diff --git a/CHANGELOG.md b/CHANGELOG.md index b7095f0c..0d3eb48d 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -5,6 +5,12 @@ ## [Unreleased] +## [0.43.0] - 2026-11-19 + +Results that say where they ran, schema fields that can refuse an unknown value, and a YAML parser that misreads less. A recorded result now names its environment, and a failure in one environment is no longer hidden by a newer pass in another. A schema field can make a value outside its allowed set fail validation, which is how pulseengine.eu's North Star `serves:` link is pinned per project. Fourteen YAML fixes, found by sieve, the yaml-test-suite and the org's own files, stop silent misreads and needless refusals. + +This release is wrong if `rivet validate` passes a project whose schema sets `allowed-values-severity: error` on a field while an artifact holds a value outside its `allowed-values`, as a scalar or as a list item; if `import-results --environment` writes two environments' results to one file; or if a yaml-test-suite case at the pinned commit that this release counts as matched parses to a different value. + ### Added - **A schema field can make an unknown value fail validation, and list values are checked** (REQ-418). `allowed-values-severity: error` on a field turns a value outside `allowed-values` from a warning into an error. A field holding a list, such as `serves: [O1, O9]`, is now checked item by item; before, a list was never compared with the allowed set. This carries pulseengine.eu's North Star `serves:` link: a project schema adds `serves` to `feature` with the objective ids it accepts (see docs/schemas.md). @@ -14,7 +20,7 @@ ### Fixed - **YAML: a line holding only spaces between block entries is blank** (#1289). A whitespace-only line whose spaces reached the entries' indentation failed with `expected mapping key, found Some(Newline)`; PyYAML, yaml-rust2 and GitHub accept it, and two of kiln's workflows have it. Shallower, deeper and tab-holding blank lines between entries, at the root and in sequences, now read as blank too (yaml-test-suite H2RW, DK95/04 and DK95/05 now match). A tab where a block scalar's indentation goes is still refused (Y79Y/000). - **`import-results --environment` refuses a name that would collide** (REQ-417). The name goes into the results file name; characters other than letters, digits, `-` and `_` used to be replaced by `-`, so `hil bench` and `hil-bench` wrote the same file and the second import overwrote the first. Such names are now refused. -- **JUnit import keeps the whole failure body** (REQ-009). A failure or error body was cut at its first entity: `expected a < b` was recorded as `expected a`, because quick-xml reports an entity as its own event and only the first text event was kept, and per-event trimming dropped the spaces around it. The body now collects text, entity and CDATA events and is trimmed once when the element ends. Found while moving to quick-xml 0.42. +- **JUnit import keeps the whole failure body** (#1291; extends REQ-009). A failure or error body was cut at its first entity: `expected a < b` was recorded as `expected a`, because quick-xml reports an entity as its own event and only the first text event was kept, and per-event trimming dropped the spaces around it. The body now collects text, entity and CDATA events and is trimmed once when the element ends. Found while moving to quick-xml 0.42. - **YAML: text after a closing quote is an error, not dropped** (#1246). `k: 'abc' def` read as `abc`, and `k: 'it's'` as `it`, with no diagnostic. A quoted value must now end its line, apart from spaces and a comment. - **YAML: a key at a column no mapping uses is an error, not moved** (#1247). A line dedented between two levels (`a:` / ` b: 1` / ` c: 2`) joined the outer mapping, and one indented past its siblings joined theirs, so a field moved to another parent with no diagnostic. The error is reported once per misplaced block. This also catches a more-indented `key: value` after a scalar line (#1249, first case). In every local pulseengine checkout the set of files that fail to load is unchanged; synth's already-broken files report fewer errors. - **YAML: a tab indenting a block entry is an error** (#1248). `\ta: 1` and `a:` / `\tb: 1` were accepted with the tab counted as one column, so the structure could differ from what an editor shows. A tab before a flow collection or a scalar on a value's next line stays allowed, as YAML specifies. No local pulseengine checkout uses tab indentation. @@ -27,9 +33,20 @@ - **YAML: a line cannot continue a quoted value** (#1259). `k: "q"` followed by an indented `more` read as `"q"` and dropped `more`; it is now an error, at the root, under a key and in a sequence item. - **YAML: JSON-style flow mappings, root block scalars and root plain scalars that continue at column 0 are read** (#1259). `{"a":1}` (no space after a quoted key's `:`), `--- |` and `plain\nspans` were refused. `["a":1]` read as `["a", ":1"]`; it is now an error until single-pair flow entries are supported. - **YAML: invalid block scalar headers are errors** (#1258). `|0`, `|10`, `|--` and `|#c` were read without a diagnostic. -- **Recorded results: a newer pass on one environment no longer hides a failure on another** (#1241). Readiness, `trace-results` and the results view kept one result per artifact across all runs, so a test failing on `rpi5` and passing later on `macos-arm64` read as passing, also under `release.require: results`. The latest result is now kept per `run.environment`; results files without one read as before. `rivet import-results --format junit --environment ` records where the tests ran. Declaring which environments a requirement must pass in is still open. +- **Recorded results: a newer pass on one environment no longer hides a failure on another** (REQ-417, #1241). Readiness, `trace-results` and the results view kept one result per artifact across all runs, so a test failing on `rpi5` and passing later on `macos-arm64` read as passing, also under `release.require: results`. The latest result is now kept per `run.environment`; results files without one read as before. `rivet import-results --format junit --environment ` records where the tests ran. Declaring which environments a requirement must pass in is still open. - **YAML: plain scalars inside flow collections fold across lines** (#1252). `[a\n b]` read as `"a b"`, and in `{ t: hello\n world }` the second line became a key `world`; they now read `"a b"` and `{t: "hello world"}`. +### Security +- **wasmtime 48.0.9** (GHSA-5m8g-xhq4-4h93, #1280). wasmtime is built only with the `wasm` feature, which the release binaries do not enable, so the published binaries never contained the affected code. +- **chrono's RUSTSEC-2026-0543 (unmaintained) is recorded** in the cargo-deny configuration (#1276). + +### Dependencies +- quick-xml 0.42 (#1291; names and text are `&str`), lsp-server 0.10 (#1292; `Response` carries one `response_result`), and minor bumps of thiserror, regex, serde, log, petgraph and the GitHub Actions in use (#1281, #1262, #1238, #1265, #1271, #1273, #1274, #1234, #1264). + +### CI +- **No fork pull request reaches a self-hosted runner** (#1287). Every self-hosted job uses a fail-closed same-repository guard, and `tools/ci/check-selfhosted-fork-guard.rs` enforces it in YAML Lint, with a planted unguarded site as its negative control. +- **GitHub-hosted jobs run on `ubuntu-24.04`**, not `ubuntu-latest` (#1283), and the compliance action is pinned to a commit, not `@main` (#1243). + ## [0.42.2] - 2026-11-13 The first published release of this work. The v0.42.0 and v0.42.1 tags exist but nothing was published under either: no GitHub Release, no binaries, no npm or Marketplace version. The v0.42.0 run failed at VSIX packaging; the v0.42.1 run was cancelled because GitHub no longer accepts the `macos-14` runner. diff --git a/artifacts/requirements.yaml b/artifacts/requirements.yaml index 49fa95c3..a441ad33 100644 --- a/artifacts/requirements.yaml +++ b/artifacts/requirements.yaml @@ -1469,7 +1469,7 @@ artifacts: created-by: ai-assisted model: claude-opus-4-7 timestamp: 2026-04-22T20:29:48Z - release: v0.43.0 + release: v0.44.0 - id: REQ-084 type: requirement @@ -7832,7 +7832,7 @@ artifacts: created-by: ai-assisted model: claude-opus-4-8 timestamp: 2026-07-01T00:00:00Z - release: v0.43.0 + release: v0.44.0 - id: REQ-248 type: requirement @@ -8443,7 +8443,7 @@ artifacts: description: "Every mutating command writes through `rivet-core/src/yaml_edit.rs`, a ~2000-line line-oriented editor built on indentation heuristics. Hand-rolled deliberately (it preserves comments, key order and human formatting, which a serde round-trip destroys in version-controlled files) — but it has produced four data-loss bugs, each found by a user rather than a test: #573 flow-style `fields:` maps, #618/#625 block-list orphaning, REQ-287 unquoted tag emission, and #806 multi-line values in `fields:`. All four violate one invariant: replacing a key must consume that key's ENTIRE on-disk extent and change nothing else. #806 is the proof that examples are not enough — `set_field` upholds the invariant via `field_block_end` + `splice` and its comment names the history, while the custom-fields loop 700 lines below did a single-line assignment and had no multi-line-value test at all. The property must be stated as EXACT equality: parse before, mutate, parse after must succeed, every path except the target must be byte-identical, and the target must equal the intended value. 'It parses' is only a precondition — #806 mode 1 produced a perfectly parseable file whose value silently became 'short first body line second body line', so a parse-only assertion would green on the data loss. Corpus axis is scalar STYLE, not field name: plain, single/double-quoted, folded, literal, |- / |+ / |2, empty value, flow map, flow seq, block seq, nested block map, body containing colon / hash / leading dash, internal blank line, anchors, CRLF, tabs, last-field-of-last-artifact, unicode. Mutations: every --set-* flag, --add-tag/--remove-tag, --where bulk, plus `rivet add` and the MCP/batch writers. Lands as a proptest in rivet-core (fast, shrinks, runs in CI) plus a thin CLI-level test for flag wiring. Deliberately does NOT include migrating the writer to the in-tree rowan-yaml CST: that is an architecture decision with real cost for the maintainer to weigh, and the harness is the right thing to have in place BEFORE such a migration so it can prove behavior is preserved." fields: corrective: true - release: v0.43.0 + release: v0.44.0 links: - type: traces-to target: REQ-287 @@ -8520,7 +8520,7 @@ artifacts: title: "Humans have no assisted way to author artifacts — YAML editing is the only path (#546)" status: draft description: "User-reported and the largest gap in the v0.34 customer-UI theme, missed in the first triage pass. rivet's mutation surface is excellent for agents (add, modify, link, batch, next-id, MCP) and its read surface is excellent for everyone (serve, query, sql, trace, export), but a human authoring or correcting an artifact has exactly one option: hand-edit YAML. The reporter's words are the requirement — the editing is cumbersome, and we are missing the UI for helping humans work with it. Note the irony this repo has just paid for: hand-editing is also the path with no guard rails, and the write path underneath has produced four data-loss bugs (#573, #618/#625, REQ-287, #806), so the humans doing it by hand are the ones least protected. DESIGN-FIRST — this must not go straight to implementation. At least three plausible shapes exist and they differ enormously in cost and in who they serve: (a) extend the vscode-rivet VSIX with form-based editing and completion, which meets people where the YAML already is but is limited to VS Code users and the webview has no htmx; (b) make `rivet serve` writable, which reaches any browser and reuses the existing dashboard but turns a read-only viewer into a mutating service with all the authz questions that implies; (c) a guided terminal flow (`rivet add --interactive`) which is cheap and composes with the existing CLI but does not address the reporter's likely mental model of a UI. A design decision artifact must land before any REQ is scoped for implementation, and it must state which of these it is NOT doing and why." - release: v0.43.0 + release: v0.44.0 links: - type: traces-to target: REQ-287 @@ -8796,7 +8796,7 @@ artifacts: title: "Externally-owned verification: a criterion only another project can run has no home but prose (#800)" status: proposed description: "A requirement can carry a verification criterion that only a DIFFERENT project can run. `rivet verify` already refuses to promote without evidence, which is the correct block, so the mechanism is present; what is missing is a way to say WHY it is unrun and WHO owns it, so the set is queryable rather than living in a description field. `external-anchor` is not it — that models the INBOUND boundary, requirements handed to us by a supplier, with `source-of-truth`, `expected-derived-types` and `received-status`. This is the outbound direction: evidence we are waiting on from someone else. Same family as REQ-313, which gave a declared-blocked acceptance criterion a machine-readable home instead of shouty caps inside the string, and as REQ-308, which is about `verify` advancing on the existence of evidence rather than its sufficiency. Deliberately NOT scheduled into a release yet: the three overlap enough that solving them piecemeal risks three near-identical mechanisms for 'this obligation is declared, owned elsewhere, and not yet discharged'. REQ-308 already needs clauses to become addressable, which is a schema migration across every artifact; this should be designed alongside it rather than bolted on first. Backlog pending that design decision." - release: v0.43.0 + release: v0.44.0 links: - type: traces-to target: REQ-313 @@ -8995,7 +8995,7 @@ artifacts: title: "strict-mode merge serialization turns a throughput cost into a correctness risk on the two files that always collide" status: proposed description: "Branch protection sets strict: true, so every merge marks every other open PR out of date and forces a rebase plus a full CI re-run. N queued PRs therefore cost N SEQUENTIAL CI cycles rather than one parallel pass. Measured over the 2026-09-06 to 09 drain: five PRs (#897, #903, #906, #908, #909) were landed one at a time, each requiring a rebase and a fresh cycle, and a single cycle has run past 75 minutes under label-partitioned saturation (REQ-317). The wall-clock cost is the smaller half. The real cost is that serialization MANUFACTURES conflict-resolution events, and this repository's two hottest files collide at the tail on every concurrent branch — requirements.yaml and rivet-cli/tests/cli_commands.rs. Three of those rebases hit mid-statement splits. Two were two artifacts sharing one provenance block, where the naive keep-both resolution produces VALID YAML that rivet validate passes while one artifact silently loses its provenance — no gate in the chain would catch it, and it was found only by checking for the field by name. One was a FALSE overlap where git aligned two different test function bodies because both open with the same line; surgery there yields code that COMPILES and is wrong, and the branch had to be rebuilt from main instead. So the policy converts a throughput problem into a correctness risk, at a rate proportional to how many PRs are in flight. Candidate fix: a GitHub merge queue, which batches candidates and tests them against the prospective merge result without forcing a manual rebase per PR — keeping the guarantee strict mode exists to provide while removing the per-merge rebase. Alternatives worth weighing: relaxing strict and relying on the required CI Gate alone; or keeping the current policy and formalising the one-PR-at-a-time discipline the autonomous loop already follows, which is the status quo made explicit and costs the same time. Whichever is chosen, the mid-statement conflict shapes deserve their own guard independently, since they are latent whenever two branches touch those files. THE FIX ALREADY SHIPS AND RIVET DOES NOT USE IT, found 2026-09-09. SourceLayout::PerId exists in rivet-core, and its own doc comment states the case: one file per artifact id in the source directory, so two PRs adding different artifacts touch different files and never conflict. The SingleFile variant is documented as the historical default where two parallel adds collide on the file's tail. `rivet shard ` performs the migration, verifies the per-id files load identically before removing the original, and is reversible. rivet's own rivet.yaml sets no layout on any source, so every source is SingleFile, and artifacts/requirements.yaml carries 301 artifacts in one file — which is precisely why five rebases in one drain collided at its tail and why two of them silently dropped a provenance block. This reframes the requirement: a merge queue addresses the THROUGHPUT half by removing the per-merge rebase, but per-id layout removes the CORRECTNESS half outright, because the conflicts stop existing rather than being resolved more carefully. They are complementary and per-id is the cheaper and sharper of the two. Applying it to this repository is dogfooding a shipped feature its author never adopted, and would also remove the collision on artifacts/requirements.yaml that has been the single most common source of manual conflict resolution in this loop. SCOPE DECISION 2026-09-09: deferred from v0.37.0 to v0.38.0. The requirement identifies per-id source layout as the structural fix, and adopting it means migrating a single 311-artifact file into one file per artifact — a repository-wide restructure of the project's own source of truth. That is not work to run inside a release cut, where a mistake in the migration is indistinguishable from a mistake in the release. The cost of waiting is real and should be stated rather than assumed away: three tail collisions occurred in the 2026-09-09 session alone, and one of them placed a single shared `provenance:` block after the conflict's closing marker so that the naive keep-both resolution produced VALID YAML which `rivet validate` and `yamllint` both pass while REQ-339 silently lost its provenance. It was caught only by checking the field by name and reconstructing the block from origin/main. Every further cycle without the migration is another opportunity to lose data in a way no gate in the chain detects, so this should be early in v0.38.0 rather than late." - release: v0.43.0 + release: v0.44.0 tags: [ci, process, merge-queue] links: - type: traces-to @@ -9009,7 +9009,7 @@ artifacts: title: "No visual review-and-approval surface for a PR or release, and no typed record of who approved what at which commit" status: proposed description: "Maintainer request, 2026-09-09, thinking-out-loud captured before design hardens. Wanted: for a PR or a release, see the differences visually in a browser or the VS Code extension, leave approval or comments there and hand it back; a GitHub Action linking to the plugin so a reviewer can see it and get status; the view must invalidate when the commits change; and it should form part of the release view, with a notion of approving in a specific ROLE. Hard constraint: there is no server, so this must run locally or in CI. SURVEY FIRST — most pieces already exist and this is mostly assembly. `rivet check review-signoff --role ` is implemented, not a future oracle: it fires when a released artifact lacks a reviewer distinct from its author, optionally matching a role. The aspice schema already carries a review-roles context and per-oracle reviewers lists with change-control: pr-review. `rivet diff --base --head --format json` already computes what changed between two artifact trees. `rivet export --format html` already renders a static site with no server. serve has views/api/components, and the VS Code extension has a webview shell. ASPICE PAM v4.1 supplies the vocabulary rather than us inventing it: SPL.2 BP5 requires release approval before delivery, information item 13-13 is Product release approval, 18-06 is Product release criteria (conditions to be satisfied prior to authorizing delivery), and 11-03 lists approval for delivery by responsible roles as release-note content. GP 2.1.7 and GP 2.2.4 are the reviewer-not-author rule already enforced. PROPOSED SHAPE. The store is git and approval is an ARTIFACT, not a server record — it survives the PR, lands in the trace, and works offline. The approval names the exact commit SHA it approved, which makes staleness mechanical and identical to the pattern already shipped three times (REQ-309 exempt-when-field, REQ-320 unmodelled-rules, REQ-331 externals fallback): a declaration that stops being true must FAIL rather than go quiet. If HEAD moves past the approved SHA the approval is stale and the gate reopens, which is exactly the maintainer's requirement that changes must invalidate the view. CI materialises both sides with a git worktree, runs `rivet diff`, renders with `rivet export --format html`, publishes it as a workflow artifact or Pages, and posts a check-run linking to it; the extension renders the same static output locally. OPEN QUESTION, and the reason this is not yet a plan: with no server a browser cannot write to the repository, so 'put my approval in there and hand it back' has two honest routes and they differ in more than plumbing. Either the extension writes the approval artifact locally and the human commits it — traceable and offline, but read-only in a browser — or a GitHub PR review is the input and a CI job translates an approving review into an approval artifact recording reviewer identity and SHA, which works in a browser and matches the hand-it-back flow but makes GitHub the authority for who approved. The second is likely what was meant; it should be a recorded decision rather than an implementation detail, because it determines whether the approval record is verifiable without trusting GitHub. KNOWN CONSTRAINT: the extension webview loads only mermaid and a shell shim, so htmx attributes are inert there — any interactivity must go through the shim, and one anchor can serve both browser and extension. `rivet diff` takes directories rather than git refs, so CI has to materialise the base side itself. CORRECTION, maintainer, 2026-09-09, before any code: writing the approval as an artifact INTO the branch under review is self-defeating. If developer A opens a PR and reviewers B, C and D must commit their approval artifacts to A's branch, then B, C and D become contributors to that PR — and a contributor cannot independently approve it. GitHub blocks the self-approval, and GP 2.1.7 and GP 2.2.4 require reviewer identity distinct from author identity. Committing the approval destroys exactly the independence the approval exists to evidence. Reviewers also typically lack push access to the author's branch, so the flow is not merely wrong but usually impossible. The paradox dissolves once two DIFFERENT approvals are separated, which the first draft conflated. FIRST, peer review of a change: GP 2.1.7 / GP 2.2.4, reviewer not author, happens BEFORE merge and gates the merge. GitHub already implements this and rivet must not re-implement it — rivet's job is to RECORD it as evidence, never to collect it from inside the branch. SECOND, release approval: SPL.2 BP5, information item 13-13, approval for delivery by responsible roles, happens before DELIVERY rather than before merge, and applies to an assembled release of already-merged content. That one is safely an artifact committed on the default branch, because the approver is not the author of the thing being approved and no PR independence is at stake. So: release approval is authored directly as an artifact and gates the release. Peer review is EXTRACTED from the PR after merge by CI — reviewer login, review id, timestamp and the exact reviewed SHA — and written to the default branch by automation, so no reviewer ever commits to the branch they are reviewing. Consequence to accept honestly: an extracted peer-review record cannot gate its own merge, since it does not exist until after that merge. It is evidence, not a gate at that point. The gate that remains available is on the default branch — every merged commit touching traced paths must carry an extracted review record — which catches an extraction that silently failed rather than pretending the record blocked the merge. VERIFIABILITY, which is the residual cost of this route: a bot-written record saying B approved SHA X is only as trustworthy as the bot unless it stores enough to be re-checked — review id, reviewer login, timestamp, commit SHA — and even then it is auditable only while GitHub retains the review, and PR reviews can be dismissed. Anyone needing an approval record that is verifiable without trusting GitHub needs signed approvals, which is a larger decision and is not in scope here. Recording it so the limitation is chosen rather than discovered. SURFACES, maintainer 2026-09-09: the diff and the other review views are needed in the VS Code extension AND in the CLI, not only as rendered HTML, because an AI agent may be asked to perform a review and cannot read a webview. That makes this a CONSISTENCY requirement rather than a convenience one. The review view must be computed ONCE in rivet-core and rendered three ways — CLI text and JSON, extension webview, static HTML — because this repository already has a shipped instance of the opposite: #788, where plain coverage and coverage --tests computed the same evidence separately and disagreed by a hundred points, with nothing telling a release gate it was reading the narrower view. Three surfaces computing a review independently would reproduce that defect with higher stakes, since the thing they would disagree about is what a reviewer approved. The CLI rendering must therefore be a first-class consumer, emitting JSON an agent can act on, not a text pretty-printer bolted on afterwards. AI REVIEWER INDEPENDENCE, which follows directly and is not yet handled: if an agent performs the review, the record must capture WHICH agent — the model identity — and the reviewer-is-not-author rule has to be judged on that, not on a blanket label. rivet already stores provenance.created-by and provenance.model and already compares provenance.reviewed-by against provenance.created-by in `check review-signoff`, and common.yaml already carries an ai-generated-needs-review conditional rule. But that rule fires only when status equals active and its severity is warning, and the comparison is between identity strings, so an artifact created by ai-assisted and reviewed by ai-assisted is caught only because the labels happen to match — while the same model reviewing its own work under two different label spellings, or a second agent of the same model, would pass. For a repository that ships EU AI Act artifacts and stamps every artifact with a model name, an AI reviewing its own output is exactly the independence failure GP 2.1.7 exists to prevent, and the data to detect it is already stored. Scope note: this is adjacent to REQ-308, which is about evidence sufficiency rather than reviewer identity, and should not be folded into it. REVIEWER-AUTHORED APPROVAL PRs, maintainer 2026-09-09, and this supersedes the CI-extraction route above for peer review. Each reviewer commits their OWN approval as a separate pull request rather than into the branch under review, referencing the reviewed commit SHA. Independence is preserved because no reviewer touches the author's branch; the complete review assembles afterwards, and the maintainer has explicitly accepted that timing. The decisive advantage over CI extraction is verifiability: the reviewer's own commit — and its signature, since this project already signs tags with an ED25519 key — IS the approval evidence, so the record no longer depends on trusting a bot or on GitHub retaining a review that can be dismissed. That was the residual cost recorded above and this route removes it. It also needs no new conflict machinery, because SourceLayout::PerId already exists and is documented as one file per artifact id so two PRs adding different artifacts never conflict. Concurrent approvals from reviewers B, C and D therefore land as three separate files with no tail collision, which is the same mechanism REQ-332 identifies for requirements.yaml. An approval source should be per-id from the start rather than migrated later. SCHEMA NEEDED, as the maintainer put it — a way to approve an artifact and store it elsewhere. The type must carry at minimum: what is approved (artifact ids, or a commit SHA, or a release tag), the reviewer identity, the ROLE approved in, a timestamp, the exact reviewed SHA, and a verdict distinguishing approved from changes-requested. Stored in its own source path so approvals never share a file with the artifacts they approve. The SHA field is what makes staleness mechanical, per the invalidate-on-change requirement. AI REVIEWER INDEPENDENCE IS THE PRIMARY JUSTIFICATION, NOT AN ADJACENT ONE — maintainer, 2026-09-09: this machinery is needed even if the review is delegated to an AI, and more strongly so. That inverts the usual reading. With human reviewers, independence is self-evident and the record is bookkeeping: B is visibly not A, permanently. With an AI reviewer independence becomes UNOBSERVABLE WITHOUT A RECORD, because four situations that differ enormously all report identically as a review performed — a different model on fresh context, the same model on fresh context, the same model with the AUTHOR'S CONTEXT INHERITED, and the same session at a later turn. The last two are the author reviewing itself, and nothing in a rendered review distinguishes them from the second. The tooling makes the bad case one flag away: a forked subagent inherits the parent's full conversation by design, which is the natural reach when a reviewer should know what the work is about, and is precisely the inherited framing a clean-room review exists to exclude. So the fields this requirement already demands — reviewer identity, role, SHA, verdict — are necessary but not sufficient for an AI reviewer. The record must also carry MODEL IDENTITY rather than a generic ai-assisted label, since two agents of one model are not two reviewers, and an explicit statement of whether context was inherited plus a run identity distinct from the author's. A review lacking those is unrecorded rather than merely undocumented, which matters because this is consumed at a release gate. AND IT NEEDS A NEGATIVE CONTROL, which is the half with no analogue in human review: an AI reviewer that structurally cannot dissent yields independent-review-present forever, which is exactly the weak-green class v0.36.0 was themed around. Periodically hand the reviewer a known-bad artifact — a cited test that does not exist, a verified status with no verifying evidence — and require that it DISSENTS; record when that control was last exercised. A reviewer that passes the poisoned artifact is not a reviewer. The methodology half of this is filed upstream as pulseengine.eu#196; this requirement owns the schema half, and the schema cannot repair a practice that records the wrong thing. ASPICE grounding read from source: SUP.1.BP1 Note 1 carries the parenthetical (no self-monitoring), which is the standard's own one-phrase statement of the rule." - release: v0.43.0 + release: v0.44.0 tags: [review, release, approval, aspice, vsix] links: - type: traces-to @@ -9042,7 +9042,7 @@ artifacts: title: "A review surface that lists every check uniformly trains reviewers to tick blind — the proven and the unproven must be partitioned, and the partition must itself be auditable" status: proposed description: "Presentation half of REQ-333, from two grounded research passes commissioned 2026-09-09 after the maintainer left visual appearance open for the first version. THE MEASURED CASE, from the Automotive SPICE PAM v4.1 read from source: information item 13-19 Review evidence — the standard's own record of what a review produced — documents ten characteristics, and SEVEN of them are already proven mechanically by checks rivet ships today. Only three need a human: justification, non-conformances found, and improvement suggestions. Rendering all ten as a uniform checklist therefore asks a reviewer to attest to seven things a machine already decided, which is precisely how a reviewer learns to tick without reading, and it buries the three that carry the entire value of the review. This is the weak-green class of v0.36.0 relocated into the UI: a surface that reports more assurance than it has. THE PAM LICENSES HIDING THE PROVEN CHECKS IN ITS OWN WORDS, which matters because the instinct in a regulated context is to show everything defensively. Table B.1 states that it is not intended to use the listed characteristics as a checklist. GP 2.2.1 Note 2 states there may be types of work products for which no review or approval is required. And attached to every bidirectional-traceability base practice, recurring eighteen times, is the sentence that traceability alone, for example the existence of links, does not necessarily mean that the information is consistent — which is the standard itself distinguishing a mechanical link check from the judgement a reviewer owes. So an assessor-facing surface that SUPPRESSES what is proven is more standard-conformant than one that lists it, not less. REQUIRED SHAPE: partition rather than a flat list. Four buckets — what BLOCKS, what NEEDS JUDGEMENT, what is SETTLED MECHANICALLY, and what is NOT REVIEWABLE here. Settled and not-reviewable collapse by default and are counted rather than enumerated. THE PARTITION RULE IS THE LOAD-BEARING PART AND MUST BE AUDITABLE, because a surface that hides items is only safe if a reader can ask why a specific item was hidden and get the name of the check that settled it plus its result. An item may enter SETTLED only by naming the mechanical check that discharged it; anything without such a check is NEEDS JUDGEMENT by construction, so the failure direction is toward showing too much rather than hiding something unproven. Absent that rule this requirement would ship the very defect it exists to remove. THE HUMAN RESIDUE IS A SHORT AND STABLE VOCABULARY, which is what makes the judgement pane designable at all: across every information item examined it reduces to rationale or justification, adequacy, intended use, impact, root cause, summary, non-conformances found, and improvement suggestions, plus the four adjectives information item 17-00 demands of a requirement — unambiguous, testable, measurable, necessary. The judgement pane should therefore be organised around those words rather than around item identifiers, because the reviewer's task is the same shape regardless of which item raised it. DENSITY IS WILDLY ASYMMETRIC AND THAT ASYMMETRY IS THE LAYOUT: information item 13-16 Change request is nine of twelve human and wants a prose pane; 13-07 Problem is two of twelve; 13-08 Baseline has exactly one human element and it sits in the definition sentence rather than the bullets, so it wants a single line, not a form. A uniform per-type template would be wrong for all three. GROUNDING NOTE, recorded so a later reader does not assume the research was exhaustive: Annex C.4 Figures 9 and 10, the software and hardware consistency maps and the most on-point content in the PAM for a traceability tool, did NOT extract from the PDF and were deliberately not reconstructed from memory. DO NOT MECHANIZE information item 13-52 Communication evidence — its documented characteristics include chat protocols and verbal statements, so the pull-request conversation IS that item; link to it rather than building a form that would duplicate it worse." - release: v0.43.0 + release: v0.44.0 tags: [review, ui, aspice, weak-green] links: - type: refines @@ -9056,7 +9056,7 @@ artifacts: title: "rivet cannot answer what a change forces you to re-verify — the PAM requires re-verification criteria and coverage against release scope, and no shipped check computes either" status: proposed description: "Highest-value UNBUILT check found by the Automotive SPICE PAM v4.1 research pass, 2026-09-09, and filed separately from REQ-333 and REQ-335 because it is a core traceability capability rather than a review-surface concern. Two places in the PAM demand it. Information item 08-58 lists among its documented characteristics the criteria for re-verification, for example in the case of changes. And SWE.4.BP2 together with SWE.6.BP2 both require sufficient coverage according to the release scope — a scoped judgement, not a global percentage. rivet ships seven mechanical checks and none of them answers either question. The capability is squarely within what rivet already models and is arguably the sharpest thing it could compute that no adjacent tool can: GIVEN A CHANGE SET AND THE LINK GRAPH, DERIVE WHICH VERIFICATION MEASURES THE CHANGE INVALIDATES. Everything needed is present — the typed link predicates give reachability from changed artifacts to the tests and oracles that verify them, `rivet diff` already computes what changed between two artifact trees, and the release field already delimits scope. Today a reviewer or release manager infers that set by hand, which is guesswork that fails silently and in the unsafe direction, since an unrecognised invalidation looks exactly like a requirement that is still verified. Note the shape it shares with the rest of this release line: a declaration that has quietly stopped being true must FAIL rather than go quiet, the pattern already shipped in REQ-309, REQ-320 and REQ-331. A verified status whose supporting evidence predates a change to what it verifies is exactly such a declaration. Scope note: this is adjacent to REQ-308, which concerns whether evidence is SUFFICIENT, whereas this concerns whether existing evidence is still CURRENT with respect to a change; they should not be folded together. Design question to settle before implementation, and the reason this is proposed rather than planned: whether the invalidation set is computed from artifact-tree diffs alone, which rivet can do offline today, or must also incorporate source-file changes reached through the verification markers, which is more faithful to what actually invalidates a test result but requires the change set to span code as well as artifacts." - release: v0.43.0 + release: v0.44.0 tags: [coverage, verification, aspice, traceability] links: - type: traces-to @@ -9474,7 +9474,7 @@ artifacts: created-by: ai-assisted model: claude-opus-5 timestamp: 2026-09-15T19:29:56Z - release: v0.43.0 + release: v0.44.0 - id: REQ-361 type: requirement @@ -9749,7 +9749,7 @@ artifacts: created-by: ai-assisted model: claude-opus-5 timestamp: 2026-09-24T07:13:56Z - release: v0.43.0 + release: v0.44.0 - id: REQ-375 type: requirement @@ -9764,7 +9764,7 @@ artifacts: created-by: ai-assisted model: claude-opus-5 timestamp: 2026-09-24T15:33:28Z - release: v0.43.0 + release: v0.44.0 - id: REQ-376 type: requirement @@ -9803,13 +9803,13 @@ artifacts: created-by: ai-assisted model: claude-opus-5 timestamp: 2026-09-24T21:09:51Z - release: v0.43.0 + release: v0.44.0 - id: REQ-378 type: requirement title: 62 of 67 unsafe control actions name their control action in free text that resolves to nothing, and validate passes status: proposed - release: v0.43.0 + release: v0.44.0 description: "Found while testing REQ-340's grounding claim on rivet's own corpus, and a defect in its own right independent of any formalization work. STPA's analytical value depends on each unsafe control action being attributable to a SPECIFIC control action issued by a specific controller. In rivet that attribution is prose. Measured through the store with rivet list --format json --full, guarded by an assertion that fields are populated: of 67 unsafe control actions, 62 name their control action in FREE TEXT THAT RESOLVES TO NOTHING, and only 5 use a string that is an artifact id. UCA-C-1, for example, names its action as Build link graph, validate artifacts, generate metrics, a sentence that matches none of the 20 control-action artifacts that exist. rivet validate passes over all of it, because uca.control-action is declared as a plain string field rather than a link, so there is nothing for validation to resolve. The same holds one level up: all 8 controller process models are lists of English phrases such as Current state of local artifact files, with no ids, so a UCA's context can only reference the controller belief it invalidates by paraphrase. TWO CHANGES, in order. First, make the UCA-to-control-action relation a typed link to a control-action artifact, so an unresolvable reference becomes a broken-link error rather than silently passing, and migrate the 62 existing references, most of which will need a human to choose which of the 20 actions each one means. Second, give process-model entries ids so a UCA can cite the belief it is about. Both improve STPA rigour on their own merits, and both are the groundwork REQ-340 identified as the precondition for ever revisiting autoformalization. Note the migration is itself a grounding exercise of exactly the kind the paper describes: reconciling differently-worded references to the same thing, and it should be done by a person rather than by string similarity." tags: [stpa, traceability, weak-green, grounding] fields: @@ -9828,7 +9828,7 @@ artifacts: type: requirement title: no AI bill of materials — per-artifact model provenance exists but is never aggregated or emitted in a standard form status: proposed - release: v0.43.0 + release: v0.44.0 description: "Split out of REQ-344 by DD-079. Issue 104 asks for an AI bill of materials: which models were used, at which versions, and which artifacts and code each generated, emitted in a standard form such as CycloneDX ML-BOM. No type in any rivet schema models this. The nearest existing mechanism is per-artifact provenance, which records created-by, model, session-id and timestamp, and therefore records THAT a model was involved in one artifact, but it is not a bill of materials, it is not aggregated across a release, and it emits nothing in a standard format. DD-079 separated this from the supply-chain bridge deliberately: that half has a complete schema and evidence the release already produces, while this half has neither, and bundling them would hold the ready half hostage to the unready one. Scoped with REQ-059, AI provenance with model id and session context, since an AIBOM is the release-level aggregate of exactly that per-artifact provenance. The provenance model assessed for REQ-341, which separates generated-by from a sequence of verified-by events, is the relevant prior art: an AIBOM that cannot say which AI output was subsequently checked, and by whom, is a list of models rather than evidence." tags: [aibom, supply-chain, ai-provenance, eu-ai-act] fields: @@ -9852,7 +9852,7 @@ artifacts: type: requirement title: bounded, pre-registered autoformalization spike on rivet requirement pairs, with the fidelity check and negative control the paper lacked status: proposed - release: v0.43.0 + release: v0.44.0 description: "The bounded spike decided by DD-080 after the maintainer reviewed REQ-340's assessment. PURPOSE: first-hand evidence on whether autoformalizing rivet requirements into Lean yields anything a reviewer would not already see, to inform a later adopt-or-not decision. It is a spike: its output is a findings document and a measurement, not production code or a new command. PRE-REGISTERED, so the spike cannot grade its own homework. The procedure, the measurements and the success criterion are fixed here BEFORE any formalization is run. CORPUS: four to six PAIRS of rivet requirements chosen because they plausibly constrain the same behaviour, so that equivalence, implication or inconsistency is a meaningful question for each pair. Each pair's selection rationale is recorded before formalizing. PROCEDURE, the paper's pipeline plus the step it lacked. One, an LLM formalizes each requirement into a Lean 4 proposition using core Lean only, no Mathlib, matching the paper's encodings. Two, a human writes the grounding hypotheses equating differently-named symbols, and every one is counted. Three, attempt to prove equivalence or implication between the pair, recording proved, refuted by counterexample, or search failed; a failed search is NOT a disproof and must not be reported as one, which the paper did. Four, THE FIDELITY CHECK THE PAPER LACKED: back-translate each Lean proposition into English and have a human compare it with the original requirement, recording any mismatch. MEASURED PER PAIR: compiles or not; number of grounding hypotheses; proof outcome; fidelity-check outcome; whether the proof went through only because the grounding made the two sides syntactically identical, as in the paper's own showcase. NEGATIVE CONTROL, and the most important element: one pair is deliberately seeded with the exact error the paper shipped, a greater-or-equal formalized as a strict greater-than. The fidelity check MUST catch it. If it does not, the procedure cannot be trusted to catch real mis-formalizations and every other result is void. SUCCESS CRITERION, fixed now: the spike SUCCEEDS if it surfaces at least one equivalence, implication or inconsistency between real rivet requirements that a human reviewer confirms is genuine and was not already recorded, AND the negative control is caught. It FAILS USEFULLY if every proof is carried by the grounding, which would replicate the paper's result on rivet's corpus and settle the adoption question in the negative. Either outcome is a result. TIME-BOXED; if the time-box expires, whatever was measured is reported as it stands." tags: [research, formal-methods, spike, lean] fields: @@ -10003,7 +10003,7 @@ artifacts: type: requirement title: "gates that pass vacuously: the traceability gate cannot fail on a traceability regression, and the WASM render tests skip themselves (#957)" status: proposed - release: v0.43.0 + release: v0.44.0 description: "Umbrella from the weak-green audit. Items include: ci.yml runs rivet validate, which exits 0 on the real corpus because every traceability finding is warning or info, so appending an implemented requirement with no links still passes; the WASM adapter tests render_aadl_via_wasm and load_spar_wasm_component skip on a hardcoded path that exists nowhere and one swallows Err, so call_render never executes (confirmed by two MISSED mutants on a72637c), leaving the read-only preopen and the fuel limit untested. REQ-375, the oracle dispatcher, carries its own item. ACCEPTANCE: each item has a negative control that now reddens its gate; the WASM tests run a real component in CI and fail rather than skip when it is absent there; a probe guest demonstrates a refused write through the preopen and a trap on fuel exhaustion. PROGRESS, WASM item: the tests take the component only from SPAR_WASM_PATH (the hardcoded path is gone); with RIVET_REQUIRE_SPAR_WASM set a missing component fails them, and locally they print 'SKIPPED (proves nothing)'; render_aadl_via_wasm fails on a render error instead of printing it. The wasm-seam CI job builds spar-wasm at the Cargo.lock commit and runs the wasm_runtime tests with the requirement set; Cargo.lock now triggers that job. With the real component, a call_render that always errors fails the render test (it used to pass). Remaining: the probe guest demonstrating a refused write through the preopen and a fuel-exhaustion trap, and the traceability-gate item. A1 (maintainer decision 2026-10-09, step one of the traceability ratchet): traceability rules whose required backlink is verifies now credit a '// rivet: verifies ' source marker, the evidence rivet verify and release readiness already accept (REQ-329). LinkGraph carries the marker set apart from its links (with_marker_evidence), so coverage and the validator's rule check credit it wherever they run; the salsa validate path is filtered with validate::credit_marker_evidence; markers are scanned only by commands that compute coverage or verification (validate, coverage, context, supplier check, release readiness), at dashboard load and reload, and at MCP load, because a scan takes about 4 s in a debug build. On this repository requirement-verification went from 29/401 (7.2%) to 177/401 (44.1%), overall coverage from 50.5% to 59.4%, and warnings from 1,611 to 1,478; CLI, context, dashboard API and MCP agree on 177 (MCP and the API count 400, without the one external artifact, as before). Remaining for the gate: the ratchet itself. Ratchet (maintainer decision A): the CI traceability job runs tools/ci/traceability-ratchet.py on every pull request. A finding is a validate diagnostic at warning or error from a rule rivet coverage reports, keyed by rule and artifact, counted only while the artifact is at implemented, verified, accepted or released in that tree; the pull request fails when its head adds one its merge base lacks. tools/ci/traceability-ratchet_test.sh pins each filter with a negative control (removing the status, severity or rule-scope filter each turns a case red). Run over the last six commits of main it flags REQ-400, marked verified with no satisfies link." tags: [ci, testing, weak-green, wasm] fields: @@ -10064,7 +10064,7 @@ artifacts: type: requirement title: "self-hosted runner disk exhaustion fails every compile gate, and the post-job cleanup hook frees nothing (#567)" status: proposed - release: v0.43.0 + release: v0.44.0 description: "On the self-hosted pool a full disk makes every compile gate fail at once, and the post-job cleanup hook that exists for this reclaims no space. ACCEPTANCE: the hook's reclaim is measured with df before and after on a real runner, and it frees the per-job build output; a low-disk precondition fails a job early with a clear message rather than as a compile error; the fix holds across at least three consecutive runs on the same runner." tags: [ci, runners] fields: @@ -10080,7 +10080,7 @@ artifacts: type: requirement title: "Linux release binaries require glibc 2.39, which excludes Ubuntu 22.04, Debian 12, RHEL 9 and every static image; ship musl builds (#980)" status: proposed - release: v0.43.0 + release: v0.44.0 description: "Measured on v0.37.0's Linux x86_64 asset: dynamically linked, with GLIBC_2.39 as its highest symbol version. varve's pulseengine realm ingests these binaries, and a layer is only as portable as its highest floor, so rivet alone sets it for the whole toolchain. ACCEPTANCE: the release publishes statically linked musl builds for x86_64 and aarch64 Linux alongside the existing assets, listed in SHA256SUMS and covered by the cosign signature; the release verifier confirms each is static and runs rivet --version inside a musl or distroless container; the release notes name the asset to use on older distributions." tags: [release, portability, varve] fields: @@ -10095,7 +10095,7 @@ artifacts: type: requirement title: "no VEX: a consumer cannot tell whether a published advisory affects the rivet release they run (#983)" status: proposed - release: v0.43.0 + release: v0.44.0 description: "An SBOM says what is in a release, not what an advisory means for it. rivet already publishes a CycloneDX SBOM inside the signed sums, runs cargo-audit, and records not-applicable assessments as artifacts (REQ-376 for the salsa advisory). varve is adding ingestion of each payload's supply-chain documents. ACCEPTANCE: each release publishes a CycloneDX VEX document whose vulnerabilities carry analysis.state and a justification, generated from cargo-audit findings plus the recorded assessments, listed in SHA256SUMS and signed; an advisory published after a release can be answered by re-issuing that release's VEX without rebuilding it; an advisory with no recorded assessment is published as under investigation, never omitted." tags: [release, supply-chain, security, varve] fields: @@ -10144,7 +10144,7 @@ artifacts: type: requirement title: "the VS Code extension is published only to the Microsoft Marketplace, so VSCodium, Cursor and Windsurf users cannot install it from their registry" status: implemented - release: v0.43.0 + release: v0.44.0 description: "Found by the 2026-09-30 sphinx-needs comparison: useblocks' ubCode began publishing to Open VSX on 2026-09-28. rivet's release workflow publishes vscode-rivet to the Microsoft Marketplace only; no Open VSX reference exists in .github or the extension. ACCEPTANCE: the release workflow publishes the same VSIX to Open VSX under the pulseengine namespace, from the SHA256SUMS-covered artifact; a missing token is a visible warning, as for the Marketplace; the release verification confirms the version is listed on Open VSX after a release. The Open VSX account, namespace and token secret are the maintainer's to create. IMPLEMENTED in release.yml as publish-vsix-openvsx: it publishes the same VSIX artifact as the Marketplace job with ovsx, is continue-on-error like that job so it cannot block release-npm.yml, warns visibly when OVSX_PAT is not configured, and after publishing polls https://open-vsx.org/api/pulseengine/rivet-sdlc/ until it answers 200, failing the job otherwise. The URL scheme was checked against a listed extension (useblocks/ubcode 0.36.0 answers 200, a missing version 404). TRUSTED PUBLISHING (maintainer's choice, 2026-10-01): the job now runs tools/release/openvsx-publish.sh, which picks its mode explicitly: an OVSX_PAT secret publishes with the PAT; otherwise the repository variable OPENVSX_TRUSTED_PUBLISHING=true publishes with --trusted-publishing (ovsx 1.2.0, pinned), exchanging the job's OIDC token (permissions id-token: write, environment openvsx) for a five-minute single-extension token, so no long-lived secret is needed; with neither, the job warns and skips. A registration without the id-token permission is an error, not a skip. Open VSX registers trusted publishing only for an owned namespace and an extension with an active version, so the first version arrives by PAT or by a manual upload; then the maintainer registers pulseengine/rivet, workflow release.yml, environment openvsx, sets the variable, and removes the PAT. tools/release/openvsx-publish_test.sh covers the mode choice in CI; negative controls redden it. NOT YET VERIFIED: that needs the pulseengine namespace, a first version, the registration and a release run; the status stays implemented until the listing check passes on a real release." tags: [release, vscode, distribution] fields: @@ -10510,13 +10510,14 @@ artifacts: - id: REQ-417 type: requirement title: "recorded results keep the latest result per environment: a newer pass on one target hides a failure on another (#1241)" - status: implemented + status: verified description: |- A results file says where a run happened (run.environment), but ResultStore::latest_for kept one result per artifact across every run, and evidence_for, trace-results and the results view classified from it. TEST-1 failing on rpi5 at 10:00 and passing on macos-arm64 at 11:00 read as Passing, and under release.require: results (REQ-413) the requirement counted as ready although it never passed on the Pi. rivet import-results --format junit always wrote environment: None. ACCEPTANCE: evidence_for keeps the latest result per (artifact, environment), so a failure in any environment outweighs a pass in another; a run with no environment is one more environment, so existing results files read as before; a newer result in the same environment still replaces an older one. Views that show one status per artifact (trace-results, the results page) show the newest failing per-environment result if there is one. rivet import-results --format junit --environment records the environment and puts it in the run id, so two environments' imports of one JUnit file do not overwrite each other; --environment with a format that imports artifacts is refused. Out of scope here: declaring where a requirement must pass (a verify-on field or variant bindings) and failing readiness for a required environment with no result, which wait on that design choice. tags: [results, release, gate, environment] fields: + independent-review: "2026-11-19: acceptance met after #1290. Reviewer: a fresh-context subagent (claude-opus-5-5, new context, not a fork of the author's session; the same model family as the author, so not a different-model review). It read the source and tests and ran the named tests (all pass). It found one counterexample: `--environment \"hil bench\"` and `hil-bench` wrote the same results file, so the second import overwrote the first. #1290 refuses such names and adds a CLI test importing one JUnit file for two environments. Noted, not contradicting the acceptance: the results page itself shows per-run rows; the one-status-per-artifact view is the verification view." category: functional priority: must verification-criteria: "rivet-core unit tests: results.rs latest_results_are_kept_per_environment (one per environment, newest first, no environment as its own, failure and error win), result_trace.rs a_newer_pass_on_one_environment_does_not_hide_a_failure_on_another (evidence, trace status and verdict fail; a newer pass on the same environment clears it), junit.rs set_environment_records_it_and_keeps_run_ids_apart and set_environment_refuses_a_name_that_would_collide. CLI tests in rivet-cli/tests/import_results_environment.rs: two imports with --environment give two results files and a failing trace verdict; one JUnit file imported for two environments writes two results files, and a name that would collide (`hil bench`) is refused and writes nothing; --environment with reqif is refused." @@ -10529,17 +10530,19 @@ artifacts: created-by: ai-assisted model: claude-opus-5-5 timestamp: 2026-11-18T05:40:00Z + release: v0.43.0 - id: REQ-418 type: requirement title: "a schema field can make a value outside allowed-values fail validation, and a list field is checked item by item (North Star serves:)" - status: implemented + status: verified description: |- pulseengine.eu defines North Star objectives O1-O4 and wants a feature or requirement to declare `serves: O2`, checked against the ids each repo pins in its own schema. Project schemas could already add a `serves` field with `allowed-values` to `feature` (same-named types merge), but a value outside the set was only a warning, so `rivet validate` passed, and a list value such as `serves: [O1, O9]` was never compared at all. ACCEPTANCE: a schema field may set `allowed-values-severity` (info, warning or error); absent, a value outside `allowed-values` stays a warning as before. A field holding a sequence is checked item by item, in the default view and in variant overlays, and each item outside the set gets its own diagnostic at the field's severity. A missing optional field produces no diagnostic. With `allowed-values-severity: error`, a planted unknown objective id makes `rivet validate` exit non-zero. tags: [schema, validation, north-star] fields: + independent-review: "2026-11-19: acceptance met, clause by clause. Reviewer: a fresh-context subagent (claude-opus-5-5, new context, not a fork of the author's session; the same model family as the author, so not a different-model review). It read the source and tests and ran the 6 named tests (all pass). Untested but not contradicting the acceptance: the `info` severity and an end-to-end variant-overlay run through validate_variants." category: functional priority: must verification-criteria: "rivet-core unit tests in validate.rs: allowed_values_checks_each_item_of_a_sequence, allowed_values_severity_error_fails_an_unknown_value_only, variant_overlay_checks_sequence_items_at_the_declared_severity. CLI tests in rivet-cli/tests/allowed_values_severity.rs: a planted O9 (scalar and in a list) fails validate with an ERROR line; O2, [O1, O4] and no serves pass; without the severity key O9 stays a warning." @@ -10550,3 +10553,4 @@ artifacts: created-by: ai-assisted model: claude-opus-5-5 timestamp: 2026-11-18T11:30:00Z + release: v0.43.0