grid_assets 0.7.0-dev.2

The_grid's opinion assets: the agent/verify/land Capability impls, the code circuit, and the git SourceControl, extracted out of the opinion-free engine.

0.7.0-dev.2

• Breaking: the dependencies filing requirement is a PROJECTION of the dependency rows bd holds for the bead, and bead PROSE is never parsed for blockers again (Nico, 2026-09-13, under the_grid#the-grid-is-a-beads-controller; power_station#the-dependencies-row-is-a-projection-of-bd-dependency-rows). The declaration grammar — a segment opening with Blocked by / Blocked on / Depends on, its tokens read as ids by prefix or digit tail — and the declared-vs-wired comparison it fed are DELETED, with no compatibility arm. Blocked-by: pow-x and Blocked by pow-x now project identically, because neither is a row. New public surface: DependencyProjection, ExternalBlocker, ExternalResolution and noArmedSubstations. Migration: wire the row (bd dep add <blocked> <blocker>, or link for a cross-store capability) — a sentence never declared one.

• Breaking: FilingContract.evaluate, FilingService.inspect/check, ApproveService.approve, UnparkService.unpark and the filing/approve/unpark commands take armedSubstations — the station's roster by NAME, which an external:<project>:<capability> row's project resolves against. It is FAIL-CLOSED: a project the roster does not arm refuses the row, and so does a roster that was never supplied (noArmedSubstations, the default), each naming its own remedy. No station threads the roster yet, so every external: row refuses both verbs until one does — a composition gap, named in the refusal.

• Breaking: --state-root retires from filing, approve and unpark, and the stateRoot parameter is gone from FilingCommand, ApproveCommand and UnparkCommand. The option existed on those verbs for ONE reader — the grid home's cross-store link beads — and that read is gone (the_grid#447) with the dependencies row now reaching no second store at all; a verb that takes a root it never reads teaches an operator that the root matters to its answer. park and show keep the option and still reach the state store's session-lifecycle beads. Migration: drop stateRoot: from those three constructors and pass the station's roster as armedSubstations instead; drop --state-root from filing/approve/unpark invocations.

• Breaking: a standing grid.approved_rev over a bead whose description NAMED a blocker re-derives, because the basis dependencies member changed shape — from the ids the prose named with their proofs to the rows bd holds, typed local or external. Those receipts read STALE and their beads need re-approving; the governor sweeps in-flight work at the boot checklist. The basis PREFIX stays filing:v1:sha256: — a new prefix is a v2 receipt scheme and its own bead. The ROSTER is deliberately excluded from the basis: arming a substation must not revoke a governor's approval of a bead nobody edited, so the mount gate needs no roster.

• Changed: ExactSubstationBeadSource.readExact reads bd's RECORD surface (BdCliService.queryGraph), which carries the bead and its dependency rows in ONE spawn, and no longer calls bd dep list or normalizes its two row shapes. That surface is the only one an external: target survives on — bd dep list resolves each row to an issue record, and a cross-project target has none in this store. The two surfaces are reconciled by beads_dart's own externalDepRowsFrom, which REFUSES a record surface that dropped rows the resolving read still holds; that control is the only remaining second spawn. ExternalDepRef and its parser are CONSUMED from beads_dart — this package mints no second spelling of bd's vocabulary.

• Changed: the vended intake-refinement corpus teaches the projection in both overlay legs — a blocker is a declaration bd holds, the two refusal details verbatim, and the roster composition gap where the row is taught and again where the refusal is corrected.

• Changed: adopts the 2026-09-13 the_grid dev.3 wave — genesis_tree ^0.4.0, grid_engine ^0.4.0-dev.3, grid_sdk ^0.4.0-dev.3, grid_runtime ^0.2.1-dev.2, grid_trajectory ^0.2.1-dev.2 and beads_dart ^0.3.0-dev.2 (pow-abaw).

• Breaking: GitSourceControl.provisioner is a StationGitRepository, not a StationGitService, and GitSourceControl implements the new SourceControl.baseShaFor by delegating to it, so a committee pins its review diff to the exact commit the provisioner cut from (grid_runtime 0.2.1-dev.1, the_grid#436). GitGridAssets watches StationGitRepository and GitServices .provisioner carries one. Migration: a delegate that provided Provider<StationGitService> provides Provider<StationGitRepository>(StationGitRepository(service: <the service>)); every other SourceControl implementer adds baseShaFor, returning null when it cuts no worktree.

• Breaking: the state-store cross-link surface is GONE with grid_engine's (the_grid#447). CrossLinkBlockerSource is deleted, FilingService drops links and its stateRoot argument, FilingContract.evaluate drops linkedBlockers, kUnconsultedCrossStoreDetail is deleted, and ApproveService.approve / UnparkService.unpark drop stateRoot. A named FOREIGN blocker is now judged by the bead's OWN outgoing blocks edges like any other and reported missing, fail-closed; a cross-store blocker is authored as bd's external:<project>:<capability> row (grid link). The --state-root option stays registered on filing/approve and is VALIDATED there — the ONE spelling of the grid home across the verb set — but nothing reads through it any more; only the park pair does. Migration: drop the arguments; re-prove a cross-store blocker with a bd dependency row.

• Breaking: a bead whose approval was taken with --state-root CONSULTED and an open cross-store link MATCHED is RE-DIGESTED, and its standing grid.approved_rev is stale. The v1 basis SHAPE is unchanged — _approvalRevisionOf keeps its linked member rather than dropping it, which would re-digest every approved bead in every store — but the member is now pinned false, and the knownPrefixes a matched link contributed are gone with it. Two effects, and the quieter one is the worse: a digit-tailed foreign id (pow-f6pc) stays named and its row fails CLOSED, while a DIGITLESS one (pow-abaw named from a space- bead) is no longer read as an id at all, so it leaves the basis and the row passes VACUOUSLY — re-digested, still green, and no longer considering the blocker. Measured on space_station space-7tj: filing:v1:sha256:2ff365a6… becomes filing:v1:sha256:caf7bb15…, passed false on missing outgoing blocks edges: pow-f6pc. Consequence: approve and unpark REFUSE the fail-closed shape until a governor re-approves, and silently re-stamp the vacuous one against a basis that proves less; the stale revision stands in the audit trail and in approve's staleness report either way. Mount is unaffected today only because mountEligibilityFindings carries the revision without enforcing it. Migration: re-run the approve verb over every bead an open cross-store link names as BLOCKED, and re-declare the blocker of any bead whose row now fails. test/filing/filing_contract_test.dart pins BOTH cases — the locally wired golden that does not move, and the cross-store shape that does, against the measured pre-cut digest.

• Changed: the vended intake-refinement, discover and station-operations skills teach bd's external: row and the post-cut link verb; the retired type=link bead, grid.link.* metadata, crossLinkTypeRefusal, StationJoinBridge._applyCrossLinks and applyBlockGuard are named nowhere under station_overlay/.

• Breaking: the decision gather returns ONE evidence record instead of a list. gatherDecisions, commandDecisionIndexSource, AnchorsCapability.decisions and buildCodeRegistry's discoveryDecisions are typed on the new DecisionGatherEvidence ({decisionEntries, decisionLookups}), and DecisionSurfaceEvidence.decisions / namedElsewhere carry entry IDS (List<String>) that index into that map instead of inlining a DecisionEntryEvidence per surface, so an entry named by several surfaces is resolved and carried once. New public surface: DecisionGatherEvidence, DecisionReferenceCodec, DiscoveryAnchors.decisionEntries / decisionEntryFor, DiscoveryLensPromptAssembly, kMaxDecisionLensPromptBytes, kDecisionLensPromptOmissionReserveBytes and kDecisionLensEvidenceOmissionMarker (#310, #327). Migration: resolve a surface's entries as gather.decisionEntries[id] for each id in surface.decisions; a source function returns DecisionGatherEvidence(decisionEntries: ..., decisionLookups: ...) rather than a bare list, and DecisionSurfaceEvidence.fromJson / toJson take the codec that resolves those ids.

• Breaking: composePrimeContext takes a PrimeHandoff?, not a SeatHandoff?, and the free function handoffNamingLine is deleted. The naming line is now a member of the handoff record the launcher hands over, because a note the LAUNCHER already consumed and archived names itself differently from one still live on the disc (#328). Migration: pass PrimeHandoff.consumed(<the body delivered in kConsumedHandoffEnvironmentVariable>) or PrimeHandoff.unconsumed(<the SeatHandoff>), and read handoff.namingLine instead of calling handoffNamingLine(...); composePrimeContext also grows an optional handoffDiagnostic.

• Added: the seat archive-and-succession surface — SeatDisc.writeHandoffOnce, SeatDisc.newestHandoffState, SeatDisc.verifyIndexIntegrity, SeatArchiveSink, seatArchiveDirectoryName / parseSeatArchiveDirectoryName, seatArchiveStamp, seatArchiveDisposition, seatArchiveRetentionDisposition, seatArchivesToPrune, seatHandoffAgeDiagnostic, seatHandoffDeliveryRefusal, kSeatArchiveRetention, kSeatArchiveSubdirectory, kConsumedHandoffEnvironmentVariable, kHandoffWriteRemedy, SeatHandoffWriteException, SeatDiscIntegrityException, the new SeatSuccessionReport members (archive, archivesPruned, body, pruneRefusal, sink) and the injectable now on PrimeCommand, SuccessionCommand and SeatSuccessionService (#311, #312, #315, #328).

• Added: the release circuit registration — kReleaseCircuit, ReleaseCircuitRequest, ReleaseGateCapability, ReleasePromotionRouteCapability, ReleasePackageTarget, ReleaseCommandInvoker / InProcessReleaseCommandInvoker / ReleaseCommandInvocation, kReleaseGateCapabilityId, kReleasePromotionRouteCapabilityId, kReleaseReceiptKey and buildCodeRegistry's releaseCommands (#319).

• Added: AgentCapability.gitRunner and AgentCapability.completionContract, kAgentStep, kArgvTransport, BoundedTextBudget, the ArtifactFencedSession failure trio (failureKind, blockedDiagnostic, probeErrorDiagnostic), kNoRoundCommitDiagnostic and kRoundCommitUnreadableDiagnostic (#313, #314, #320, #321). kShowOutputCapBytes is now deprecated in favour of the shared output bound.

0.7.0-dev.1

• Changed: the seat launcher CONSUMES the handoff itself. Before it primes a child it archives the disc, proves the archive, deletes the note and its one MEMORY.md pointer line, and hands the successor the body it just consumed; a succession that refuses stops the launch instead of starting a successor over an unresolved disc. Consuming was an instruction the successor was asked to follow, and four sessions did not — one of them could not, because its disc was gitignored. The succession verb is unchanged as the by-hand recovery path (pow-d5ol).

• Added: succession picks its archive sink from the disc's own tracked state, probed with git check-ignore. A tracked disc takes the pathspec-scoped commit exactly as before; an IGNORED disc takes a local .grid/seats/<seat>/.archive/<utc-stamp>/ holding the handoff and MEMORY.md byte-identical, proved by reading the copies back, and reported as ARCHIVED-LOCAL <path>. git add -f is invoked on no path in either sink: a station ignores its seats tree for the PII a disc accretes, and forcing it into history would trade a lost handoff for leaked PII. An answer that is neither ignored nor tracked refuses and deletes nothing (pow-d5ol).

• Added: the consumed body reaches a HOOK-primed harness. A SeatPrimeMode.hook environment — claude, the seat --env default — takes no prompt segment and is primed by the station's own SessionStart hook, and the note that hook used to read off the disc is the note the launcher just deleted. The launcher now declares the consumed body in GRID_SEAT_HANDOFF and prime injects it from there; that declaration wins over the disc, and a note still ON the disc is by construction one no launcher consumed, so prime names it as the hand-started session that IS owed the succession verb. Without this the default environment archived the handoff, deleted it, and started a successor with nothing (pow-d5ol).

• Changed: prime's handoff naming line is PrimeHandoff, which renders the line for the origin it has — a consumed body says the launcher archived and deleted the note and that no verb is owed; a note on the disc says it is still there and names the succession verb. handoffNamingLine is gone and composePrimeContext takes a PrimeHandoff. The old single line told every successor to "run the succession verb in this turn", which after the launcher's consume sent it after a note that no longer exists (pow-d5ol).

• Added: seatHandoffDeliveryRefusal — the launcher proves the transport BEFORE it consumes. An environment that declares primeMode: prompt with promptMode: none and no session adapter has nowhere to put the body, and the launcher now refuses with the note untouched rather than archiving and deleting a handoff it would hand to nobody (pow-d5ol).

• Added: the LOCAL archive sink is bounded and never collides. Each succession keeps the newest kSeatArchiveRetention (ten) .archive/<stamp>/ directories and prunes the rest oldest first, deleting only names of the shape it writes; retention runs last, on an archive already proved byte-identical, and a prune that fails is reported as PRUNE INCOMPLETE without failing the succession, because blocking a handoff on a stale directory would re-create the defect the verb closes. A stamp already taken now takes the next ordinal — <stamp>-2, <stamp>-3 — where it previously REFUSED: the stamp has second resolution, the launcher's relaunch loop is not paced by a human, and stranding a seat on a clock tick is the same failure in a smaller window. seatArchiveDirectoryName, parseSeatArchiveDirectoryName, seatArchivesToPrune and seatArchiveRetentionDisposition are the pure surface; SeatSuccessionReport gains archivesPruned and pruneRefusal (pow-d5ol).

• Changed: the vended handoff skill carries ONE signal line — Handoff written: <path>; exit, the launcher relaunches this seat. — on both overlay legs. The /clear and /compact cases are gone: a cleared or compacted context is the same occupant carrying on, not a successor. Its Resume section now tells the successor the launcher already consumed, and names succession only as the recovery path. The cut is station-wide — the vended governor and refiner role definitions each told the seat to write the handoff "and then /clear", the one path the launcher cannot consume, and now teach the same ending the skill does; no file under station_overlay/ instructs /clear (pow-d5ol).

• Breaking: AgentArming.seats is now Iterable<SeatPreference> (was Iterable<ModelPreference>), and the kAgentsTargetHead / kClaudeTargetHead fields are removed (pow-ycoi, #298). Migration: a station that read seats as model preferences reads SeatPreference values and takes each seat's provider from SeatPreference.provider(); code that named the two target-head constants derives the head from the overlay mapping instead.

• Added: the open seat mechanism from pow-ycoi (#298) — SeatPreference, SeatProvider and CriticSeatProvider are the published way a station arms a typed seat over Nest; AgentArming survives only as an Iterable<SeatPreference> shim. Published at the dev rung on the 0.3-line grid pins so space_station can consume it (space-qyw) ahead of the genesis_tree 0.4 wave.

• Also in this line since 0.6.0: #292 #294 #295 #296 #301 #302 #303 #304 #305 #306 —

  • fix(assets): surface governor-owned blockers in sweeps (#306)
  • feat(agent): arm a relay seat by presence in the tree (#305)
  • fix(specify): type validation plan read failures (#304)
  • fix(code): run the validation plan as a child script (#302)
  • fix(specify): parse-check validation plans before filing (#303)
  • fix(grid-assets): anchor test readers on package root (#301)
  • refactor(agent): open typed-seat arming over Nest (#298)
  • feat(filing): vend a bounded one-bead show command (#296)
  • fix(assets): allow boot runner to be configured separately (#295)
  • feat(overlay): add agents_root mapping for repository root AGENTS.md (#294)
  • fix(grid_assets): format registries at package language version (#292)

• Added: explore-decision prompt assembly is bounded at 256 KiB of UTF-8 (kMaxDecisionLensPromptBytes). An oversize evidence bundle is clipped at a record boundary, says so in the prompt with a marker naming what was withheld and how to ask for it, and reports evidenceTruncated through the new DiscoveryLensCapability.assembleLensPrompt; buildLensPrompt is unchanged and delegates to it. The lens's freshness stamps and its absolute file-write instruction are reserved before any evidence is admitted and are never clipped, because a report missing them is discarded by the read fence while a clipped bundle is one the lens can narrate. Before this, a bead naming many surfaces against a mature register produced a request the harness answered api_error_status 400: Prompt is too long, which the circuit could only read as a model step exiting with no artifact. The byte arithmetic is the describe manifest's own, extracted as the public BoundedTextBudget (ceiling, omission reserve, clamp at a line boundary) rather than written a second time; buildDescribeManifest's output is byte-unchanged.

• Changed (BREAKING, DecisionSurfaceEvidence, DecisionIndexSource): the discovery gather carries each recorded decision body ONCE per round instead of once per touched surface. DiscoveryAnchors gains a gather-level decisionEntries index keyed by the body's own evidence id plus a loud decisionEntryFor; a surface lookup now carries ordered REFERENCES into it, and a DecisionIndexSource answers with the whole DecisionGatherEvidence rather than a list of lookups. Every roster-qualified surface of a repo resolves to the same entries, so a bead naming the full twelve anchors wrote twelve copies of each: a measured anchors.json was 3 201 633 bytes, 3 069 457 of it decision lookups, and the downstream decision lens then exceeded the model's context and exited with no artifact three rounds running. The per-surface selection order, the 96-entry bound, every state and every failure posture are unchanged — only the carriage is. The artifact is schema 3 and a schema-2 gather is REFUSED rather than adapted, because its lookups carried entry objects where a reference is now read; the round-aware sweep re-gathers one. A reference is written as the body's ORDINAL in the index's lexically ordered keys rather than as the ~120-byte evidence id repeated per surface, so a gather's reference overhead is a function of how many entries it carries and never of how long its register identities are; DecisionSurfaceEvidence.toJson/fromJson take the new DecisionReferenceCodec that owns that order, so a lookup cannot be serialized apart from the index its references point into.

• Added: runRecall takes an optional workingDirectory — the root its durable recall corpus path resolves against, defaulting to the process working directory, so the vended tool/search_recall.dart invocation is unchanged. A caller that owns a corpus elsewhere (the pack's own recall suite, which exercises record mode against a disposable copy) now names that root instead of assigning Directory.current. That property is process-global while dart test runs test files in concurrent isolates of ONE process, so the three suites that moved it raced the eleven that read package-local source by a relative path: a read scheduled inside another file's window threw PathNotFoundException or resolved a different tree, nondeterministically and never in isolation. The suite now anchors every package-local path on one cwd-independent package root and assigns the process working directory nowhere.

• Fixed: the asset generator formats a generated registry at the CONSUMING PACKAGE's own Dart language version — the inclusive floor of its environment.sdk — instead of --language-version=latest, so ONE committed lib/src/assets/grid_asset_pack.dart is current AND format-clean under every SDK a consumer runs. Under latest the rendered bytes were a function of whichever SDK happened to run the generator: an unchanged commit that was current on Dart 3.12 reported STALE lib/src/assets/grid_asset_pack.dart on 3.13, and no single committed file could satisfy both. The SDK dart format executable stays the formatter, and a pubspec whose environment.sdk is missing, malformed, or carries no inclusive floor is now refused LOUD before either output is rendered, compared, or written (pow-10fv; extends pow-5ifa, which put the SDK formatter executable in place; the duplicate tg-txbs was misfiled against the_grid and is closed).

0.6.0

• PROMOTED from 0.6.0-rc.26. This is the stable release of the 0.6.0 line; the code is the candidate's, unchanged. Every family dependency constraint is rewritten from its prerelease form to the stable one, because pub refuses a stable package that depends on a prerelease.
• Consumers on a ^0.6.0-rc.N constraint resolve this automatically: a caret range admits the release above its own prereleases, so no downstream pubspec edit is required to pick it up.

0.6.0-rc.26

• Changed: the refiner and governor agent definitions no longer present PUBLISHING as an action the seat does not take. Both now carry the org ruling: publishing a PRERELEASE is ordinary agent work with no per-release ask, candidates included, while the human owns the PROMOTION — beta to rc, and rc to a non-prerelease version. dev to beta stays agent work on its machine-checkable condition. Every other gate in those lists is unchanged: merging to main, the first live arm, persistence and credential changes (org-6si, pow-2cnz).
• Added: the release skill gains the RUNG LADDER it had no vocabulary for — dev/beta/rc as a per-package property distinct from the semver move, the --rung and --promotion-intent flags the vended verb takes, --change rc as the compatibility spelling of --change breaking --rung rc, and the rule that a breaking change enters at dev rather than being forced onto a candidate. Its trigger now admits an agent-initiated prerelease rather than only "when the human says release".
• Added: the release skill records a MEASURED limit on the stale-rc demote — 0.6.0-beta.1 sorts BELOW 0.6.0-rc.25 and pub resolves the highest, so a demote on a package with higher candidates publishes a version nothing selects.
• Changed: the station-throughput decision's human-gate enumeration is NARROWED where the agent definitions inherited their wording. It read "anything outward-facing beyond a branch push and PR", which swept in prerelease publishing — neither irreversible nor a human act, since pub excludes prereleases from stable caret ranges. It now names promoting a release instead, carrying an updated-by edge to the two org entries. No new decision minted.

0.6.0-rc.25

• Added: {{bootRunner}} — the RESIDENT-BOOT invocation hole, independent of {{runner}} and DEFAULTING to it. A station has two runtimes and the overlay could name only one, so a station whose runner is on PATH could not spell its seat verbs with the global name without also re-spelling its resident boot — which needs the JIT run form, because that is what carries --enable-vm-service and therefore hot reload, the reload verb and the leonard attach. station-operations declares the hole on both targets and uses it at its five boot sites; seat, down and status stay on the verb hole, none of them needing a VM service. Purely additive: a station naming only one runtime renders byte-identically (lunar_station-cu7, #288).
• Fixed: the overlay template substitution existed TWICE — the materializer's _render and the loader's _mustache — so a default bound in one was refused as an unbound hole by the other. Both now share the public renderOverlayTemplate, which is public because the two-runtime contract is worth asserting directly (#288).
• Changed: the build-agent working agreement makes a command, skill or other vended asset DONE only when a real runner or install path REACHES it, and requires the acceptance criteria to prove that reachable surface. A verb could previously ship in an asset, be named by a vended skill, and pass a green plan while no runner composed the command — so the installed skill taught a first step that could not run (pow-yt8n, #283).
• Changed: handoff board state and rulings carry bead TITLES next to ids, so a handoff read on a phone needs no store lookup; both overlay legs updated identically (Nico 2026-09-10, #282).
• Changed: floors dart_grid_assets at ^0.2.0-dev.1, the rung-split release service (#284).

0.6.0-rc.24

• Added: the specify capability stamps spec.author = specify on the spec it authors through writeSpecifyAuthoredSpec (the grid_runtime rc.19 chokepoint), so a rework re-authors specify text while preserving hand-written and governor-restored design / acceptance_criteria (pow-m8v5, #279; closes the loop opened by the_grid#390).
• Added: park and unpark filing verbs that reclaim a stalled session slot in one command, composed over the engine voidRetireMetadata and liveness seams with a marker-based park predicate (pow-6fqs, #280).
• Added: seat succession, the safe handoff-consumption verb (pow-s2sk, #276).
• Fixed: the asset generator normalizes generated Dart with the SDK dart format executable instead of the consuming workspace resolved dart_style, so --check and dart format --set-exit-if-changed agree (pow-5ifa, #277).
• Changed: floors grid_runtime at ^0.2.0-rc.19 and grid_sdk at ^0.3.0-rc.22.

0.6.0-rc.23

• Fixed: a decision-surface record is TRUNCATED only when its entry SET was clipped at kMaxDecisionEntriesPerSurface or a named entry failed to resolve; a decision doc over kMaxDiscoverySnippetChars keeps its own truncated state on the entry body and no longer fails the whole surface. With rc.22's register-wide resolution, 49 of 125 the_grid and 48 of 87 power_station decision docs are over the bound, so every spec round on those registers held at discovery with no narrow exit (lunar epoch 60: tg-nidl r3) (pow-jidn, #270).

0.6.0-rc.22

• Fixed: explicit decision citations resolve REGISTER-WIDE, never per surface. On rc.21 a real accepted decision cited from a surface that does not declare it read as absent and failed the surface, so every bead citing a sibling-surface decision held at discovery with no exit (lunar epoch 56: tg-nidl, lenny-dgp); a record indexed elsewhere in the gathered register now resolves and is annotated namedElsewhere (pow-9xkj, #263).
• Fixed: adopted worktrees fast-forward their base to the root checkout's default branch on re-adoption, best-effort with a receipt, so a surviving worktree no longer pins the base its first mint cut (pow-1g7, #260).
• Test: acp_session_adapter_test compiles the ACP probe fixture once per suite and bounds its waits by wall clock instead of fixed iteration counts; four different tests had failed across a lane, a merge-queue run and a PR check on 2026-09-08 under load (pow-qw6e, #266).

0.6.0-rc.21

• Fixed: the discovery decision-surface lookup selects bead-named records FIRST and raises kMaxDecisionEntriesPerSurface from 12 to 96, so a bead citing decisions on an 81-entry surface no longer holds at discovery on a clipped index; more than 96 named records is a FAILED surface, not a silent clip. Explicit references are only canonical register#slug tokens under a register the gathered index contains, or ADR-nnnn ids; a bare legacy A<n> token never fails a surface (it only orders), and a canonical-shaped token under an unknown register is prose. DecisionIndexSource gains the requesting Bead as a third positional argument (pow-mrg8, #257).
• Added: the interactive refiner role definition and its seat assets, vended with the baseline pack (pow-7nmo, #254).
• Added: design round classification and verification — the spec round records whether a design is a verify or an infer round and the committee reads it (#251); the verify inference wiring is pinned by test (#253).
• Fixed: FormatCleanCapability classifies a dirty-format verdict as work, never infra (#250).
• Fixed: the agent capability reports a provider capacity refusal as a typed non-result instead of an end turn (#248).
• Test: token-keyed WorkBead persistence is asserted (#249).

0.6.0-rc.20

• Fixed: FormatCleanCapability reports a DartFormatDirty verdict as CapabilityFailureKind.work and declares a supervisionPolicy that spends ONE deterministic attempt on it before parking at a gate whose reason names the would-change files. A formatter that ran and refused is substantive work, not an environment refusal: on lunar epoch 51 the untyped non-result was classed infra, counted five silent harness exits, flared harness.throttled and hid the offending file under the flare's underlying field (pow-7jvc). An undecidable probe — missing or unreadable pinned scope, a formatter that would not launch — stays a non-result on the circuit's own budget.
• Fixed: MountEligibilityAssets expires a cached fresh-read failure after kMountEligibilityReadRetryBackoff (30 s, doubling per consecutive failure of the same bead, capped at 5 min; a seed value with an injected clock) and re-arms the read on the next evaluation, and flares bd's own BdTimeoutException through kMountEligibilityReadTimeoutFlare; on lunar epoch 43, 280 of 500 refusals were boot-burst read failures cached until the bead was edited (pow-l73v, #246).
• Fixed: the discovery lens freshness fence carries the session id, so a re-minted session sweeps the PRIOR session's lens reports and waits for its own instead of joining a stale verdict; a missing SessionHandle fails fast (pow-l2r5, #245).
• Added: HarnessProvider mounts InheritedSeed<SiteBinding> above the availability seed and runs EnvironmentRegistry.validate as the ONE boot-eager gate, whose site-binding tail now covers only the environments an arming reaches plus custom entries — a bare station with builtin pi unbound still boots, an armed environment with a missing endpoint refuses with SiteBindingError (pow-2eg, #244).
• Added: the committee selector's Ok payload and the shadow route's Advance payload persist the selection run and receipt (17 and 34 reserved keys; the six pre-existing keys byte-identical) (pow-1nl.1.3, #242).
• Fixed: decision lookups render as cd '<grid home>' && <verb> so critic lanes and the architect can run the roster-mode decisions index from a worktree (pow-96ei, #243).
• Floors beads_dart to ^0.2.0-rc.10, grid_engine to ^0.3.0-rc.20, grid_runtime to ^0.2.0-rc.13 and grid_sdk to ^0.3.0-rc.16.
• Tests: the PDR §7 (e) spawn-in-flight probe admits its bead through StationAdmissionAuthority before mounting the SessionScope directly, and answers grid_engine rc.20's mount-attempt read and record before the gated session create — the exported GatedCreateBdRunner still answers bd list with an object, which rc.20's mint prelude cannot parse.

0.6.0-rc.19

• Fixed: the mount gate no longer compares a bound approval stamp's revision against a fresh filing evaluation. rc.18 (pow-9rah) bound grid.approved_rev over the bead's design, acceptance criteria and notes and refused a mismatch as approval: stale; the station's own specify step writes design and acceptance on every first round and rework appends notes, so every bead approved on rc.18 un-approved itself at its own spec write and the engine's content gate disposed the running round (lunar epoch 43, 2026-09-05: four of four fresh rounds, only legacy raw-sha stamps survived). mountEligibilityFindings still requires a COMPLETE stamp of either form and still accepts evaluatedApprovalRevision, which is recorded but not enforced; the approve verb keeps writing the bound revision for the audit trail. Re-instating a comparison over a basis the station does not itself rewrite is pow-lr8n (governor patch, precedent pow-gcx9).

0.6.0-rc.18

• Fixed: the explore-decision lane renders a decision entry's body ONCE however many roster-qualified surfaces cite it — the first surface carries the body, every later surface cites the entry by identity as "also governs this surface". The roster-mode decisions index answers every surface of a repo with the same entries, so a bead touching 12 power_station surfaces carried 12 copies of each (550 KB for 12 distinct bodies on pow-ed1c) and the cheap lens died prompt_too_long, holding the round on a failed lane. Per-surface lookup records and the evidence-id profile (rc.17, pow-bvui) are unchanged (pow-alwh, #239).

0.6.0-rc.17

• Fixed: DiscoveryAnchors.fromJson no longer refuses the gather the deterministic gather itself writes — the colliding-profile guard counted a decision ENTRY once per roster-qualified surface while evidenceIds collapsed the same body cited under several lookups, so on rc.16 (the first release whose decision lookups resolve) every discovery round with decision entries decoded to an empty gather and held at discovery-route on `gather:round -1 / gather:workBeadId ``. Decision entry ids are counted distinct; a duplicated lookup id or a body id reused by a foreign record still refuses (pow-bvui, #237).
• Fixed: MountEligibilityAssets bounds its fresh read — kMountEligibilityReadDeadline (60s) wraps the bd.query in _readFresh, a timeout is routed through the existing failure cache as a refused decision naming the bead, store root and deadline, and is flared by name (kMountEligibilityReadTimeoutFlare) so the operator can see WHICH store's read hangs; genesis and decisions beads sat at fresh mount-eligibility read pending forever across nine lunar boots (pow-7i3k, #236).
• Fixed: the work bead's OWN citation fields (description, design, acceptance, notes) are carried WHOLE into the gather — boundDiscoveryEvidence gains truncateSnippet (default true; only BeadFieldEvidence opts out), so a description over kMaxDiscoverySnippetChars no longer holds its own discovery round as TRUNCATED; foreign evidence keeps the bound (pow-xfm0, #235).

0.6.0-rc.16

• Fixed: the shelled decision lookup accepts the decisions index envelope spec: 2 beside spec: 1 (_acceptedDecisionIndexSpecs); against decisions_grid_assets 0.2.1 every explore-decision surface FAILED on rc.13–rc.15 and the round held at discovery-route (pow-rokz, #231).
• Fixed: approval receipts are complete and bound to the filing basis — ApprovalStamp.tryParse requires a nonblank grid.approved_by, a UTC grid.approved_at, and a recognized grid.approved_rev (a legacy raw git sha, or the filing:v1:sha256: digest the approve verb now writes over the bead's filing contract); a lone grid.approved_at no longer mounts, and mount eligibility re-checks the fresh bead against the stamped revision (pow-9rah, #232).

0.6.0-rc.15

• Fixed: the discovery gather no longer treats a record clipped at its OWN bound (kMaxDiscoverySnippetChars, kMaxHistoryCommits) or a clipped anchor/symbol extraction (kMaxAnchors) as a deterministic gap — every mature surface exceeds those bounds, and the override held every round at discovery-route while the lenses reported zero gaps. A clip is rendered as context the lens narrates; only failed overrides a lens, and only the lens's own insufficient-evidence outcome holds the round on a clip. gatherHistory records unavailable over an empty resolved path list instead of logging the whole repository (pow-gcx9, #225).

0.6.0-rc.14

• Added: the grid: block accepts an optional teaches: sequence per skill asset (the deterministic commands its prose teaches), parsed by parseGridBlock, emitted by the registrant generator only when non-empty, and authored for the baseline skill/command pairs; unclaimed packs generate byte-identical output. Floors grid_sdk ^0.3.0-rc.15 for GridAssetDefinition.teaches (pow-prw4, #228).
• Fixed: the discovery gather's roster decision lookup qualifies surfaces from the SESSION's substation (metadata['rig'] is a session-bead field no work bead carries), never shells a <repo>-prefixed surface (recorded unavailable instead), and runs the station's verb from the composing grid home (overlayArgs['gridHome'] ?? devRoot) rather than the work worktree — every explore-decision lens held its round on rc.12/rc.13 (pow-974y, #226). Stations pass 'gridHome' beside 'runner' in buildCodeRegistry(overlayArgs:).
• Fixed: --state-root on the filing verbs takes the GRID HOME its help names — the resolver appends .grid when the root holds one, accepts a state store unchanged, and refuses a root holding neither child loudly; an unconsulted cross-store edge is reported as UNCHECKED, never as missing (pow-ixag, #227).
• Fixed: the code-validation gating lane preserves the validation failure output the way land/revalidate does — the shared captured-output leaf recognizes [E] beside Error: / Failed to load, and the full log lands in .grid/critique/code-validation.log (#224).

0.6.0-rc.13

• Breaking: asset availability is resolved ONCE — resolveGridAssets over a repository-observed SubstationFactsSnapshot defines the tree's selected definitions and BOTH overlay writers (assets install, OverlayMaterializer) and the landing pre-rebase guard consume that same resolution; --check reports drift. Un-migrated stations without facts keep the pre-resolution path (pow-4peu, #216). Migration: compose a station's assets through resolveGridAssets/GridAssetRegistry — a writer must never build a second catalog; github_grid_assets ^0.1.0-rc.13 carries the matching SubstationSeed.
• Breaking: operatorSkillIds is a FUNCTION over the station-resolved GridAssetRegistry (operatorSkillIds(registry)), not a getter over this package's own pack — a downstream pack's audience: human declaration is now withheld from a build agent's brief (pow-vtts, #218). Migration: pass the resolved registry; vendedSkillIds is unchanged.
• Added: prime --hook-json (echoes bd prime and injects an operator seat's newest handoff on SessionStart sources startup/clear/compact — never resume) and seat <name> (the harness-neutral operator-seat launcher composed over AgentEnvironment); five nullable seat declarations on AgentEnvironment (drivenArgs, roleAsset, roleArgs, memoryDirArgs, primeMode) — the driven-only flag moved from args to drivenArgs, one-turn argv byte-identical. The vended station overlay's SessionStart hook is now {{runner}} prime --hook-json; a station adopting this rc MUST compose PrimeCommand/SeatCommand into its runner (pow-lv6t, #220; space_station space-31x).
• Added: SHADOW stage-specific committee selection — CommitteeSelectionPolicy (eight deterministic rules over the round-stamped discovery evidence / pinned diff, a closed classifier allowlist, fullFallback on a non-result), CommitteeSelectionCapability as a dependency of NOTHING, CommitteeShadowRouteCapability wrapping every route, typed JSON receipts with lane input digests and counterfactual totals; the full committees stay authoritative. Depends on grid_trajectory ^0.2.0-rc.4 for its report vocabulary (GateDisposition, LaneReport, UsageSample) — value types only (pow-1nl.1.1, #222).
• Added: usage cost is DERIVED from a declared per-model price table (ModelPriceTable, kUsageModelPrices, keyed by the ladder's model ids) when a harness reports tokens but no total_cost_usd (codex), stamped UsageCostSource.derived vs reported; an unknown model keeps its tokens, reports null cost and flares (pow-zetn, #221).
• Fixed: the land/revalidate reason leads with the Dart front-end's Error: / Failed to load lines (deduplicated) before the tail and writes the full output to .grid/critique/revalidate.log (pow-gvfx, #217).

0.6.0-rc.12

• Breaking: the vended asset surface is DECLARED in the package's own pubspec.yaml grid: block and CODEGEN'd into a typed Dart registrant (GeneratedGridAssetRegistrant); the hand-maintained kVendedSkills/kOperatorSkills mirrors are retired and extension/mcp/config.yaml is generated from the same block (pow-u6hj, #205). Migration: a downstream pack that listed its assets in the const mirrors declares them in its grid: block and runs the generator (dart run tool/generate_grid_assets.dart); nothing else changes for stations that only compose the vended packs.
• Breaking (floor): grid_engine ^0.3.0-rc.15 — critic verdict failures ride the typed CapabilityFailure seam (CapabilityFailureKind.invalidResult), CriticCapability declares its own SupervisionPolicy (retry its lane only, never grade F, gate on exhaustion), and the six Failed.nonResult call sites are gone (pow-dzc, #214).
• Added: the station Dart registrant is generated from the resolved package closure — every dependency's grid: block is unioned once (pow-bafc, #212).
• Added: discovery evidence is gathered ONCE into a bounded, round-stamped DiscoveryDossier with provenance and explicit truncation, and each inference lane receives a capability-specific projection; the decision-index gather runs through the composed decisions command or the station's runner, and an ABSENT tool is unavailable, never a gap (pow-ri9c, #208).
• Added: parseSpecContract — the typed record grammar for specs (AC-n ids, labeled steps, exact AC-n -> command -> expected validation mappings) measured in SHADOW by spec_contract_shadow.dart; the five live presence checks remain the A/F gate (pow-5ufz, #207).
• Added: format-clean, a deterministic code-review step before the critics that gates unformatted Dart naming the files (pow-jicn, #209).
• Added: PR-description inference receives a bounded (16 KiB) deterministic change manifest — commit subjects, diffstat, change shape, bead identity, receipts — instead of the raw diff (pow-c7lb, #203).
• Fixed: the decision-alignment brief renders the roster lookup from the STATION's runner (overlayArgs['runner']) and the rubric uses the {{runner}} hole — no more hardcoded space decisions index (pow-q7ty, #213).
• Floors tightened to dart_grid_assets ^0.1.2.

0.6.0-rc.11

• Fixed: SpecifyCapability is harness-neutral — it routes the spec seat through its AgentSessionAdapter (a shared _resolveRun feeds both spawn and a new createSession), and the ACP bridge now reports its child's real exit status and a bounded stderr tail instead of a bare "output closed". landReasonTail, kRevalidateReasonTailChars and planOutputWithoutPubAdvice move verbatim from the landing circuit to a zero-import src/agent/captured_output.dart leaf (same names, same behaviour, newly exported from the barrel) beside the new capturedOutputReason assembler; usageEnvelopeJson/writeUsageEnvelope render FT-2 telemetry a channel harness has no wrapper to produce, and AgentSessionAdapter.launch gains an optional usageOut to carry the path across the seam (pow-39tl, #197).
• Changed: the vended governor agent overlay states the seat's COST posture — ranked under throughput, with the model/effort calls it implies — pinned by governor_posture_test.dart (pow-8dwh, #198).
• Added: the /handoff ritual vends as an operator-audience skill on BOTH overlay legs (station_overlay/claude/skills/handoff/ and station_overlay/agents/skills/handoff/), with the MCP config.yaml entry and the asset_loader wiring that carries it (pow-pry0, #199).
• Fixed: scaffold restore MERGES directories instead of clobbering them, and a failed provision UNWINDS what it created rather than leaving a half-cut worktree behind — the shape that wedged fresh worktrees on a scaffold collision (pow-gnrm, #200).
• Fixed: the discovery circuit fences lens reports on ROUND. Every lens prompt stamps round beside nodePath and one shared fence reads both on every read path, the round-stamp parser is promoted out of committee.dart as the shared stampedRound, AnchorsCapability's blanket .grid/discovery wipe becomes a round-aware sweepStaleDiscovery, and the route join classifies an artifact-less lane (decided-with-no-artifact vs. merely LATE) instead of dropping it (pow-3yo, #202).

0.6.0-rc.10

• Fixed: AgentSession.onRuntimeEvent handles RuntimeEvent.sessionOrphaned. The arm is an OBSERVATION, not a state change: it flares agent.sessionOrphaned (sessionId, pgid, memberCount) on the injected ExplorationTransport and returns, so the session stays live and supervised and the Exited/Died after the provider's bounded grace is still the terminal. AgentSession gains an optional transport parameter, wired in AgentCapability.createSession from the ambient ServiceBundle.transport; absent means no flares, never a failure. The switch stays exhaustive with no default arm, so the next lifecycle variant is caught the same way.
• Fixed: this pack's three ProcessGroupController fakes implement the groupMembers(pgid) member the same upstream change added, so the package analyzes and tests clean again.
• Breaking (floor): grid_runtime is floored at ^0.2.0-rc.10. A single source cannot be exhaustive over RuntimeEvent under both rc.9 (no SessionOrphaned) and rc.10 (with it), so this pack now requires the candidate that carries the variant and ships with that wave.

0.6.0-rc.9

• Added: the typed-seat arming MECHANISM is vended — AgentArming (the pure per-seat VALUE), TypedEnvironmentProvider (the ONE seed both the station rung and the per-substation rung mount) and SeatEnvironments (the offline projection of all four resolutions at a point in the tree) now ship from lib/src/agent/seat_environments.dart, beside the seat vocabulary they wrap. AgentArming is the typed-seat shape: four typed seats and no role rung. A composing station keeps its own named environments and ladders: mechanism is vended, posture is not.

0.6.0-rc.8

• Fixed: the filing verb reads the grid home's cross-store link beads. It gains --state-root <path> with the same injected default approve carries, and both verbs now register and resolve it through one seam (lib/src/filing/state_root_option.dart: kStateRootOption, kStateRootHelp, noStateRoot, addStateRootOption, resolveStateRoot) — a blocker wired by an open link bead used to pass approve and fail filing. The named-blocker parser also tightened: a blocker is DECLARED by a segment that OPENS with Blocked by / Blocked on / Depends on (a mid-sentence mention declares nothing), and a <prefix>-<tail> token is a bead id only when its prefix is known or its tail carries a digit, so cross-store is no longer reported as a missing blocker.
• Breaking: approval IS the grid.approved_* stamp. mountEligibilityFindings no longer reads the grid.approved LABEL — its clause is now if (!isApprovalStamped(bead)), refusing with approval: not approved - run the approve verb — the kApprovedLabel constant is deleted, and the approve verb writes only the three stamp keys (grid.approved_by, grid.approved_at, grid.approved_rev) in its one bd update, with no --add-label. A bead carried four encodings of "not yet" and the label and the stamp were the same act written twice; a label any writer could add mounted work ahead of its blockers, while a hand-added one silently never mounted at all. Migration: every open bead holding ONLY the grid.approved label stops mounting — re-approve it with <runner> approve --actor <name> <bead-id>, which stamps it. Beads already stamped by the verb keep mounting untouched; the now-inert label needs no removal. ApprovalStamp, the three key constants and isApprovalStamped are unchanged and still exported, and the filing preflight is untouched. The vended station overlay teaches the stamp rule.

0.6.0-rc.7

• Breaking: the role map is retired — AgentRole, roleEnvironments, modelForRole, stationModelFor, tierFor and defaultModelFor are gone; resolveAgentConfig maps role to tier and model environments are selected by value through typed seats (seat_environments.dart) (#161, #168, #170). Migration: author the agent posture as typed seat environments instead of a role map; see lib/src/agent/seat_environments.dart.
• Breaking: author-owned spec-review verdicts route to the human gate (pow-hxme, #164), and committee metadata changes route through docs review (#162).
• Live environment availability is published to the arming surface (pow-n6n.3, #165).
• The decisions register: design decisions are recorded as slug entries under docs/decisions/ and the lens excludes views/ (#169).
• The approve verb stamps grid.approved_by/at/rev beside the label, and the overlay skills teach approval through the verb (pow-5ch, #163).
• Declared-tests base-gates bare prose test mentions (#158).
• Requires grid_runtime ^0.2.0-rc.8 and beads_dart ^0.2.0-rc.5 (the wave-2 the_grid train).

0.6.0-rc.6

• Breaking: a grid.approved label without a grid.approved_at stamp no longer mounts. The new approve verb (ApproveCommand) gates its atomic receipt write on filing completeness — sentence-scoped blockers and state-store links included — and records the approver, the UTC instant and the repository revision, so mount eligibility can tell a verb-issued receipt from a bare label (pow-kps, #159). Migration: re-stamp every open grid.approved bead with the approve verb; a bare label is refused at the gate. The governor already ran this org-wide on 2026-09-02.
• The bead filing contract is enforced by a read-only filing service and FilingCommand, which reject mechanically incomplete author-side filings; discover calls the command, and the defer and mount-eligibility boundaries are unchanged (#148).
• ACP-backed agent sessions: Copilot and Codex drive through one long-lived ACP adapter carrying structured progress, completion, usage, steering, permissions and model selection, with hermetic protocol-conformance coverage and an opt-in live worktree proof (#153).
• Channel-backed agent sessions: harness-specific session adapters, bead-routed fenced steering and environment opt-in wiring; agent briefs, structured results and usage travel the long-lived channel while every builtin stays on the one-turn path (#147).
• The architect agent role gives specification agents an independent environment role, with a build fallback so existing build-environment armings keep working (#154).
• Copilot one-shot telemetry: the Copilot environment declares silent JSON output and keyed resume, and projects premium-request consumption plus session duration through the generic usage report path (#145).
• Critic verdict rounds are authored at the capability boundary: a canonical verdict must be proven to belong to the current critic incarnation before it replaces a model-authored round, the model value is preserved for diagnostics, and an unresolved durability probe flares (pow-uok, #157).
• Critic verdict artifacts are written through same-directory atomic replacement, and respec ledgers are fenced by their owning session root, so a concurrent or stale artifact cannot poison a live join (#140).
• specify completion is gated on a fresh exact-id readback through the owned bd client, failing closed when authored acceptance or design is absent or unreadable (#149).
• Mount eligibility preloads the owning store's bead snapshot and rechecks a tentative refusal against the fresh bead, so first-refusal clauses derive from current fields; eligible snapshots stay synchronous (#152).
• Declared tests: extraction is narrowed to authored declarations and ignores run commands, quotation contexts and unchanged/restore statements (#150); bare Test: run references are split from authored declarations and consult the pinned base only as the fallback set (#155); package-relative declarations resolve against repo-relative pinned-diff paths by path suffix (#138).
• Readiness, specify, spec review and discovery prompts search both local decision homes (the ADR directory and docs/decisions) through one missing-safe command set, and packaged assets accept legacy clauses as well as decision slugs in a citation (#146).
• The vended skills teach the enforced approval-label transition in place of defer staging, with pinned source rendering and operator installation (#151).
• Tests: the specify environment assertion names the architect environment rather than a bare codex argv (#156).

0.6.0-rc.5

• New declared-tests-present code-review lane: confidently-declared test paths in the design are compared against the pinned diff; omitted files hard-block the round (#131).
• CompletionContract.artifactDurability adopted for every critic: recovery lives in the probe (the tg-291 stdout salvage recovers and persists canonically), result()'s unreachable envelope/fail-closed tiers are deleted, and the artifactless SiblingView cache fallback is removed — the join waits on durable artifacts, never a cached completion (#132).

0.6.0-rc.4

• Critic verdict artifacts are strict-decoded (non-object root, off-ladder/blank grade, blank rationale/nodePath, non-integer round all refuse); a present-but-malformed verdict fails the lane loudly (AllocationFailed, reason-prefixed) instead of silently grading F; unknown read exceptions fail the same way. Repair rides criticRepairInstruction on engine-supervised restarts (#128).
• Bundle derivations converted to ServiceBundle.derive — new bundle fields compile-error instead of silently dropping (#126).

0.6.0-rc.3

• DeliveryMethod seam additions backing the grade-gated landing postures (#121).

0.6.0-rc.2

• MountEligibilityAssets — the composable mount gate (pow-50l, #114). A station that mounts this seed admits a work bead only when it carries a driveable type, a validation_plan, and the grid.approved label. Without it the gate is INERT and every ready bead mounts, which is what 0.6.0-rc.1 shipped: the class exists on main but is absent from the published 0.6.0-rc.1 archive, so consumers resolving from pub could not compose the gate at all (pow-w83). This release is that fix — the version moves so the archive and the source stop disagreeing at the same number. It stays an -rc because grid_assets still depends on pre-release grid_engine / grid_runtime / grid_sdk / beads_dart / grid_exploration, and pub requires a package depending on a pre-release to publish as one.
• The git composition collaborators are watched from the tree rather than passed as constructor params (#113), matching the seat-facing const-services direction.
• Terminology: the human approval gate is "approve/approval" throughout (#118).
• Tests: the invariant-2/3 acceptance suites assert chokepoint creates by SHAPE rather than by a hard total, so they hold under both published-dep and path-override resolution (the_grid tg-zlfu adds a mount-attempt write).

0.6.0-rc.1

• Breaking: the GitHub implementations are REMOVED from this package and now live in github_grid_assets (pow-2ua, power_station #109). Six exports are gone: GitHubAppPrOpener, GitHubPrDelivery, GitHubGridAssets, GitHubReconciler/GitHubReconcilerRuntime, GitHubReconcilerCursor/ GitHubCursorStore/FileGitHubCursorStore, and NormalizedGitHubEvent. Migration: depend on github_grid_assets ^0.1.0-rc.2 and import them from package:github_grid_assets/github_grid_assets.dart. The abstractions they implement stay here — DeliveryMethod, DeliverRouteCapability, SourceControl, PrComposition and every *Capability are unchanged.
• Breaking: this package no longer depends on github_grid_assets. The dependency direction is inverted per the org rule: grid_assets holds the generic assets and the abstractions other asset packages implement, domain implementations live in their own domain package, and the edge runs implementation -> abstraction. Anything that reached a GitHub symbol transitively through this package must now depend on github_grid_assets directly.
• The MINOR moves rather than the patch specifically so ^0.5.0-rc.1 resolvers do not silently inherit the removals.

0.5.0-rc.1

• Breaking: adopts the_grid's 0.2.0-rc.1 prerelease wave — beads_dart ^0.2.0-rc.1, grid_runtime ^0.2.0-rc.1, grid_engine ^0.3.0-rc.1, grid_sdk ^0.3.0-rc.1. Published as a prerelease because pub requires a package depending on a prerelease to be one itself.
• Breaking: BdExportBeadSource no longer shells bd export --all, which is refused in proxied-server mode and whose API was deleted upstream. It now issues ONE all-status bd query --all --json per store. The contract is unchanged — one spawn per store, a read-only probe that never mutates, and closed beads are still included.

0.4.0

• Breaking: rides the 0.2.0 substrate wave — grid_engine/grid_sdk ^0.2.0, genesis_tree ^0.2.0 (foundation diagnostics; ext.leonard.* namespace).

0.3.1

• CodeCircuitResolver accepts an optional pre-classification overrideFor policy: a non-null override roots that circuit without cursor classification; the null path is byte-for-byte unchanged. Enables subclass stations to route selected beads (e.g. burn orders) to non-code circuits.

Changelog

0.3.0

• Breaking: overlay assets now ship from VISIBLE source directories (extension/station_overlay/claude/, agents/, github/, …) mapped to dot-targets (.claude/, .agents/, .github/, …) at install time — dart pub publish strips hidden directories, so 0.1.0/0.2.0 tarballs shipped HOLLOW (no operator files at all). Default mappings cover claude/agents/github/copilot/codex; the overlay manifest can declare its own. Migration for OverlayInstallService.install overriders: the overlayRoots: List<String> required parameter is now optional and superseded by overlaySources: List<StationOverlaySource>; root providers return List<StationOverlaySource> instead of List<String>.

0.2.0

• Breaking: SearchCommand and AssetsCommand now OWN the --grid-home flag, its absolute-path guard, and normalization; the delegate factory seam changed from a zero-arg closure to GridDelegate Function(String gridHome). Migration: drop your own --grid-home option registration and resolve/guard block, pass gridHomeDefault, and curry the factory with the command-resolved home — delegate: (gridHome) => MyDelegate(gridRoot: gridHome). The AssetsCommand install leg treats the flag as an explicit OVERRIDE only: when absent the default remains mount-then-read-ambient GridRoot via mountedGridHomeOf.
• Added runnerInvocation as a first-class parameter on AssetsCommand and OverlayInstallService, so a JIT-launched station renders the {{runner}} template hole without subclassing. Omitted, behaviour is unchanged (runner.executableName remains the default).
• Added buildComputeServeCommand() / buildComputeLeaseCommand() — the compute asset now vends its own fully-wired ServeCommand / LeaseCommand instead of every station copying the assembly block. The --allow list stays a caller-supplied parameter (station security policy).
• Generalized codedRosterOf off the station-specific factory typedef, with dispose-on-throw fenced.
• Fixed: spec verdict rounds are sourced from circuit params.

0.1.0

• Initial release: the_grid's opinion assets — the agent/verify/land Capability impls, the code circuit, and the git SourceControl.

0.2.1

• Store station overlay assets in publish-visible source directories and map them to harness dot-directories at install time.
• Allow asset manifests to override the default station overlay mappings.
• Warn during install and release dry-run when hidden overlay source directories would be omitted by dart pub publish.