing stay unchanged.

[core][2.27.0] — 2026-09-27

2026-09-27T00:00:00Z

Ask the agent to explain a diff, branch, commit, or pull request and you get back one self-contained offline HTML lesson — background, intuition, a code walkthrough, and a five-question quiz — written for that specific change rather than poured into a shared template. The agent reports the exact file path it wrote.

[experience-design][4.0.0] — 2026-09-27

2026-09-27T00:00:00Z

The copy layer is now one skill with three modes. content-design covers message and narrative structure, per-surface acquisition copy goals, and the brand-level register. You pick a task, not a registration — the skill selects the mode from your request. The copy-direction and tone-of-voice skills are removed.
Nothing you have already written moves. All three artifacts keep their paths and their type: values: content/<slug>.md, copy/<surface-slug>.md, and the reserved copy/brand-register.md, which still carries scope: brand-level beside its type:. An existing document needs no migration.
Shared copy references were reconciled rather than merged. Six references that shared a name across the folded skills disagreed; each was resolved deliberately, and the two whose variants differed only by scope are now one body with a named scope parameter that binds per mode.

[core][2.26.46] — 2026-09-26

2026-09-26T00:00:00Z

Base freshness can now be unverified without stopping every work-loop. When network access, authentication, timeouts, or enterprise Git-metadata policy prevent the check from completing, the loop tells you freshness was not verified and continues so you can update the branch separately. A proven stale base or unsafe repository state still stops the run.

[core][2.26.45] — 2026-09-26

2026-09-26T00:00:00Z

Intent preambles can now say an outcome is shared with another intent by adding Outcome co-owner: as a typed peer pointer. The corpus and traceability checks reject malformed, missing-target, and self-referential peers without turning the peer into a parent or delivery edge.
Decomposed: can now record a finished childless decomposition with YYYY-MM-DD closed-empty. It stays distinct from the literal no, so an author can tell "we deliberately closed this with no child work" from "we decided not to decompose."

[product-engineering][0.13.19] — 2026-09-26

2026-09-26T00:00:00Z

The intent template and product-engineering how-tos now show both closure declarations authors can copy: optional Outcome co-owner: for a shared outcome, and closed-empty for a decomposition that is complete with no child work.

[core][2.26.44] — 2026-09-26

2026-09-26T00:00:00Z

A `Shipped` brief now requires a `Cut-closed:` declaration. When you close a brief as Shipped, add a Cut-closed: preamble field recording that no further slices are coming: an ISO 8601 date then evidence, as in 2026-08-25 All nine slices are Shipped. It must sit above the brief's first ## heading; written below one it reads as absent. The coverage lint also refuses a Draft brief that carries the field, and every other state may carry it or omit it.
A brief that passed the coverage lint yesterday can fail today, so this is a gate change to plan for rather than a new option to adopt. Any repository holding a Status: Shipped brief without the new field fails until it is backfilled — run it from your repository root as python3 <skill>/scripts/lint-brief-coverage.py --root . before your pipeline does — it resolves docs/product/briefs beneath the root you pass, so pointing it at the briefs directory itself finds nothing and exits 0.
A brief's states and legal moves are now documented once, so the reference guide and the lint cannot disagree about which transitions are legal or what each status requires of its children.

[experience-design][3.0.0] — 2026-09-26

2026-09-26T00:00:00Z

information-architecture now owns the surface-genre routing that previously required six separate registrations: analytical-design, conversion-design, documentation-design, informational-design, marketplace-design, and workspace-design.
The removed registrations' methods now live as genre references under information-architecture, so adopters should update direct skill references and invocations to call information-architecture with the relevant surface-genre: value.

[frontend-engineering][0.3.3] — 2026-09-26

2026-09-26T00:00:00Z

Frontend pre-flight now routes every surface-genre: value through the consolidated information-architecture handoff when experience-design is co-installed, with interaction-design still owning transactional journeys.
The design-system handoff now names the shipped design-system skill slug, and experience-design >=3.0.0 is the recommended floor for the genre-aware route.

[atlassian][0.10.1] — 2026-09-26

2026-09-26T00:00:00Z

You can now see what each Jira Epic delivered beside what that work was meant to change, in one read-only view. Ask for a project's Epics and their outcomes: the view groups the project's work by Epic, counts delivery from the flow-metrics skill's own per-issue rows, and renders each Epic's recorded outcome next to it. It writes nothing to Jira and leaves no file behind.
An Epic with no recorded outcome says so, and asks you for one. Instead of a blank row, the view states plainly that nothing is recorded and names the one place to write it — the block under an Outcome heading in the Epic's description. Hand your answer back with --outcome EPIC-KEY=<text> and your exact words come back as text you paste into Jira yourself, so the outcome stays something your team said.
A `flow-metrics` run can now touch the on-disk cache not at all. The new --inert-cache mode performs no cache read, write, directory creation or stale-temporary-file cleanup, which --no-cache still did. It is off by default, so every existing flow-metrics run behaves exactly as before.

[core][2.26.43] — 2026-09-25

2026-09-25T00:00:00Z

workspace-status can now tell you which delivery contracts are retirement candidates. It reads the repository, reports every blocker holding each candidate back, and names the migration obligations a candidate must discharge before its container can go. Being listed authorizes nothing: a human selects, and the separately confirmed prune executes.
Evidence the run could not read withholds eligibility rather than passing silently. Each blocker declares the input corpus its absence rests on, and a refusal naming any member of that corpus suppresses every candidate that blocker is evaluated over — including candidates carrying no blocker of their own, which are reported rather than dropped. A shorter list never means a cleaner one.
Ages come from recorded change history and are labelled as such. They are not the thirty-day cooling clock, which runs from a delivery-completion event this capability does not read, so the report is a triage signal rather than a disposition.

[experience-design][2.0.10] — 2026-09-25

2026-09-25T00:00:00Z

The creative-direction skill now exposes five named operations — frame, explore, visualize, converge, and refine — each with a dedicated reference, so craft detail loads only when the operation needs it.
A route rule (inherit, extend, originate) selects how much invention the work warrants; inherit runs no divergence and no visual step when a direction already exists for the surface.
Divergence generation in explore draws referents from the audience's own world — publications, instruments, maps, architecture, packaging, signage — rather than from other software products, so candidate directions stop converging on the category default.
The visual step in visualize is capability-gated: it defaults to a text schematic in any harness that cannot produce images, with a named skip rather than a blocker, and produces a rendered comp only on the originate route when the harness supports it.
Asking to refine an existing direction now reaches the skill and amends the document in place, instead of starting a second direction for the same surface.

[core][2.26.42] — 2026-09-24

2026-09-24T00:00:00Z

Three re-entry edges now require a wave reopen before you re-enter implementation. When gates fail, a code review finds remaining issues, or a blocker is applied, the loop now refuses the edge and names the tasks that still hold a live dispatch record. Run loop-cohort wave reopen <spec-dir> --expect-run-id <id> to mark the current wave's records superseded and unblock the edge.

[core][2.26.41] — 2026-09-24

2026-09-24T00:00:00Z

An intent that says the work is finished now has to say who decided that and on what evidence. Status: Fulfilled requires both an Accepted: and a Fulfilled: record, and Status: Cancelled requires Accepted:. Intents that passed the corpus lint yesterday can fail today, so this is a gate change to plan for rather than a new option to adopt.

[core][2.26.40] — 2026-09-24

2026-09-24T00:00:00Z

A review finding is now blockable when something outside it establishes the defect — a test, a lint, a schema, a resolvable reference or a stated constraint that disagrees with the target — even when several repairs would fix it. The old wording asked instead whether the fix was fully determined, so a finding like "this count says three where the tree holds five" could be capped at Concern purely because you could change the count, add a baseline, or drop the claim. How much that mattered depended on which model read the rule. Two models were asked which of the same 50 recorded blocking findings should have been capped: under the old wording they answered 2 and 8, and under the new one both answered 1.

[governance-extras][0.11.2] — 2026-09-24

2026-09-24T00:00:00Z

`new-adr`'s infra mode no longer claims its seven topics cover the whole governance index. Each of the seven maps to a domain, and observability is named as the domain with no authoring topic of its own — so you can see the gap instead of assuming the list is complete.

[governance-extras][0.11.1] — 2026-09-24

2026-09-24T00:00:00Z

Superseding part of a decision record is now explained where you write it. The four supersession fields shipped with the ability to carve one numbered constraint out of an older ADR, and the shape check has always enforced it — but the template's one-line examples never said whose D-IDs each half names, so the only way to get it right was to find an existing pair and copy it. The template now states the rule beside the fields: both halves name D-IDs defined by the *older* record, and a partially superseded record keeps Status: Accepted.

[product-engineering][0.13.18] — 2026-09-24

2026-09-24T00:00:00Z

The `align-value-stream` layout page now names every skill that reads its output setting. It lists all eight [product] readers, calls out lean-canvas's scan-only read, and makes clear that ux-writing reads [design] instead.

[core][2.26.39] — 2026-09-23

2026-09-23T00:00:00Z

Saying that an intent was replaced now takes two plain fields instead of one packed line: the status reads Superseded, and a Superseded by: field names the bet that took over. Each is refused without the other, so a pointer can no longer be left behind when a status moves on.
The Accepted: and Fulfilled: lines now have to carry more than a date. Each takes a date, a space, and text — a date on its own is refused, because these lines exist to hold the reasoning. What that text should say, who decided and on what evidence, is for you; the check asks only that it is there.

[frontend-engineering][0.3.2] — 2026-09-23

2026-09-23T00:00:00Z

A surface now declares how much contract it owes, and `explore` is the default. The journey opens by asking you to pick the risk tier, then tells you the contract fields, the states, the gates and the captures that tier carries. A tooltip and a checkout flow walk the same stages at different depths, so the cheap path is a stated choice rather than a corner cut.
The cheap tier never drops an accessible surface. Every state whose absence fails WCAG 2.2 AA is either already in the explore set or binds conditionally wherever its trigger fires. Depth buys evidence and completeness; it does not buy a surface someone cannot use.

[experience-design][2.0.9] — 2026-09-23

2026-09-23T00:00:00Z

The say-this table now tells you which steps you can skip. Every row carries one optionality verdict, resolved against the how-to guide that owns that skill, so the two surfaces can no longer disagree about whether a step is optional.
A named minimal viable thread replaces guessing at what to drop. Four steps make the output coherent; the journey says so, and says what a skipped step actually costs rather than warning against shortcuts in the abstract.

[product-engineering][0.13.17] — 2026-09-23

2026-09-23T00:00:00Z

The contract this pack carries now defines what each depth owes. Its copy of the experience contract gained the shared state-coverage map and the frontend-engineering owner label, so a team reading the contract alongside an intent finds one definition of a risk tier rather than a bare word. The pack's skills are unchanged and still do not read the contract themselves.

[product-strategy][0.2.7] — 2026-09-23

2026-09-23T00:00:00Z

The contract this pack carries now defines what each depth owes. Its copy of the experience contract gained the shared state-coverage map and the frontend-engineering owner label, so the tier annotations a strategist reads there mean one thing across all four packs. The pack's skills are unchanged and still do not read the contract themselves.

[agentbundle][0.49.0] — 2026-09-23

2026-09-23T00:00:00Z

`agentbundle catalogue sync` can now bring a derived catalogue up to date, not just check it. Run it with neither --dry-run nor --check and it asks once for confirmation, then applies the plan: unedited files update, your edits keep their bytes with the source's version saved beside them, and paths the source dropped are removed — guarded so a run never touches anything outside what the catalogue already recorded. Four new flags — --pack, --profile, --guides, --package — let you sync just part of the catalogue, and --yes skips the prompt for a script or CI job.

[core][2.26.36] — 2026-09-22

2026-09-22T00:00:00Z

A reference from one artifact to another now names exactly one target, and an ambiguous name is refused rather than guessed. A pointer written as a bare name used to be resolved by picking the alphabetically first match when several fitted. Nothing told the author a choice had been made. Such a name is now reported with every candidate it matched, so the author decides rather than the sort order.

[agentbundle][0.48.0] — 2026-09-22

2026-09-22T00:00:00Z

You can now check a derived catalogue against its source without touching a file. agentbundle catalogue sync --dry-run prints what would update, what would keep your edits behind a companion file, and what the source no longer has. agentbundle catalogue sync --check answers whether your catalogue still matches its source. Neither writes anything.

[core][2.26.29] — 2026-09-22

2026-09-22T00:00:00Z

An intent now states who owns it, what altitude it sits at, and where it stands — and a check refuses one that does not. Reading an intent's opening lines used to leave the important questions open: two of the fields a reader needs most were optional, and one was buried in the body. Anyone can now learn who is accountable, how high the bet sits, whether it is still live, and whether it has been tested and broken down, without reading further.
A stopped bet can say so. An intent that was withdrawn or cancelled now records that as its own standing, rather than being indistinguishable from one still waiting. A rollup that counted a cancelled bet as delivered is the outcome this prevents.

[experience-design][2.0.8] — 2026-09-22

2026-09-22T00:00:00Z

A brief produced from the shipped template now carries the mode the next skill reads. copy-direction takes communication_mode from the brief to decide whether its anti-AI-smell scan applies to the copy goals it names. The template had no slot for the field, so an author following it produced a brief that skill could not route, and the eval harness did not catch it.

[product-engineering][0.13.16] — 2026-09-22

2026-09-22T00:00:00Z

The intent template now seeds every field an intent is required to carry, and three that record how far its shaping got. An author filling the template in produces something the checks accept, instead of discovering the gaps afterwards.

[desk-research][1.1.9] — 2026-09-22

2026-09-22T00:00:00Z

arXiv searches now find the paper you asked for. Free text used to be pasted straight into arXiv's query language, where a phrase quietly became a loose word match: asking about instruction adherence in agent config files returned 124,103 papers with an unrelated one first. The retriever now composes the query properly and widens it only as far as it must, returning 97 matches with the right paper at the top — and it tells you which attempt produced them, so you can see when a match was loose.
You can ask arXiv precise questions. Search by title, author, abstract or category, combine them, sort by newest or by relevance, and restrict to a submission-date window. If you already know arXiv's own query syntax, pass it through untouched.
Fetch one paper exactly, instead of searching for it. Give the retriever an arXiv identifier or a link and it returns that record — including papers from before 2015 and the older slashed identifiers — with its abstract, authors, categories, DOI and journal reference where arXiv publishes them. Ask for the full text and you get named sections within a stated budget rather than a whole paper.
A paper's submission date and its revision date are now separate. The previous version showed only one, so a 2017 paper revised in 2023 could be read as a 2023 paper.
Rate limits no longer end a research session. Requests are spaced and retried, so the retriever keeps working where it previously became unavailable mid-investigation.

[core][2.26.28] — 2026-09-21

2026-09-21T00:00:00Z

A knowledge topic can now be recorded as a durable lesson that owns no file. Some lessons are about how a tool behaves, not about a defect living in particular bytes, and pinning them to a file made them go stale on every unrelated edit to it. Such a topic can now be written with no owning source at all, which is a state the store could already hold and read but its writer could not produce.

[agentbundle][0.47.3] — 2026-09-21

2026-09-21T00:00:00Z

A commit made for you by a workspace session now stages only that item's own files. If the directory you configure for a pack's output contains *, ?, [, { or }, it cannot describe a commit scope — a * in particular made the commit reach every uncommitted file under the directory above it. Such a value is now refused with an error naming the section to correct, and the rest of the session's git tools keep working. Rename the directory to one without those characters to commit again.

[frontend-engineering][0.3.1] — 2026-09-21

2026-09-21T00:00:00Z

Verify mode's gate list now names the rendered-page inspection. The inspection gate has shipped for some time, and both the visual QA checklist and the evidence manifest already required it. The verify procedure now enumerates all five gates and counts five.

[core][2.26.27] — 2026-09-21

2026-09-21T00:00:00Z

A small defect noticed mid-loop now gets routed while both of its routes are still open. Deciding where it goes happens at the DECIDE pass, which the loop reaches at a definite point, so a fix that could have shipped in this change no longer reaches the close with nowhere left to go.
You can now read one row of the close-time table and know where your defect goes. A ready-now fix that can ride along and one that needs its own review unit sit on separate rows, so telling them apart no longer means reading past the table into the paragraph below it.

[core][2.26.26] — 2026-09-21

2026-09-21T00:00:00Z

A work-loop close can now record a specific, real piece of leftover work that is genuinely blocked, instead of losing it. When a defect, an open question, or a decision is blocked on someone else's call, an instrument you don't have, elapsed time, or a dependency, the close captures it as a work-item — with what "done" looks like and why it couldn't be finished now — rather than dropping it or writing a vague note that a later reader can't act on.
A work item is checked before it's written, not after. Each one runs through a cold review with no memory of the session that noticed it; an item that isn't real, isn't necessary, or is missing what its shape requires is refused with a specific reason instead of being written silently.
A blocked defect may carry its own reproduction command, bounded to a small, read-only set of tools (cat, wc, grep, ls) with no options and no path traversal, so a captured item can point at exactly how to confirm it without granting the record any power to write or execute.

[agentbundle][0.47.2] — 2026-09-21

2026-09-21T00:00:00Z

What the status surface reports about your output paths now follows the same rules as everything else it reports. If the directory you configure for a pack's output uses anything beyond letters, digits and . _ - /, or points at the repository root itself, the status surface reports no output pattern and tells you which section to correct, rather than passing the value through. Renaming the directory to a plain, repository-relative path restores the reported path.

[core][2.26.25] — 2026-09-21

2026-09-21T00:00:00Z

Closeout asks five questions instead of seventeen. work-loop's finish checklist now reads as five assertions — acceptance criteria satisfied, verification passed, review passed with residuals surfaced, artifacts consistent, repository state valid — with the existing obligations grouped underneath each one. Nothing that blocked a release before stops blocking it; the list is easier to hold in mind and easier to answer honestly.
Knowledge capture leaves the critical path. Routing a session's scratch notes is no longer a condition of declaring done. ## Capture keeps its routing rules and now runs after the loop closes — at session end, an explicit handoff, or before compaction. A capture obligation enforced at the completion boundary turns "is this lesson worth keeping?" into a blocking judgment, and the cheapest way past a blocking judgment is a low-value note. Recorded against RFC-0077 § Errata.

[agentbundle][0.47.1] — 2026-09-21

2026-09-21T00:00:00Z

The status surface and the commit scope agree on where work goes. workspace_status reported an item type's built-in default output path even when agentbundle-layout.toml configured a different output_dir, while the git tools already scoped commits to the configured one. Anyone reading the status surface saw one location and got another. Nothing was written to the wrong place — the label was wrong, not the write.

[core][2.26.24] — 2026-09-21

2026-09-21T00:00:00Z

A path that hands your work to another skill is fixed, not configurable, and the reference now says which paths those are. workspace-status' layout page documented a shaping key nothing reads and left out output_dir, the one every consumer does read. It now describes the real key, explains that the shaping folder is composed under it rather than configured beside it, and states plainly that your briefs and intents folders are hand-offs — redirect them and the artifact stops being found.
A new intent now gets a readable name, not just a slug. An intent admitted through work-intake at a recognized altitude lands as CAP-0004-external-tracker-projection.md rather than external-tracker-projection.md, so you can say which one you mean out loud. The number is per altitude, so a capability and a feature can both be 0001.
Admission never stops because a number was unavailable. If the altitude has no prefix, or the allocator cannot be sure of its answer, the intent is still written and registered under its plain name and records why — so nothing you were doing is lost, and the number can be added later. The slug is the identity; the number is a label for humans.
Nothing is renamed. An intent already on disk keeps its filename whatever its altitude, so existing links and registry entries keep resolving.
The layout reference answers your question instead of citing ours. Where it used to say that two of our internal proposals disagree about whether a skill offers the default path before asking you, it now tells you what to do: do not rely on either order, and set output_dir explicitly if the distinction matters.

[product-engineering][0.13.15] — 2026-09-21

2026-09-21T00:00:00Z

`frame-intent`'s layout page now tells you which intents it places and which it does not. An intent you author for yourself, including in a personal vault, goes where you configured. A repository intent goes to the folder core's admission reads, and that path is fixed — the same treatment briefs already had, now written down for intents too. Authoring privately and admitting to the repository are two steps of one flow, not two rival destinations.
The page also names its real readers. It claimed two skills read [product]; eight do, and one reads a different pack's section entirely.

[core][2.26.23] — 2026-09-20

2026-09-20T00:00:00Z

A review pass runs when something names a reason for it. work-loop selects reviewers from one table instead of several scattered triggers: adversarial-reviewer always, and security, user-facing, architecture and quality lenses only as the change reaches them. quality-engineer joins that rule rather than running on every full-mode loop — it is warranted when the change warrants at least one operational-safety module, is structural, or is explicitly requested. The three conditions reuse routers the catalogue already ships; none is a new predicate. adversarial-reviewer keeps correctness and scope; testability, reliability, observability, maintenance cost and test-strength judgment belong to quality-engineer alone, so work tripping none of the three conditions ships without that lens — a deliberate trade of the lens for speed. Recorded against RFC-0025 § Errata.

[core][2.26.22] — 2026-09-20

2026-09-20T00:00:00Z

Ordinary work stays on the short path, and the loop stops when the work is done. A work-loop risk trigger that names something you can modify — a security or governance boundary, a structure or public interface, a persistent representation — now fires when a change modifies that thing. Reading it, calling it, testing it, documenting it, or regenerating a projection of it no longer routes the work to full mode. The triggers that name an act still fire on the act, even when everything they touch is unchanged.
The loop stops searching once the work is done. Resolve the findings the current intent requires, and do not expand the frontier unless a finding blocks correctness, security, or a stated acceptance criterion. Meeting the stop condition ends the iteration: do not continue searching for additional improvements.
Not knowing a file no longer sends ordinary work down the heavy path. The Unfamiliar trigger asks whether you can predict the design, not whether you know the territory. Write down the design you intend and what grounds it; the trigger fires when you cannot, or would be guessing.

[core][2.26.21] — 2026-09-20

2026-09-20T00:00:00Z

A work-loop no longer drops a small fix just because it sits outside the task at hand. When it notices a mechanically verifiable fix along the way — one it can prove correct with a command, a search, or a check against a named authority — it now lands that fix instead of discarding it or writing it down for later.
Whether a fix rides along turns on its risk, its behavior, and how it was checked — not on how close it sits to the change already in flight.
A design call that is already settled — because a rule says so, or because the owner answered in one line — no longer disqualifies a ride-along. Anything that would still change a convention, a contract, or a published interface keeps requiring its own separately reviewed change.
Nothing noticed during a loop is dropped silently. At the end of a loop, every item gets routed somewhere; anything genuinely stuck is written down together with the one fact needed to unstick it later.

[product-engineering][0.13.14] — 2026-09-20

2026-09-20T00:00:00Z

Arriving with an unshaped idea, you can pick your shaping route from frame-intent alone, instead of inferring it from three skills' descriptions.
discovery-lead names the skill it runs at every gate, so a discovery can be driven from the agent definition without reading the loop's gate table.
A discovery hands off through work-intake whether or not Core advertises the handoff object, so the route into delivery no longer depends on the envelope.
Writing an assumption-test or delivery-contract slot, you can read its field set and its starting data classification off the sidecar schema.
diverge-solutions no longer asks for a minimum option count, so options are generated to expose the decision rather than to reach a threshold.
ux-writing output follows the same rendering directives as its sibling product-engineering skills, instead of falling back to unguided formatting.
frame-situation no longer tells you two shipped skills are unshipped.

[experience-design][2.0.7] — 2026-09-20

2026-09-20T00:00:00Z

The experience-design pack's own description of the copy handoff points at a skill that exists, so a reader following it reaches ux-writing.

[architect][0.15.14] — 2026-09-20

2026-09-20T00:00:00Z

`architect-design` now reaches for a system-shape concern when your design turns on one. When an open decision in your design depends on how a system shape coordinates — an event-driven system, a layered application, a distributed service — the skill now loads that shape's concept in full, so the concern reaches the document without you having to name the shape yourself. A design with a live decision in each of several shapes loads every one of them. Where no decision turns on a shape, none is loaded and the working receipt says so, so you can tell the axis considered your design and found nothing to add.

[architect][0.15.13] — 2026-09-20

2026-09-20T00:00:00Z

The paragraph-length check now sees a bold lead-in. Counting sentences needed a space straight after the full stop, so any closing mark standing between the two hid the break: an emphasised or code-spanned sentence, a parenthesised aside, or a closing quotation mark each read as no sentence end at all, and a paragraph that emphasised every sentence read as one. A long paragraph written with bold lead-ins could therefore pass the check it should have failed. Measured over the thirty-seven design documents in this repository's own architecture folder, the check reported eighty-one over-long paragraphs where the true number is one hundred and twenty-seven — it missed forty-six that crossed the budget unseen, and the mark that hid them was a bold span in one hundred and twenty-one of the one hundred and twenty-six mis-counted paragraphs. Those paragraphs are counted now, and the check reports the same file and line it always did.
A closing mark in any script closes a sentence. The marks recognised are every Unicode close-punctuation and final-quote character, so a document written with corner brackets, guillemets or angle brackets is counted the same way as one written with parentheses — and so is a strikethrough span. The set is those Unicode categories rather than every character that looks like a closer: an ASCII > stays out, because in Markdown it closes an autolink around a URL, and admitting it would misread more documents than it would fix. One shape stays undecidable and is unchanged: a single-letter sentence end inside a code span reads as an enumeration label, because an enumeration label written the same way has to keep reading as one.
The decomposition rubric names the third place a candidate can land. It offered a document of the part's own or a row in the element catalogue. A part can satisfy the qualifying rule and still belong above the document holding it, so the rubric now says to route it upward and how to record that. The case this catches is a document that carries the changes it is asking of a ratified parent: ratifying it would accept those changes by implication, because its own reviewer has no standing to accept them. The qualifying rule now says plainly that it does not settle which document holds the result, and that none of the six criteria will tell you: they ask whether a candidate differs from its parent, not at what altitude the people who must accept its decisions hold that authority. So standing is a question to ask separately every time the rule is met.
A design review no longer reports a correctly routed part as a missing document. The reviewer's decomposition checklist asked whether a qualifying child had a document of its own. A child whose decision only someone above this document can accept — a ratified parent's reviewer, say — has a link instead, by design, and would have been reported as a defect. The checklist now looks for either, and says which correction the part needed turns on standing. It also gains an item of its own for the case that started this: a document carrying the changes it is asking of a document above it that someone else accepts.

[architect][0.15.12] — 2026-09-19

2026-09-19T00:00:00Z

A design document now gets a runnable check for two of its ten quality gates. A gate script flags any prose paragraph over three sentences and any document over a 3,300-word size bound, both reported with the exact file and line, so those two checks no longer depend on a reviewer catching them by eye.
Reviewing a design document now asks for a verdict on all ten named quality checks, every time. The checks cover present tense, named cross-references, model-before-prose ordering, single-sourced concerns, removed decision history, diagram clarity, buildability from the models, linked evidence, and document size — so a check nobody looked at and a check that passed no longer read the same in the result.
Seven of those checks now give an author narrowing hints before a design document is ever sent out for review, including future-tense wording, an unnamed cross-reference, and a model-after-prose section — a reviewer still decides, but an author sees them first.

[core][2.26.20] — 2026-09-18

2026-09-18T00:00:00Z

A pull-request template ships with the pack. work-loop now carries assets/pull-request-template.md, with a copy step for GitHub and for GitLab in the adapt-to-project guide. It is not installed for you: a repository that already has a pull-request convention keeps it. The template states three things that do not belong in a body — what you tried, the method behind a result, and the diff restated — in comments that do not render, so an author reads them while writing and a reviewer never sees them.
The template ships with no checklist. Its verification section asks for commands and their observed results, which a checkbox cannot supply.
A `Review focus` section is available and optional. It names the one decision the author wants checked, as a question about the code.
Worked pull-request bodies and writing rules. A new references/pr-authoring.md carries seven rules and two worked bodies of different size, so the shape is shown scaling down rather than only described.
The end of a run asks before it offers. work-loop decides whether it can open a pull request from the exit status of gh api user and the value of viewerPermission, and when either refuses it completes the checklist and says nothing. A blocked credential store makes gh auth status report an invalid token and gh repo view report a connection failure, so neither message states the cause — an exit status makes no such claim.

[core][2.26.19] — 2026-09-18

2026-09-18T00:00:00Z

A shipped spec is the record of what was agreed, not a rule the code must keep obeying. An agent that meets an older spec disagreeing with today's change now reads that as the system moving on, and says so instead of raising it as a defect. The spec directory your repository is seeded with says the same thing in plain words, so a person opening it meets the limit of a spec's authority beside its conventions.

[core][2.26.18] — 2026-09-18

2026-09-18T00:00:00Z

A design decision in a plan now names the task that builds it. Every ## Design (LLD) sub-section carries an Owned by: field alongside its existing Traces to:, and the pre-review walk checks that every interface, type, symbol or ownership decision has an owning task. The traversal runs one way only — design down to tasks — because over a third of tasks have no design origin and a reverse rule would fire on every one of them.
A plan that uses the field gets its ownership checked. Once any sub-section carries Owned by:, the contract-alignment lint reports a body-bearing sub-section that names no task, and a named task that no heading defines. A plan written before the field is reported as predating it, never failed, so existing plans need no migration.
The plan template no longer promises an edit the engine refuses. It said an implementer corrects the design in place "without an amendment"; approval hashes the whole plan, so that edit was always rejected. The permission is now bounded to before approval, and the two post-approval cases are named: grounding for a seam the plan marked no stub (implementation-discovered) goes to the verification ledger, and a settled decision that execution falsified is a plan error taking controlled amendment.
A review finding is now classified by cause depth before it is answered. A finding whose cause is a design decision repairs the design and every task that decision owns, rather than patching the one place it was spotted.

[core][2.26.17] — 2026-09-18

2026-09-18T00:00:00Z

A plan now names the production code each of its checks drives. A task's Tests: bullet already owed the suite it lives in, the fixture carrying a join key, and the shipped assertion the change moves; it now owes the production code the check exercises as well. A reader of a plan written from here on can see what each check is meant to hold against without opening the test.

[core][2.26.16] — 2026-09-18

2026-09-18T00:00:00Z

A spec now opens with what changes, not with an essay. The Objective section is renamed Outcome and capped at two sentences — who this is for, and what success looks like for them — and a new What Changes section sits right under it carrying the delta as bullets. A reader landing on a spec cold gets the shape of the change in the first screen instead of hunting for it.
Boundaries is now called Agent Rules. Same three tiers, same meaning, a name that says what the section does. Specs written before this keep the old heading and stay valid: every reviewer that reads the section now recognises both names.
The Assumptions section holds open questions only. It used to be an audit trail — every fact the author checked, with a citation for how. Those facts now go where they do work: a fact that bounds the delivery goes into the spec body, a fact that shapes how it is built goes into the plan's design. What is left in Assumptions is what nobody has answered yet, so a reader sees the gaps instead of scrolling settled ground.
A plan's Changelog records approvals and nothing else. The dated diary of how the approach evolved is gone; the two approval entries, which are the only record of when the contract froze and on whose authority, stay exactly as they were.
A plan task writes an Approach only when it has a decision to record. Where a task's Tests and Done when already say what to build and how it is observed, the Approach was restating them. It is now written for an ordering or a seam decision a reader cannot otherwise infer.
New advisory guidance on prose that reads as generated. A prose-discipline reference gives the author a signal-word scan, a list of canned rhetorical shapes, and the rule that the repair is restructuring rather than swapping synonyms. It is advice, never a gate, and it defers the form rules to the output-rendering block each skill already carries.

[architect][0.15.11] — 2026-09-18

2026-09-18T00:00:00Z

Design docs now match the size of what they describe. Writing about a whole application or system, a single subsystem, or a change to an existing architecture each gets its own template, so a reviewer no longer has to fill in eight generic sections that half apply and skip the rest.
Every modelled section leads with its model, not a page of prose about it. Each template opens with the question the section answers, and where the section carries a diagram or table it puts that first and only then explains why it looks that way — so a reader can see the shape of a decision before reading the argument for it.
A new rubric says how many documents a design needs, not just what goes in one. Six criteria decide whether a part of a subsystem deserves a document of its own, so a design doesn't grow past its own boundary just because a subsystem's story kept expanding, and a parent document says plainly what it still owns once children split off.
Existing links to the earlier design-doc template keep working. It stays at the same path and now says plainly what it is: a kept pointer to the three templates, never chosen for a design itself.
Design and review now speak the same three scopes. The review checks ask the same scope-shaped questions the templates were written to answer, so a design that follows its template's model doesn't get marked down against a checklist that expected different sections.

[architect][0.15.9] — 2026-09-18

2026-09-18T00:00:00Z

Choose the right Mermaid layout for your renderer. The diagram guidance now explains that releases before Mermaid 12 default to dagre and need a separate ELK package, while Mermaid 12 and later default to bundled ELK.

[core][2.26.15] — 2026-09-18

2026-09-18T00:00:00Z

A wave can no longer be signed off without saying who did the work. The step that closes a wave used to pass no matter what: it returned success for any state it could read, so a wave whose tasks were never dispatched looked exactly like one where every task was done. It now asks for one short record per task in the wave — either "this was dispatched" or a declared reason it was not — and refuses the exit by name when a record is missing, listing the tasks it could not account for rather than a count. The sibling command that steps the wave pointer forward asks for the same thing, so the check cannot be walked around by moving the pointer first.
Nothing already in flight breaks. A run whose state predates this feature carries no records at all, and the exit passes it with a notice saying enforcement is off rather than blocking it, so an upgrade mid-run still finishes. loop-cohort status reports whether records are being enforced.
A run that genuinely has no implementer can still proceed. A wave task can be declined for one of two declared reasons — no implementer is installed, or a human directed the work another way — and a declined task counts as accounted for.

[frontend-engineering][0.3.0] — 2026-09-18

2026-09-18T00:00:00Z

The frontend pre-flight now reads the design handoff your team already wrote, instead of picking an aesthetic direction of its own. Point [design] output_dir at your design directory and it takes the aesthetic direction, the per-screen brief and the token taxonomy from there. Its own canonical reference list is now only the fallback, and only for a slot no artifact filled.
It keys on no section name inside your files. Design templates are scaffolds people edit, so a reader that looked for named sections would find nothing in most real documents — it takes the first heading, the frontmatter as it finds it, and the body whole.
A file under a read path that is not the artifact — a different type:, a draft, another kind of design record — is skipped and the scan continues. A design directory holding many kinds of document is expected, not an error.
When the directory is a personal vault or anywhere outside the repository, it asks you about each artifact before using it, and shows you what it found. Nothing in these files says which product they belong to, so that confirmation is the check — it is never reported as though a test had passed.
Every refusal stops the whole read and halts, by name. It does not tidy up a bad value, pick a different file, quietly downgrade to a skip, or fall back to the canonical list. [Read the design handoff](../../guides/frontend-engineering/how-to/read-the-design-handoff.md) says what each skip and each refusal means, and where the limits are.

[experience-design][2.0.6] — 2026-09-18

2026-09-18T00:00:00Z

Compare creative directions on shared terms. Each direction now records its commitments across fifteen named axes, including seven structural axes, so teams can identify a meaningful difference instead of arguing from taste.
Test whether a direction is distinctive before carrying it forward. A counterfactual pass reworks a similar brief, then records changes wherever the result falls back to a generic default.
Start with a named visual direction. Swiss / International Typographic, editorial broadsheet, and Bauhaus directions provide ready-made starting points for exploration.
Check that candidate directions truly diverge. A divergence audit identifies the closest pair, showing whether the options differ in substance rather than only on average.
Review a surface through two published instruments. Design reviews can now rate perceived quality and perceived genericness separately.

[core][2.26.14] — 2026-09-17

2026-09-17T00:00:00Z

A plan now records who approved it, and when. Approving a spec and plan froze the contract and left no trace in the artifacts of when that happened or on whose authority. Reading a shipped delivery months later, you could not tell which claims were in the contract at approval and which arrived during the work — the plan's own history started at drafting and resumed at execution, with the gate itself unmarked. Each approval is now a dated entry in the plan's changelog, written by the approver in the same edit that sets the status.

[governance-extras][0.11.0] — 2026-09-17

2026-09-17T00:00:00Z

New ADRs now record how easy a decision is to reverse and which areas it touches. The new-adr template adds Reversibility and Areas fields, plus four supersession fields that stay none until a decision is actually reversed. Coining an Areas value nothing else in your docs/adr directory already uses now asks for your explicit confirmation first, so a typo does not quietly start a new label on its own.
Reversing a decision no longer means editing the old record. Point a new ADR's Supersedes: field at the one it replaces; the old record's own Status and Superseded by: fields pick up the link. Its text stays as written — history, not something the new decision rewrites — and the generated index shows the pointer from both records.
A shape check for your own decision records. new-adr and new-rfc bundle a script that reads every record's metadata block and reports anything missing, unmatched, or shaped wrong, over the whole directory in one pass rather than stopping at the first problem.
Correcting an accepted decision has one place to write it. Add a dated entry under ## Errata instead of editing the decision itself; entries are only ever added, never rewritten or removed.

[core][2.26.13] — 2026-09-17

2026-09-17T00:00:00Z

Your agent decides where an unsettled assumption belongs. While drafting a spec, it surfaces the assumptions it is making and checks each one before filing it. Some cannot be settled at that point — what an existing helper really does, what a generated file actually contains, what a check truly proves. Each of those is now placed by a rule rather than written down as-is. The rule lives in the new-spec skill's assumptions step, under the anchor load-bearing-claim-routing; it has one home, and this entry does not restate it.

[core][2.26.12] — 2026-09-17

2026-09-17T00:00:00Z

The spec checker now notices when a criterion was reworded and its assertions were not. It could already do this, but only if you knew to hand it a revision to compare against. Without one it said nothing and reported a clean run, so the check was there and almost never ran. It now works out what to compare against on its own — the point your branch left the default branch. Nothing new can fail your build: this check describes what it found and never sets the exit code. Pass --no-since if you want it left alone.

[core][2.26.11] — 2026-09-17

2026-09-17T00:00:00Z

Your agent now reads one conventions file instead of two. The seeded AGENTS.md was a short file that pointed at a much longer docs/CONVENTIONS.md for the rules that actually govern work. Two subagents loaded that longer file in full on every dispatch — roughly 21,900 tokens of context spent before either had read your code. AGENTS.md now states the rules an agent needs to start work: the commit format, the four pull-request questions, the privacy rule, how changes get scoped, when to confirm before a destructive step, and what to do when documented guidance and code disagree. Nothing was dropped to make it fit — the rest moved to whichever artifact already owned it, so the answer is in the guide, skill or architecture document you would have reached for anyway.
A new `docs/README.md` tells you what belongs where. Installing the pack gave you four documentation directories and nothing explaining which one a new document goes in. The map names each area, what belongs there, and its lifecycle class — living, frozen, or governance — and it carries a row for you to extend when you add an area of your own. The seeded AGENTS.md routes to it, so an agent entering your repository has a way into your documentation rather than a dead link.
The seed stopped offering you sections it already has. Its trailing comment listed Documentation, Security considerations and Scoped instructions as guidance to add once you had a reason. All three are now real sections, so the comment no longer suggests adding what you have. Repository structure is still offered, because the seed still does not carry it.

[experience-design][2.0.5] — 2026-09-17

2026-09-17T00:00:00Z

Four design skills now write where you point them. creative-direction, design-principles, information-architecture and design-system each declare one canonical target under the [design] output_dir you set in agentbundle-layout.toml. Before this release design-principles wrote to a fixed docs/design/principles/... path that ignored your configuration, information-architecture used a location no skill declared, creative-direction named none at all, and design-system produced no file — it now writes a token taxonomy your frontend work can read instead of re-deriving tokens every session.
Treat the pack's confinement as a strong default, not a boundary. Each of the four is instructed, before writing, to state the approved output directory, the file it read that value from, and the target path it composed. One full run was captured against the shipped wording — creative-direction, which stated all three — so this is a measured instruction for that skill and an unobserved one for the other three. The check that keeps a write inside that directory is an instruction the agent has to carry out, not something the pack enforces. Against a target reached through a symlink it was skipped in 2 of 8 observed runs, and every skip wrote outside the approved directory. The step was rewritten to require an executed real-path resolution and has refused on every run observed since, but a skipped check still leaves no trace: the file exists and looks ordinary. If you need a guarantee rather than a strong default, enforce confinement outside the agent — in the filesystem, or in a tool that refuses the write.
The how-to guides now say what each skill actually does. Nineteen steps published a "where it lands" path. Eight named a file the skill never writes, and several sent you to an aesthetic/ folder nothing produces. Every step now gives either the path its skill writes or a plain statement that it writes no artifact, so you stop waiting for a file that was never coming.

[core][2.26.10] — 2026-09-17

2026-09-17T00:00:00Z

Your loop no longer halts when the same findings come back. It tells you and keeps going, bounded by the iteration cap. The stop it replaced never fired: the signal it read is equality over a fingerprint that carries a finding's position, so any repair that moved a finding made the round look new.

[core][2.26.9] — 2026-09-17

2026-09-17T00:00:00Z

Your implementer looks for what you already have before it writes something new. Hand it a task whose best answer is a helper sitting elsewhere in your repository and it now finds that helper and uses it, instead of writing a second copy of the same logic a module away. When it finds something that only partly fits, its report names that helper too, so a near-miss is visible to you rather than silently passed over.
A plan that asks for more structure than the job needs no longer gets built as written. Where a lighter route satisfies the task's own "done" check, the implementer takes it and records the swap for you to see, rather than faithfully building a configuration class and a new module to do the work of one function. Structure you genuinely need survives: when several callers really do need different settings, it still builds the shared piece.
The loop is now asked to weigh each addition it declines against a stated rule. Its "tempted to add, chose against" note asks for the "Cut before adding" step behind each line, so the reasoning sits against a written rule rather than a one-off justification, and where no step covers the call — an accepted requirement or a safety control forbids it — it is asked to say that instead of fitting a rule to it after the fact. This one is guidance the note requests rather than a check that enforces it, so read the notes rather than relying on them.

[core][2.26.8] — 2026-09-16

2026-09-16T00:00:00Z

A plan that names a task you never wrote no longer runs in the wrong order. loop-cohort schedule used to drop a Depends on: entry pointing at a task the plan did not contain — no warning, no error. The dependency edge vanished, the tasks collapsed into one wave, and because gates run per wave, the gate run collapsed with them. schedule now stops, names every offending task->dependency pair, and writes no state, so you fix the plan instead of discovering the wrong order later.

[agentbundle][0.47.0] — 2026-09-16

2026-09-16T00:00:00Z

Sending your work-loop telemetry no longer needs `agentbundle` installed to build the command. jsonl-otlp-export reads [telemetry] from your repository and your user agentbundle-layout.toml itself, through its own --config and --user-config, so anything that can run a command can start an export — a hook included. Where you put your settings, and the rule that a repository value wins setting by setting, have not changed.

[core][2.26.5] — 2026-09-15

2026-09-15T00:00:00Z

Your work-loop runs can now be watched in an observability backend. Install jsonl-otlp-exporter, point it at your own OpenTelemetry Collector, and each phase change arrives as a log record — how long the phase took, which gate decided what, and how close the run is to its retry budgets. It is off until you configure an endpoint, and the thing that sends is a separate distribution you install on purpose, so a catalogue install alone can never transmit.
What would be sent is written down before you turn it on. The how-to guide names the file that is read, the records that leave, the fields that may go, and the things a pack cannot see at all — your prompts, the model's replies, and token counts.

[agentbundle][0.46.0] — 2026-09-15

2026-09-15T00:00:00Z

A self-hosted catalogue can now be recreated from its recorded recipe. A repeat catalogue init --preset self-hosted keeps the recorded catalogue identity and pack and profile selections unless you explicitly replace them.

[agentbundle][0.45.0] — 2026-09-14

2026-09-14T00:00:00Z

`agentbundle` can now work out where to send your loop telemetry. A new resolver reads your repository and personal layout files and returns the exact arguments for the separately installed sender. Repository settings win per setting, because sending data off the machine is a team decision rather than a personal one — the opposite of how personal research settings resolve.
`catalogue lint` now tells you when an optional tool a pack wants is not installed. It names the package and exits 0. It never installs anything.

[core][2.26.3] — 2026-09-14

2026-09-14T00:00:00Z

workspace-status can now prune an explicitly approved selection of delivery artifacts. A successful exit guarantees that every selected artifact directory and every workspace membership resolving to it are absent, proven in one coherent observation taken while the shared workspace lock is held. If either half survives, the command fails and names which one, rather than reporting success.
An interrupted prune is now recoverable from the command's own output: a failure names the selection and how far it got, and a busy lock names the lock file and the process holding it.
Creating a durable work artifact now coordinates with that same lock, so an intake and a prune can no longer interleave and leave the register disagreeing with what is on disk.

[core][2.26.2] — 2026-09-14

2026-09-14T00:00:00Z

`workspace-status` shows you which initiative is which again. Orientation used to list bare slugs, because the projection replaced each initiative's name and milestone with a stand-in meaning "this value was not safe to hand to an agent." That guard was added when display text from workspace.toml was treated as untrusted input. It is not: workspace.toml lives inside the repository it describes, so anyone who can reach it already has the source, and the guard was costing readability for a risk that was not there. A session now opens with ini-002 — Platform Core (milestone: P5 · Adopt (M1–M5 shipped)).

[core][2.26.1] — 2026-09-14

2026-09-14T00:00:00Z

`workspace-status` no longer tells you an initiative is called "workspace.toml". An initiative's name and milestone are free text you type into workspace.toml, and the status tool deliberately never passes them through — they always come back as the literal workspace.toml, the same stand-in every other field uses to mean "this value was not safe to hand to an agent." The skill's own instructions had not caught up and still told the agent to print them, so a status run announced ini-002 — workspace.toml (milestone: workspace.toml). Active initiatives now show their slug alone.

[agentbundle][0.44.2] — 2026-09-14

2026-09-14T00:00:00Z

A catalogue you derive now keeps a working check on its vendored `credbroker` code. catalogue init --preset self-hosted copied the vendored copy of that code but left its source behind, and the check that compares the two passed by finding nothing to compare. The copy was frozen: an edit to it went unreported, and catalogue self-host never produced the staged floor it is meant to write. Selecting the credential-brokers pack now brings the source across too, so the comparison is real again.
`--attribution white-label` no longer leaves the upstream catalogue's name in a file you commit. That mode promises no upstream trace, and the scan enforcing it allows the upstream name zero occurrences anywhere — but the scan runs before .agentbundle/self-host-state.json is written, so the name landed there unchecked. The file now records your own catalogue's name. --attribution attributed is unchanged and still records the upstream name.

[frontend-engineering][0.2.5] — 2026-09-14

2026-09-14T00:00:00Z

A surface can now declare the minimum viewport width it supports, and the rendered-page inspection stops asking for captures below it. A desktop-only tool that supports 1280 and up is asked for four captures per route in one channel instead of eight across two.
Declaring a minimum never hides a width the surface does claim. It only ever raises a band's lower bound, so a minimum that falls between the default bands leaves the wider one where it is rather than pulling a capture down into a gap that satisfies no channel.
A breakpoint above the minimum keeps its own band, so a high minimum cannot flatten a surface that really is responsive above it. Each run records the minimum it used and any breakpoint the minimum discarded, so a mistyped minimum is distinguishable from a deliberate single-channel surface.

[core][2.25.27] — 2026-09-13

2026-09-13T00:00:00Z

When a review finding says a claim promises more than its test checks, you now have two answers instead of one. Previously the only available answer was to shrink the claim, so the easy move was to weaken what you promised until the existing test covered it — quietly dropping the property from what anyone checks. Now: if any test can be made to cover the full promise, improve the test; shrink the claim only when none can.
After fixing a review finding, search twice. Guidance used to say this search was "a walk, not a text search", which read as permission to skip searching for exact strings. You now do both: search for the repeated words and names, and read for the places that say the same thing in different words. After a fix, also re-run the search for tests that pin file contents, across every file you touched — that is when an edit breaks a test in a file you never opened.
You can now move an obligation out of an intent or brief and still have something catch its deletion. Moving it requires something that fails if it is removed, and three of the four places you were told to move it to had nothing watching them. Editing Opportunity on an intent, or Rabbit holes or Design artifacts on a delivery brief, now counts as a material change, which cancels that artifact's approved review and sends it back for a fresh one.
How to prove a test actually catches a bug is now written down. To prove it, put back the original broken code — not an empty placeholder — and confirm the test fails. Undo it by editing, not with git checkout, reset, or stash. A test that still passes while the bug is back has proved nothing. This rule previously existed only if your own process happened to supply it.

[core][2.25.26] — 2026-09-13

2026-09-13T00:00:00Z

A frontend review now gets the pictures, not just the diff. When a surface has been inspected, the work-loop hands frontend-reviewer the capture set and what was observed in it, alongside the diff. A reviewer that only ever saw a diff could not see one element covering another, which is the class of defect it was most often asked about.

[frontend-engineering][0.2.4] — 2026-09-13

2026-09-13T00:00:00Z

The rendered-page inspection now looks at both sides of your breakpoints. It used to permute viewport height and scroll position and nothing else, so a complete capture set could sit entirely on one side of every breakpoint a surface has — exercising each layout rule only where it already applies. A channel is a band of viewport widths, and a set is complete only when it covers every one your surface has.
Tell it your breakpoints and it captures either side of each. Declare them and the bands they bound become the required channels, captured at the widths that straddle each boundary — declare 1152 and you get 1151 and 1152. Declare none and two bands apply, narrow at ≤480 and wide at ≥1024, and the run records that it fell back. Those two leave 481–1023 uncaptured, which is where a great many real breakpoints sit; declaring yours is what closes it.
The floor rises from four captures per route to eight. Four heights-and- scroll-positions in each required channel, and four times *n + 1* where you declare *n* breakpoints. An existing capture set becomes incomplete at this version, which is the intended signal.
A capture's width now means something to the reviewer. frontend-reviewer is told to use it: it decides which breakpoint-scoped rules a capture exercised at all, so a failure absent from one channel is no evidence it is absent from another.

[core][2.25.25] — 2026-09-13

2026-09-13T00:00:00Z

You can now ask workspace-status which explicitly selected specs still have workspace membership and which do not, without changing workspace.toml or treating absence as permission to delete anything.

[core][2.25.24] — 2026-09-13

2026-09-13T00:00:00Z

A sustained shaping finding no longer reads as an instruction to edit. Intent, delivery-brief and spec authoring point at the work-loop's ordered response ladder — cut, route, fix, hold, walked until one applies — so an author can answer a finding by dropping a claim nothing is obliged by rather than writing more careful prose. The lifecycle gates are unchanged: whatever blocked before still blocks.
Each upstream artifact states where a demoted claim belongs. An intent sends a settled ground to Opportunity and a matter decided without the authority to decide it to Unresolved questions. A brief sends a design trap to Rabbit holes and informing provenance to Design artifacts, and never to Ready gaps, which it drops on leaving Draft.
A shaping reviewer no longer reopens a decision the artifact records as settled. A pre-existing defect stays reviewable however late it is found, so an author can refuse a reopened decision without dismissing a real defect from the same round.

[core][2.25.20] — 2026-09-13

2026-09-13T00:00:00Z

A legacy-only repository can activate its knowledge base even when the migration has nothing to import. Activating was the first command such a repository runs, and it crashed on the state those repositories are actually in: an empty patterns.jsonl. The migration now produces the canonical empty topic map instead of a traceback.

[frontend-engineering][0.2.3] — 2026-09-13

2026-09-13T00:00:00Z

The pack now looks at the page, and writes down what it saw. Step 4 of the journey captures each route you name at two viewport heights, at rest and scrolled, and reports what a reader would actually meet — something covering something else, text running out of its container, a control too small to hit. The evidence manifest gains an inspection observations field that a list of screenshot filenames does not satisfy.
A skipped inspection can no longer pass as a completed one. Seven distinct result states replace a single unverified line, and the same state reaches the manifest, the step's output, and the acceptance gate — so a missing browser arrives as a decision rather than a green tick.
Measure the noise yourself. Four defect fixtures and four known-clean ones ship with a procedure for measuring the false-positive rate in your own environment. The pack publishes no rate of its own, because the rate moves with the judge and the viewport sizes and does not transfer.
A visible defect now costs something. A run that took every capture and found a banner covering the heading reports that it *ran* and that it did not *pass* — two answers, because "the browser would not start" and "the page is broken" need different fixes. The surface does not complete over an unresolved blocking finding.
The independent reviewer can finally see the page. frontend-reviewer is handed the captures, reads them, and can take its own when the ones it was given do not cover what the diff makes it suspicious of. Until now it read the diff, and no diff shows one element covering another.

[core][2.25.19] — 2026-09-13

2026-09-13T00:00:00Z

The conventions you install now say which projected paths a gate actually watches. Two entries were wrong: the repo hooks target is adapter-driven but no drift gate reaches it, and several seed-projected paths were described as hand-owned when the pipeline still regenerates them. Membership is decided by the exclusion list, not by the list in the prose.

[core][2.25.18] — 2026-09-12

2026-09-12T00:00:00Z

Shaping an intent no longer draws spec-grade craft findings on a four-paragraph artifact. The intent review answers one question — is this well-formed enough to shape further — in a vocabulary that cannot express a rewrite of your bet.
A second, optional read attacks the bet itself and may only hand back an open question with a named decider or a kill condition with its real-world trigger. It cannot ask you to reword anything, and an empty answer is a real answer.

[product-engineering][0.13.12] — 2026-09-12

2026-09-12T00:00:00Z

After framing an intent, you can have the bet attacked rather than the wording reviewed — and whatever comes back is advisory, so it never moves the intent's status on its own.

[core][2.25.16] — 2026-09-11

2026-09-11T00:00:00Z

`new-spec` now tells you where a fact belongs, and hands you the facts to decide with. Plans kept accumulating design reasoning inside task lists — why an assertion had to take a particular shape — where it went stale against everything else and generated review findings of its own. The rule is now explicit: a fact belongs in the design unless a task must implement or verify it, because tasks are jobs to be done rather than containers for facts. And when a spec reaches you for approval, the skill reports the finding trend and every remaining concern with its consequence, rather than a verdict you cannot check.

[agentbundle][0.44.0] — 2026-09-11

2026-09-11T00:00:00Z

Installing a pack now sets up where its output goes, instead of quietly doing nothing.
Your agentbundle-layout.toml keeps its comments, formatting and permissions when a pack is installed.

[core][2.25.14] — 2026-09-10

2026-09-10T00:00:00Z

You can now measure how long each work-loop phase took, straight from the run log. Every transition line in .loop-run/events.jsonl carries when its phase began and how many seconds it lasted, so time-in-phase needs no guesswork about what happened between two entries. The first line's start is the run's own start, so the opening phase is a real measurement rather than a lower bound.
The retry budget is on the line, and so is the override flag. Every line carries the implementation and review retry counters beside their caps, so you can watch a run approach a limit without opening the cohort state file. A transition that carried --allow-retry-cap-override records that it did — the flag as passed, which is not the same as a cap having been exercised.
A repair cycle no longer looks like a give-up. Each line now says what the gate decided and, separately, whether the retry budget behind it still had room, so a routine repair round reads differently from a run continuing past a spent budget. Read the reachability note below before relying on the second case: on the ordinary path it does not arise.

[core][2.25.13] — 2026-09-10

2026-09-10T00:00:00Z

A machine that cannot run `git` now gets one line of explanation instead of a stack trace. Every loop-engine and loop-cohort command needs git to confirm the spec directory sits inside your repository. When git could not be run at all — not installed, or a PATH entry shadowing it with something unexecutable — the command ended in a 33-line Python trace that also printed internal file paths, more output than a whole successful run of the loop. Both tools now answer that situation the same way, in one line.

[core][2.25.12] — 2026-09-10

2026-09-10T00:00:00Z

Review feedback is guidance about the criterion, not an edit request for the sentence. new-spec's authoring rubric now says how to repair from a finding: name the property the clause establishes, ask whether that property is right, and look for its siblings before rewriting. It also gives you a way to notice when you have been patching text instead — and explains why a check that reads a whole document can silently pass on a second copy of the sentence it was looking for.

[core][2.25.11] — 2026-09-10

2026-09-10T00:00:00Z

Checking where the work stands costs a fraction of what it did. The orientation payload now carries the decisions you act on rather than the full working set behind them, and the saving grows with the size of your workspace. Ask for the long form when you want it.
The spec and traceability checks report what they found instead of listing it. A passing run prints its summary lines and says how many items it withheld, so nothing goes quietly missing. Anything that fails still prints in full.

[core][2.25.10] — 2026-09-10

2026-09-10T00:00:00Z

The directory your intents land in is now described in the conventions you install with them. CONVENTIONS.md reads an intent on two axes: the altitude it sits at, and how far it has travelled from a framed bet to a ratified outcome. Only a feature-level intent becomes a spec directly.
Admitting an intent that was already shaped upstream adds to it instead of overwriting it. The outcome, altitude and assumptions you framed survive the handoff into repository work.

[core][2.25.9] — 2026-09-09

2026-09-09T00:00:00Z

The defects a spec review keeps finding are now named where the spec is written, not only where it is reviewed. new-spec ships a spec-authoring rubric: six failure classes worked in order before criterion shape, starting with the one no criterion craft repairs — an obligation authored where an owner already exists.

[agentbundle][0.43.1] — 2026-09-09

2026-09-09T00:00:00Z

A new catalogue now gives Claude Code the same pack and profile guidance every other agent already got. agentbundle catalogue init writes a CLAUDE.md beside each AGENTS.md it scaffolds, so a Claude session picks up your scoped instructions instead of missing them.

[agent-skill-engineering][0.4.2] — 2026-09-09

2026-09-09T00:00:00Z

A skill author deciding whether to hand work to an isolated worker, or whether to ship a set of components as one package, now gets those concepts from the corpus instead of being routed to a runtime profile that does not exist.

[core][2.25.8] — 2026-09-08

2026-09-08T00:00:00Z

A cooled delivery's parent link is now declared on its workspace entry, so reconciliation stops guessing. When a cooled spec's source.parent is missing or names no registered brief, workspace-status reports cooled_child_scope_unknown against that entry and holds every dependency on a brief until someone answers — instead of silently reading the missing value as "this delivery has no parent" and releasing the work.
Declaring `none` is a real answer. A delivery with no parent brief is declared as such once and dispatches normally, so the new refusal is something a maintainer clears rather than waits out.

[core][2.25.6] — 2026-09-08

2026-09-08T00:00:00Z

`workspace-status` now distinguishes unavailable cooling support from a malformed lifecycle record. It directs maintainers to install or repair the cooling capability instead of repairing a valid record.

[agentbundle][0.43.0] — 2026-09-08

2026-09-08T00:00:00Z

Installed direct skills can now move to later repository revisions. agentbundle upgrade --skill <name> re-runs admission, shows the capability consent surface, previews changes, and removes files dropped upstream before it records the new revision.

[core][2.25.5] — 2026-09-08

2026-09-08T00:00:00Z

`work-loop` now rejects unsafe or ambiguous reference providers before it calls them. Provider guidance is visibly enclosed as untrusted evidence, so it cannot quietly expand the planning workflow's authority.

[architect][0.15.7] — 2026-09-08

2026-09-08T00:00:00Z

`architect-design` now rejects unsafe or ambiguous reference providers before it calls them. Provider guidance is visibly enclosed as untrusted evidence, so it cannot quietly expand the design workflow's authority.

[core][2.25.4] — 2026-09-08

2026-09-08T00:00:00Z

Closeout now keeps unrelated work visible when it shares a path with a cooled legacy entry. A second entry at the same path still blocks closeout, so maintainers are not told to invoke close-work while that work remains unresolved.

[core][2.25.3] — 2026-09-08

2026-09-08T00:00:00Z

A durable owner accepting an artifact now survives the session that recorded it. The lifecycle record carries Reclassified, so "this left delivery and someone else owns it now" persists as readable state instead of being lost at closeout. A reclassified artifact keeps its file where its record says it is, drops out of ordinary orientation, is never due for a retention review, and is not counted among the obligations someone still owes delivery work against.
It is reached deliberately and only once. Reclassification is available only from a record already retained under an exception, only when the durable owner's acceptance is supplied and validated at the transition, and never again afterwards — nothing transitions out of it, and no deletion route admits a lifecycle record. It is not gated on a date, because a durable owner accepting an artifact is not a day-30 event.

[core][2.25.2] — 2026-09-04

2026-09-04T00:00:00Z

Light-mode review now runs until it is clean instead of stopping after a fixed number of rounds. A loop that is converging finishes. At the third round and every second round after, the loop asks whether findings are getting fewer and smaller, and keeps going while they are. A loop that is no longer converging asks whether the reviewed construct should exist at all and puts that choice to the person who asked for the work, rather than sending a small change through the heavier mode.

[core][2.25.1] — 2026-09-04

2026-09-04T00:00:00Z

When a work-loop task is about a skill, a skill script or evaluation, an agent loop, a hook, or a plugin, planning now reaches the installed agent-skill-engineering reference for compiled guidance and cites the topics and provenance it used, instead of answering from model memory.
Without that pack installed, work-loop records knowledge provider unavailable and completes the same planning work it did before.

[architect][0.15.6] — 2026-09-04

2026-09-04T00:00:00Z

When an architect-design task turns on how to package agent behaviour — a subagent, a hook, a skill, or a plugin — the design procedure now reaches the installed agent-skill-engineering reference for compiled guidance and cites what it used.
Without that pack installed, the procedure records knowledge provider unavailable and produces the same design work it did before.

[agent-skill-engineering][0.4.1] — 2026-09-04

2026-09-04T00:00:00Z

The shipped knowledge provider contract now states every diagnostic the seam can return, so a consuming pack can quote a receipt an adopter actually receives rather than one that existed only in a test fixture.

[core][2.25.0] — 2026-09-04

2026-09-04T00:00:00Z

A work-loop phase now decides which behavioural rules an authoring agent is taught. A registry inside the work-loop skill declares the policy families each phase carries, and a selector turns one phase into the ordered families it selects, each with its enforcement tier and a fingerprint of the file that teaches it.

[core][2.24.3] — 2026-09-04

2026-09-04T00:00:00Z

A lesson that a rule or a control has since absorbed can be retired where it lives. Distillation now carries the lifecycle and the retirement record, so a topic settles as canonicalized, enforced, obsolete, merged, or invalidated, naming the successor that took the lesson over. A retired topic stays in place for history and stops being offered as current guidance.
A topic keeps every occurrence that produced it. Reconciling a second observation into an existing topic appends to its provenance instead of replacing it, so the record of what was seen, and when, survives revision.

[product-engineering][0.13.9] — 2026-09-03

2026-09-03T00:00:00Z

Shaping skills now suggest a workspace entry that can actually be picked up. frame-situation and diverge-solutions printed a short legacy form that is never dispatchable, so an entry added exactly as instructed never surfaced as ready work. Both now print the canonical five-field entry and route registration through work-intake.
A reference to a command that no longer exists is gone. The skills pointed at a retired skill name, so following the instruction did nothing.

[product-strategy][0.2.5] — 2026-09-03

2026-09-03T00:00:00Z

The OKR cascade now writes a workspace entry that can be picked up. run-okr-cascade records each gap it finds, and it was writing a legacy form that is never dispatchable — so every gap it recorded was invisible to the shaping room it was recording into. It now writes the canonical five-field entry.

[experience-design][2.0.3] — 2026-09-03

2026-09-03T00:00:00Z

Every experience-design skill now tells you how to start it. None of the pack's twenty skills documented an example phrase, so a person had to guess the wording and the agent had nothing to match against. Each skill now carries two to four phrases someone would actually type, sharpened so neighbouring skills in a twenty-skill pack do not claim the same request.

[frontend-engineering][0.2.2] — 2026-09-03

2026-09-03T00:00:00Z

Every frontend-engineering skill now tells you how to start it. As with experience-design, none of the nine skills documented an example phrase. Each now carries example utterances that match what the skill genuinely handles.

[core][2.24.2] — 2026-09-03

2026-09-03T00:00:00Z

Evidence from a run now has a place to go that the approved plan does not. Once a plan is approved it is pinned, so a task that asked for its own observed result to be written back into it could never be satisfied. Plans now keep the obligation and a sibling verification ledger keeps the observation — the mutation that was applied, the check that went red, the digest that was compared. One convention states when substantive edits stop, and a check fails if that statement is reverted.

[core][2.24.0] — 2026-09-03

2026-09-03T00:00:00Z

Work-loop now gives each scheduled plan task a bounded sequential implementer dispatch. The controller retains lifecycle, review, and closeout ownership while the task brief carries its execution root and applicable craft guidance.

[core][2.23.2] — 2026-09-03

2026-09-03T00:00:00Z

A local dependency can keep resolving after the delivery it names has been closed out and pruned. Put a completion receipt with delivery_id, outcome, completion_event, and evidence_ref on the citing local need. Only completed satisfies the dependency; abandoned and superseded keep the refusal visible.
Completion receipts now fail before they are written when their delivery result or evidence fields do not match the published contract. This keeps a malformed receipt from becoming the only surviving account of a pruned dependency.

[core][2.23.1] — 2026-09-03

2026-09-03T00:00:00Z

Closeout no longer offers to close an initiative that still has shaping or delivery-brief work open. Eligibility was read from the initiative's spec work alone, so an initiative whose every spec had shipped was recommended for closeout while its intents and draft delivery briefs were untouched. It now reports a named blocker instead, and stops recommending the skill that distils and dispositions.
Both record shapes now count as remaining shaping work. A shaping record written in the canonical form was invisible to the check, which read only the older shape — so an initiative whose remaining shaping work was written entirely in the current form looked empty. Retiring a record now stops it counting in both shapes alike, where before the two forms disagreed. A shipped, withdrawn, or cancelled delivery brief is still correctly treated as finished.

[core][2.23.0] — 2026-09-03

2026-09-03T00:00:00Z

The deferral-token convention now states the rule the tooling enforces. An open register entry needs an artifact of its own in Draft, unless it is a defect; it cannot point at the shipped document that recorded the deferral, because a shipped or accepted carrier and an open register membership cannot coexist. The previous wording directed authors to exactly that shape, so following it produced a record the tooling refuses to route.
The convention now warns that a deferral marker outlives its own record. A (deferred: <slug>) marker left in a body that later freezes pins its register entry open permanently, because the frozen body can no longer be edited to retire the marker.
Removed a reference to a repository-only test path. The clause pointed adopters at a file that exists only in this repository, and at a count ceiling that no longer exists.

[core][2.22.0] — 2026-09-02

2026-09-02T00:00:00Z

Closeout no longer offers to close an initiative whose remaining work the status surface says it cannot route. Deciding what a lifecycle record has retired now reads the same verdict the reconciliation layer reached, instead of rebuilding a guess from the entry's name. An entry recorded in an older shape is retired exactly when its artifact is; an entry the surface declines to route at all is never quietly treated as finished.

[core][2.21.0] — 2026-09-01

2026-09-01T00:00:00Z

An initiative whose work has all been retired now reaches closeout instead of reporting unfinished work forever. Whether anything is still outstanding is decided once, by a single pass that excludes what a lifecycle record has already retired — so two halves of the same answer can no longer disagree. An artifact named by one of a record's alternative locators counts the same as one named by its main locator.
When that reading is incomplete, the status surface withholds its suggestion to close out rather than making it anyway. A record that cannot be read, or whose review date cannot be judged, is reported as a named blocker instead, so a workflow that distils and disposes is never recommended on partial evidence. The prompt an agent renders reads the same facts, so the two agree.
Repair and migration are settled as unaffected by retirement, not left pending. Planning, applying, recovering and rolling back a repair or a migration behave identically whether or not lifecycle records are present, pinned by paired runs so a later change has to alter those lines and say why.

[core][2.20.1] — 2026-09-01

2026-09-01T00:00:00Z

Quality review now spots four tests that can stay green while proving nothing. It calls out empty-set assertions without a positive control, unstubbed negative cases, source-text checks that do not exercise a security control, and prose that substitutes for a captured baseline.

[core][2.20.0] — 2026-09-01

2026-09-01T00:00:00Z

Each reviewer now owns its concerns outright, so the same defect stops arriving twice under different severities. Adversarial owns contract conformance, scope and structural fit; quality owns whether a test can actually fail and the edge cases a spec never named; security owns every threat finding. A reviewer that notices another lens's concern says so in its own lens instead of emitting that lens's finding.
A clean review no longer pays for a model call to confirm it is clean. The loop classifies a persisted reviewer report structurally rather than comparing it byte-for-byte to one sentence, so an ordinary clean pass skips adjudication. A report carrying security-reviewer's mandatory coverage-disclosure footer still takes the full adjudication path, because that footer is prose and prose is what adjudication exists to read.
A naming quibble no longer costs the same as a correctness blocker. Nits do not block readiness and are never fixed automatically. They are recorded with their citation, deferred, and acted on only when the thread means to change the code because of one — and promoted first if that repair turns out to change behaviour, architecture, dependencies, or more than one file.
Security findings are now priced against a named attacker. A finding has to be reachable on a path this codebase runs and must not rest on a capability that already gives the attacker an easier route. This narrows severity and never coverage: an unpriced real weakness is still reported, and a trust-boundary crossing, missing authentication check, leaked secret, or spec-required control never drops on likelihood alone.

[core][2.19.0] — 2026-09-01

2026-09-01T00:00:00Z

Delivery briefs now distinguish withdrawal before execution from cancellation after execution starts. Open programmes remain Executing when their current slices have shipped but later slices are not materialized, and successful closure still needs an explicit Shipped decision.

[core][2.18.2] — 2026-09-01

2026-09-01T00:00:00Z

Re-issuing a review recording no longer risks counting the round twice. loop-cohort review record takes an optional --operation-id naming the round. Repeat the same recording under that id and the command reports a completed write and changes nothing; supply a different payload under it and the command refuses rather than silently accepting it. A session that dies before it learns whether the write landed can now simply repeat the command.

[core][2.18.0] — 2026-08-31

2026-08-31T00:00:00Z

`workspace-status` now tells you what closed-out work is waiting on you, and stops showing you work that is already finished. Orientation reports how many cooling reviews are due, names each one, and lists retention exceptions with the role that owns them and the date they are due. An artifact that has finished cooling is no longer offered as something to start, and is no longer opened, so a session begins on live work rather than on settled deliveries.
A retention exception stays visible, because someone still owes work against it. Only artifacts whose lifecycle record says Cooling or Retired are excluded; Retained and ExternalAdvisory records are projected and their artifacts remain available to start.
Orientation says whether the exclusion is complete. If a lifecycle record or the cooling module cannot be read, the run still projects the workspace and reports cooling_context_visible: true. That means the exclusion may be incomplete, not that it did not happen: one unreadable record leaves every other record still cooling its artifact, and the accompanying finding says which case you are in.
A brief whose declared child has cooled holds its dependants back. When a spec names a brief through source.parent in workspace.toml and that spec cools, the brief stops satisfying kind = "brief" dependencies until its child scope is re-established. Cooling is otherwise subtractive; this is the one place it withholds work, and it exists so a lifecycle record cannot make a blocked spec dispatchable by hiding a child's real state.
Repair and migration keep their previous behaviour and still reach cooled entries; whether cooling constrains them is a later decision.

[catalogue-curation][0.4.6] / [governance-extras][0.10.5] — 2026-08-31

2026-08-31T00:00:00Z

Assimilated primitives finish against the target catalogue's own release contract. Catalogue curators now close version, inventory, release-history, and projection duties instead of stopping after the content write.
New pack proposals reuse the target catalogue's RFC workflow. The proposal adds its primitive inventory without copying or drifting the canonical RFC template.

[core][2.17.3] — 2026-08-31

2026-08-31T00:00:00Z

The spec-metadata check now looks at the specs you touched, so finishing a work loop takes about two seconds instead of about thirty. It used to re-read all 423 specs and start a Git process for each one. Two checks still cover every spec regardless: the dangling-reference scan, and the one that resolves each deferral marker — because retiring a tracked follow-up can invalidate a marker in a spec you never opened. The remaining per-spec checks cover what you changed, and --all runs the full audit — wire that into a gate.

[core][2.17.2] — 2026-08-31

2026-08-31T00:00:00Z

When saving project knowledge runs out of time, it now says so and tells you to try again. It used to report a storage-capacity problem and advise fixing the request, and on the save path it could even claim another writer was mid-migration — so a run that simply needed more time looked like a request you had to rewrite, and the step that records lessons could not finish for anyone who did not read the exit code.

[agent-skill-engineering][0.4.0] — 2026-08-31

2026-08-31T00:00:00Z

The corpus now answers how a skill composes with subagents, hooks and packaged bundles, and keeps the portable question separate from any one tool's answer. Three portable floors were admitted — skills with subagents, hooks, and plugin packages. Each states the questions worth asking before reaching for a tool's syntax, and each says out loud that event names, configuration paths and output protocols belong to the tool rather than to the floor. Sixteen of thirty-six subjects are now carried and twenty remain recorded as absent.
A capability claim about a tool now carries a lifecycle state, so nobody reads "we checked" where the truth is "we read the manual". Each claim records its first-party source, the date it was retrieved, the version that source exposed, and one of four states. A claim confirmed by actually exercising the behaviour reads verified; a claim taken from documentation alone reads experimental and is not a support claim; a claim whose verification window has run out reads stale and returns provenance instead of guidance; an absent capability reads unavailable and is recorded as a difference rather than a gap.
Claude Code is the first tool profiled, and three of its seven capabilities are backed by an actual probe rather than a citation. Skill body loading, subagent context isolation and worktree isolation were exercised directly and read verified. Nesting limits, hook scoping, managed hook policy and package-supplied agent precedence are documented but unexercised, and read experimental on purpose.

[agent-skill-engineering][0.3.0] — 2026-08-31

2026-08-31T00:00:00Z

The reference corpus now answers language and execution questions, and says out loud which ecosystem and which versions each answer holds for. Five topics were admitted: Python with pytest, TypeScript/Node with JavaScript test runners, process and filesystem cost, pack and continuous-integration critical paths, and worktrees with state locks on a shared host. Twelve of thirty-six subjects are now carried and twenty-four remain recorded as absent.
Two answers that a single ecosystem's own documentation settles are now admissible. Earlier, a claim shipped only on two independent observations or two vendors agreeing. That bar silently excluded anything true of exactly one ecosystem — a language's own test runner, for instance. Such a claim now ships only when it names its ecosystem, an explicit version range with both bounds, and the vendor documentation it comes from.

[core][2.17.1] — 2026-08-31

2026-08-31T00:00:00Z

Exact clean reviews now finish without a redundant adjudicator call, while low-risk agent-context changes can remain in light mode. Mandatory automated reviewers no longer make work multi-person, and security review is tied to changed trust boundaries, data flows, or guarding controls.

[core][2.17.0] — 2026-08-31

2026-08-31T00:00:00Z

Work-loop now supplies reusable lessons through project-knowledge without having to assemble its storage and safety details.

[core][2.16.6] — 2026-08-31

2026-08-31T00:00:00Z

Reviewers now test the evidence behind a finding before they emit it. They retain uncertain concerns, clearly marking an untraced consequence instead of silently dropping it.

[architect][0.15.5] — 2026-08-30

2026-08-30T00:00:00Z

Architecture work reuses before it creates. architect-design looks for an adequate prior design or existing capability first and creates no new artifact when no real choice remains. A Stage-0 concept is a valid final answer, not a waypoint to a full design.
Architecture reviews challenge unnecessary surface. architect-review gained a design-doc reduction pass covering unnecessary components, services, dependencies, boundaries and custom mechanisms; ignored standard, native or provider capabilities; speculative scale and configurability; and complexity no named quality attribute supports.

[core][2.16.5] / [governance-extras][0.10.3] — 2026-08-30

2026-08-30T00:00:00Z

An RFC that should not exist no longer costs you an identifier and a file. new-rfc now decides what artifact the work actually needs before it resolves an ordinal or writes anything. When the answer is "skip", "an existing decision already covers this", or "an ADR, a spec, a PR, an issue, an architecture design, or a reversible trial fits better", it names the route and stops, leaving nothing behind.
RFCs get a review mode built for prose. The adversarial reviewer now has an RFC mode that needs no code diff, and treats the draft under review as untrusted data that cannot change its scope, severity, verdict, or routing.

[core][2.16.3] — 2026-08-30

2026-08-30T00:00:00Z

Planning-only sessions can now prove TDD acceptance criteria without leaving failing tests in the repository. The approved plan carries the executable stub, and implementation sessions materialize it only when code work begins.

[core][2.16.2] — 2026-08-30

2026-08-30T00:00:00Z

Spec review now separates original defects from faults introduced by a prior repair before another edit begins. Every completed report is independently adjudicated, and a green gate is used only for the behavior it actually proves.

[agent-skill-engineering][0.2.0] — 2026-08-30

2026-08-30T00:00:00Z

The authoring workflow now offers a `knowledge-provider` mode, and the corpus behind it will not answer a question it has no evidence for. Four topics were admitted out of seventeen candidates, because admission requires two observations at distinct skill paths in distinct packs and measured retrieval that selects the topic alone. The other thirteen are recorded as absent with the reason and what would admit them, so a reader can tell an unevidenced subject from an overlooked one.

[agentbundle][0.41.0] — 2026-08-30

2026-08-30T00:00:00Z

You can install a skill straight from a repository, without a catalogue. Point agentbundle install at a skill folder, a skills/ collection, or a single pack — locally or on GitHub — and it admits the source against explicit size and shape limits, pins it to the commit its bytes came from, and shows you exactly what the publisher declared before anything is written. A collection never installs everything by default: name the skills you want, or ask for all of them.
An upgrade that widens what a skill can do now stops and tells you what changed. Added tools, changed boundaries, new payload files, or a different credentialed status each need explicit acceptance, tied to the exact list you were shown.

[core][2.16.0] / [product-engineering][0.13.8] — 2026-08-29

2026-08-29T00:00:00Z

Authors can independently check an intent, delivery brief, or draft spec before construction starts. Core's shaping reviewer tests whether the contract is bounded and observable without replacing later adversarial, security, or quality code review.

[agentbundle][0.40.3] — 2026-08-29

2026-08-29T00:00:00Z

A read-only agent's declared boundaries are now checked where they are written, and never shipped into a projection that cannot use them. Adopters get the least-privilege posture their source declares on every supported adapter.

[core][2.15.5] — 2026-08-29

2026-08-29T00:00:00Z

`work-loop` can now correct a sealed spec or plan before the first wave runs. Review that finds a defect in the window between sealing the baseline and executing wave 1 previously had no legal move: contract-amendment refused until at least one task had completed, so the only way out was resetting the run and losing its identity and amendment history.

[core][2.15.4] — 2026-08-29

2026-08-29T00:00:00Z

An acceptance criterion that hides two contracts is now caught while it is being written, not three review rounds later. new-spec names the test — parts with separate failure modes and separate remedies are separate criteria — and fixes its boundary with five worked examples rather than an adjective an author can self-certify past.

[core][2.15.3] — 2026-08-29

2026-08-29T00:00:00Z

Spec authors can keep the contract and the plan independently sharp. An acceptance criterion now states one checkable obligation, and a plan's tests name the verification mechanism an implementer cannot infer instead of repeating the criterion.
A plan now tests its riskiest assumption before anyone reviews it. One cheap throwaway check can reshape the approach before reviewers spend rounds arguing about work nobody has started.

[agentbundle][0.40.2] — 2026-08-28

2026-08-28T00:00:00Z

Catalogue checks now admit the portable rule lookups shipped by the core pack while rejecting undeclared seed paths. Adopters can receive the new bounded rule router and scoped docs guidance through the normal seed flow.

[core][2.15.2] / [governance-extras][0.10.2] / [product-documentation][0.1.1] / [user-guide-diataxis][0.3.1] / [agent-skill-engineering][0.1.1] — 2026-08-28

2026-08-28T00:00:00Z

Agents now make chat, documentation, and skill-led work easier to scan without dropping checks, warnings, evidence, or depth. Repository rules, generated artifacts, and independent skills share the same outcome-first behavior without requiring a new host primitive.

[catalogue-curation][0.4.5] — 2026-08-28

2026-08-28T00:00:00Z

Compiled OKF skills now carry the same outcome-first output contract as hand-authored ones. A generated router or procedure skill reads the same way as the rest of the catalogue without a separate rule file.

[architect][0.15.4] / [experience-design][2.0.2] / [figma][0.3.3] / [product-engineering][0.13.7] / [product-strategy][0.2.4] — 2026-08-28

2026-08-28T00:00:00Z

Architecture, design, and product artifacts now lead with the decision and keep long work easy to resume. The output stays plain and scan-friendly while preserving requested depth, evidence, constraints, and exact terms.

[contracts][0.3.6] / [converters][0.9.6] / [frontend-engineering][0.2.1] / [iac-terraform][0.1.9] / [monorepo-extras][0.1.9] / [release-engineering][0.1.10] — 2026-08-28

2026-08-28T00:00:00Z

Engineering artifacts and handoffs now stay quiet during routine tool work while keeping exact code, errors, and verification. Results lead with the current outcome and name a next action only when one remains.

[atlassian][0.9.3] / [credential-brokers][0.3.3] / [desk-research][1.1.6] / [github][0.2.3] / [linear][0.3.3] — 2026-08-28

2026-08-28T00:00:00Z

Connected-service and research workflows now ask for less at once and return the current result without routine tool narration. Safety issues, blockers, needed choices, and host requirements still interrupt silence.

[core][2.15.1] — 2026-08-28

2026-08-28T00:00:00Z

Work-loop tools now reject execution state outside the current repository and finish safely when local Git stops responding.

[core][2.14.0] — 2026-08-28

2026-08-28T00:00:00Z

Agents now cut unnecessary work before adding code or process. Core's portable Razor ladder stops at the first sufficient option—skip, reuse, standard library, platform, installed dependency, one line, or the minimum working code—while preserving security, validation, accessibility, and explicit requirements.
Repository work now enters through one clear intent and delivery path. Neutral intake routes to intake-intent for a minimum repo-owned intent, author-delivery-brief create|continue for multi-spec coordination, or new-spec for one independently shippable feature.

[core][2.13.0] — 2026-08-27

2026-08-27T00:00:00Z

Delivery work can now close without losing the reasons and promises that code and tests cannot explain. close-work verifies repository-specific durable outputs, whole-document semantic freshness, obligations, dependencies, and authority before recommending a disposition.
Immediate disposal is safe by construction and never automatic. Every file removal needs a fresh human confirmation bound to the exact current target, fingerprint, disposition, evidence, authority, and session; any drift expires it.

[agent-skill-engineering][0.1.0] — 2026-08-27

2026-08-27T00:00:00Z

Agents can now frame, create, update, review, and optimize portable agent skills through progressive, least-authority workflows. The foundation includes activation and behavior evaluations, deterministic compiled reference knowledge, and bounded failure contracts.
Compatible organization or framework knowledge can inform skill work without coupling packs to another corpus. Workflows detect declared provider capabilities, invoke an eligible provider explicitly, and retain their complete baseline when no provider is available.

[catalogue-curation][0.4.4] — 2026-08-27

2026-08-27T00:00:00Z

Pack-owned OKF routers can now expose a bounded, explicit-only knowledge provider capability for compatible workflows. Raw OKF remains governed build-time input, while generated provider metadata supports deterministic discovery without implicit corpus traversal.

[core][2.12.5] — 2026-08-27

2026-08-27T00:00:00Z

Newly sectionless specs say why without retroactively gating adopters. The work-loop linter requires an explicit reason when a new spec omits its Acceptance-Criteria section or an existing spec removes that section, while grandfathering sectionless specs already present in an adopter's base ref.

[core][2.12.4] — 2026-08-27

2026-08-27T00:00:00Z

Maintainers can identify every unsupported legacy workspace record from a cold status read. Findings now retain a safe object slug instead of collapsing unrelated records under workspace.toml.
Deferring an acceptance criterion no longer forces a legacy workspace record. A (deferred: …) anchor now resolves from a canonical {path, kind} entry as well as a legacy {slug} one, so a new deferral can be recorded in the canonical shape.

[catalogue-curation][0.4.3] — 2026-08-26

2026-08-26T00:00:00Z

Generated OKF routers now keep hostile concept metadata inside its intended index entry. Titles, statuses, and concept types are bounded and escaped before they reach compiler-owned Markdown indexes, preventing metadata from fabricating links or additional entries.

[architect][0.15.3] — 2026-08-26

2026-08-26T00:00:00Z

The shipped architect pack is discoverable again. Its architecture-lenses source now declares the required content licence, so agentbundle show architect resolves it, and it accurately describes authored source versus generated router projections.

[core][2.12.3] — 2026-08-26

2026-08-26T00:00:00Z

Adjudicators can settle a cited repository-file claim without an avoidable stop. A finding may now point to the current, confined file needed to test its own predicate, while review artifacts and all other scope boundaries remain excluded.

[agentbundle][0.40.0] — 2026-08-25

2026-08-25T00:00:00Z

Catalogue maintainers can now build portable Agent Plugins 1.0.0 packages offline. The normal build emits deterministic, schema-valid root manifests and canonical skill trees for every skills-only pack, with complete reasons for packs the route excludes.

[catalogue-curation][0.4.2] — 2026-08-25

2026-08-25T00:00:00Z

Catalogue maintainers now have an end-to-end guide for shipping pack-owned OKF knowledge. The documented path covers declaration, canonical concept authoring, deterministic compilation, drift and ownership recovery, normal catalogue verification, and inspection through agentbundle show.

[core][2.12.2] — 2026-08-25

2026-08-25T00:00:00Z

Machine-checkable reviewer claims can now be adjudicated without giving the adjudicator execution tools. The work loop may supply evidence from one predeclared, confined read-only gate; the attempt consumes the existing review retry budget and produces a complete independent replacement report.
A bad proposed fix no longer obscures a real defect. The adjudicator now distinguishes an ineffective remedy from an effective but over-broad one while staying inside the supplied finding and repository-established repair seams.

[core][2.12.1] / [architect][0.15.2] / [governance-extras][0.10.1] / [monorepo-extras][0.1.8] / [iac-terraform][0.1.8] — 2026-08-24

2026-08-24T00:00:00Z

Architecture designs, current-state architecture, and decision records can now follow each adopter's own repository or external destinations. Explicit destinations, policy, established conventions, and optional configuration are resolved by semantic role; catalogue paths remain fallback offers.
Architect stays useful as an independent user pack. It can remain chat-only, save to an exactly confirmed personal workspace, consume Core's repository resolution when compatible, or produce a truthful repository handoff when Core is absent.

[agentbundle][0.39.4] — 2026-08-24

2026-08-24T00:00:00Z

Codex agents with read-only source permissions can read and search local files again. The design reviewer, both desk-research agents, experience reviewer, frontend reviewer, and both discovery reviewers retain Codex's command-backed file access inside a read-only sandbox without gaining write authority.

[core][2.11.0] — 2026-08-24

2026-08-24T00:00:00Z

Review loops now verify reviewer findings before asking you to repair them. An independent, read-only adjudicator checks each claim against the current target and governing rules. False positives stay in an audit instead of consuming repair rounds, while missing evidence stops for your decision.

[core][2.10.10] / [product-engineering][0.13.5] — 2026-08-24

2026-08-24T00:00:00Z

A confirmed product discovery can now hand bounded delivery context to Core intake without losing its source or skipping delivery approvals. One independently shippable feature enters as a delivery contract; multi-spec or cross-repository work enters as a delivery brief. Older or absent Core installations receive the same content as a portable rendered handoff.

[core][2.12.0] — 2026-08-23

2026-08-23T00:00:00Z

Work-loop reviews now explain readiness without asking you to trust a generic “clean” verdict. Reviewers trace non-local impact and rollout safety when warranted, then hand off an evidence-bearing verdict — BLOCKED, CHANGES_REQUIRED, READY_WITH_RESIDUAL_RISK, or READY — so every residual risk is named, not hidden.

[core][2.10.9] — 2026-08-23

2026-08-23T00:00:00Z

Core workflows can resolve a semantic destination using each repository's own paths, policy, established conventions, optional configuration, or external locator. Results preserve bounded provenance and independent authority facts, while unsafe local paths and ambiguous evidence fail closed.

[core][2.10.8] — 2026-08-23

2026-08-23T00:00:00Z

You can migrate a legacy workspace entry through a reviewed, reversible path. workspace-status shows the exact retained representation and candidate routes, plans only from a human-authored selection, applies only with fresh repository-authorized confirmation, and records a durable ledger that can restore the old workspace bytes without deleting the canonical artifact.
Starting, deferring, checking, refreshing, and migrating work now follow one documented route. The shared guides distinguish source evidence, canonical artifacts, lifecycle membership, processors, local refresh decisions, and separately confirmed tracker coordination.

[agentbundle][0.39.3] — 2026-08-23

2026-08-23T00:00:00Z

Installed workspace-status engines now understand reviewed legacy work-intake migrations. The bundled engine recognizes exact legacy workspace slices, validates human-selected target routes and durable migration ledgers, and keeps planning deterministic and non-mutating. Repository status refuses linked or aliased workspace input before projecting legacy bytes, and ledger validation rejects any operation whose immutable material no longer matches its digest.

[core][2.10.7] — 2026-08-23

2026-08-23T00:00:00Z

Every successful direct core install now tells you how to reach the read-only project-readiness check. The instruction appears for repository and local scope even when a lifecycle hook is disabled, unavailable, or not trusted, and it tells you to start a new session if the skill is not yet discoverable.

[agentbundle][0.39.2] — 2026-08-23

2026-08-23T00:00:00Z

Pack installers can now provide a useful manual next step without forcing a first-run prompt. Level A packs may add an optional Next: instruction, giving adopters a deterministic handoff when runtime hooks do not execute.

[core][2.10.6] — 2026-08-23

2026-08-23T00:00:00Z

Core now adapts to a repository's own development context instead of expecting the repository to adopt the pack's document layout. The doctor finds existing architecture, contributor, command, and scoped guidance, labels inference honestly, and proposes only approved root or subtree deltas.

[architect][0.15.1] — 2026-08-23

2026-08-23T00:00:00Z

Architecture design now starts from the repository's existing sources and implementation idioms, wherever they live. A custom DESIGN.md, scoped guidance, or repository-owned framework primitive can steer the design without first being copied into docs/architecture/reference.md.

[catalogue-curation][0.4.1] — 2026-08-21

2026-08-21T00:00:00Z

Nested OKF corpora now compile and check consistently on Windows. Generated index paths use the bundle’s portable POSIX form instead of inheriting the host path separator, so Windows no longer reports clean nested indexes as stale.

[architect][0.15.0] — 2026-08-21

2026-08-21T00:00:00Z

You can now turn a broad “assess this repository’s architecture” request into a correctable current-state model, an evidence-backed attention map, focused drill-downs, and a sequenced action plan. The assessment starts with a read-only survey, lets you correct its map before deeper investigation, and can progressively examine assurance, optimization, growth, transformation, or disposition without confusing repository conventions with architectural evidence.

[core][2.10.5] — 2026-08-21

2026-08-21T00:00:00Z

Work-loop commands now run from the project root without path discovery. The skill resolves its own installed directory before invoking the loop engine, cohort, or freshness scripts, so Codex and Claude Code no longer lose a turn searching for a repository-root scripts/ directory that does not exist.

[agentbundle][0.39.0] — 2026-08-21

2026-08-21T00:00:00Z

Catalogue builders can now reason about package formats independently of runtime adapters. The installed CLI exposes a validated contract for the existing APM and Claude-plugin routes, including their layouts, component support, admission, marketplace, and lifecycle behavior.

[core][2.10.4] — 2026-08-20

2026-08-20T00:00:00Z

The changelog rule now says who "user-visible" means, so you can tell whether your change needs an entry. An entry is required when a pull request bumps a released artifact's version — a pack or a published package — and repository tooling that ships in no release needs none. The old wording asked for an entry on "any user-visible behavior change", which read as covering maintainer-only tooling and offered no heading to write it under.

[governance-extras][0.9.7] — 2026-08-16

2026-08-16T00:00:00Z

Writing an RFC or an ADR now files its reusable lessons for you, and only once the decision is actually settled. Supporting lessons are handed to project knowledge at clean handoff and accepted-decision points, so a draft you abandon leaves nothing behind. If project knowledge is not installed, authoring proceeds unchanged and writes no stand-in file.