Prepared 2026-10-03. Proposed architecture, awaiting independent/owner review under existing Eve policy. Product requirements are in the companion PRD. Names below are proposed interfaces and tables, not claims that they already exist.
Revision 4 (2026-10-03) applies Elikem's instruction to adopt the assistant's review and a Codex read-only second opinion: triage and tier records, per-membership revision selection with provenance class, a transactional content watermark that current sign-off checks, synthesis outcomes including “review completed, changes required”, finding-resolution semantics, cut-from-scope and standing-decision capture, bounded decision visibility to agents, claim labelling, a calibrated estimator, execution provenance bound to the instruction, the operational-approval bridge, realistic scale gates, and two owner pilots in the rollout sequence.
1. Existing seams and architectural choice#
The Workspace is apps/oshun/eve-content: Next/React with Payload/PostgreSQL, authenticated content commands, unlimited document versions, source records, audit events and publication snapshots. Source formats include blocks and rich text. tools/eve-content supplies format validation, canonical hashing, readable projections and an authenticated HTTPS client. tools/eve-task-board owns operational task/review/lease behavior, task→document bindings and links, task context, and standing decisions; tools/workbench-mcp/eve-server.mjs exposes the content and board tools.
Add an isolated granular-review domain beside content in the existing service/database. Reuse canonical document identity and immutable revisions. Do not add personal state to documents.state, eve_api.reviews, task status, claim ownership or publication eligibility. No operational reader is made dependent on availability of the new domain. Review decisions update only the new domain; explicit dispatch uses existing governed interfaces. Three existing mechanisms are reused rather than duplicated: task→document bindings and links (implementation status per span, and the bounded read of owner decisions through task context), the board's standing-decision record (capture from a rationale), and the content review command (the operational-approval bridge calls it unchanged with an evidence reference; the review queue gains links and a coverage summary only).
apps/oshun/eve-content/src/workspace/WorkspaceShell.tsx, Workspace.tsx, DocumentReader.tsx, Diagram.tsx, request.ts, preferences.ts, the review-queue components and existing tests are integration seams. Backend changes will add modules, collections/migrations and endpoints beside service.ts/read.ts, not enlarge the operational review state machine. Re-evaluate paths against the then-current main before implementation.
2. Domain records and invariants#
| Record | Minimum fields and invariant |
|---|---|
| Owner binding | Stable authenticated human user ID, environment, provisioned policy version. An agent cannot edit this binding. |
| Campaign | UUID, owner, title, lifecycle, selected product-version ordering, tiers (must review, spot check, context only), cutoff, manifest generation/hash, checked content-watermark generations, checked inventory generation and census classifier version, optimistic version, created/archived times. Prior campaigns are retained. |
| Version membership | Product-version key and explicit order, domain applicability, source document ID, selected revision/state/hash, provenance class of the selection (publication snapshot, release association, source lineage, none), inclusion disposition/reason, tier, shared-foundation links. Version labels are not document revisions; a membership without version-specific provenance is labelled as such. |
| Triage disposition | Owner, campaign, document, outcome (include, spot-check, historical-context, excluded with reason), spot-check sample seed, server timestamp, idempotency key, superseding event. Proposed by the census, confirmed by the owner; never edits canonical metadata. |
| Coverage manifest | Immutable versioned inventory of primary spans, contextual spans, exclusions, unavailable material, tiers, spot-check samples and baseline selection, with the inventory generation and classifier version it was computed at. Counts and the hours denominator derive from this manifest. |
| Review plan | UUID and version, campaign/document revision, ordered chunk membership, dependencies, generator and estimator versions, author provenance, predecessor/successor relation. Active plans with decisions are immutable. |
| Chunk | UUID, plan version, title, intent, default review question by kind plus suggested questions, primary source anchors and hashes, claim labels per span (proposed, asserted-existing, unknown) with their source (heuristic, assistance, owner), pinned context/dependency references, estimated duration/complexity, boundary rationale, approved-content eligibility. |
| Session/draft | Owner + campaign/plan/chunk identity, position, note/decision drafts, preferences, optimistic version, save receipt, measured elapsed time per chunk. Private to the owner; it grants no decision authority. |
| Decision event | Append-only ID, owner identity, explicit outcome (approve, changes-requested, cut-from-scope, historical-context-reviewed, defer), target revision/plan/spans/context hashes, claim labels at decision time, rationale, visibility state (submitted decisions and rationale are agent-readable for bound spans; drafts never), execution provenance and, for assistant-executed events, the instruction reference bound to this exact action and subject, server timestamp, idempotency key, superseded event ID when corrected. |
| Annotation/question/finding | UUID, owner visibility (private until dispatched; the body stays private after dispatch unless included in the confirmed payload), pinned anchor, body, state, resolution event kind (clarified, dismissed, revised-and-accepted by a new decision), rationale/evidence, optional current correspondence and dispatch links. Never an implicit board blocker; resolution never changes the original decision's subject. |
| Revision correspondence | Old/new source and plan identities, deterministic mapping evidence, content/context comparison, mapping algorithm version, exact/unchanged/unchanged-dependencies-uncertain/changed/ambiguous/deleted outcome, revision provenance (task/agent and stated reason). Rebuildable. |
| Content watermark | Monotonic generation per document and per dependency edge, advanced in the same transaction as every relevant change to an existing subject: new revision, archival or restore, eligibility metadata, relations, referenced assets. The authority current sign-off compares against for members; background projections may lag it but never substitute for it. |
| Inventory generation | Monotonic generation per campaign scope, advanced in the same transaction as document creation, archival or restore, and kind/domain/release/eligibility metadata changes of any document. The manifest records the generation and census classifier version it was computed at; current sign-off verifies it (§7) so newly eligible material outside the checked membership is detected. |
| Estimator calibration | Version, formula parameters per kind, the measured durations it was derived from, effective-from generation. Changes future estimates only. |
| Synthesis/sign-off | Frozen manifest hash, scope (baseline or current), outcome (review completed approved; review completed changes required; completed with exclusions), constituent decisions and exceptions, unresolved checks, explicit owner acknowledgement, source/plan generations and, for current scope, the member watermark generations, inventory generation and classifier version verified and any pending candidates disclosed. Historical sign-offs are immutable. |
| Dispatch receipt/outbox | Owner-approved payload hash, kind (task, proposed revision, successor, new-document proposal, standing decision), subject kind and lifecycle state, predecessor revision/hash/spans, target family/document/task, affected bound tasks shown at preview, idempotency identity, operation state, external receipt/backlink, attempt/error metadata. |
| Projection/job checkpoint | Event cursor, source revisions processed, generation/hash, last success/error, freshness and bounded retry state. Stale projections cannot claim current completeness and cannot authorize a current sign-off. |
All mutations are transactional, append an audit event, enforce optimistic preconditions and retain retry identities. Prevent accidental direct collection writes using existing domain-command access patterns. Validate all cross-record owner, campaign, revision and plan relationships on the server. Never accept a client-supplied actor as authority.
3. State and completion semantics#
Use separate axes rather than an overloaded status. Triage disposition: include, spot-check, historical-context, excluded. Workflow position: not-started, in-progress, deferred, decided. Decision: none, approved, changes-requested, cut-from-scope, historical-context-reviewed. Questions/concerns have their own open/resolved/dismissed state with the resolution kind recorded. Currency: exact reviewed revision, provably unchanged, unchanged with dependencies uncertain, needs-review, correspondence-unknown, source-unavailable. Exclusion is a manifest disposition with a reason, not an approval.
Document progress is derived; final sign-off is an explicit action with three outcomes. “Review completed, approved” requires every primary span accounted for and every required decision/question resolved under the signed scope. “Review completed, changes required” requires every required chunk examined and records that some are rejected, cut or awaiting changes; it closes the examination, preserves those decisions and follow-ups, and is never rendered as approval. “Completed with exclusions” discloses the excluded scope. Approval coverage measures unique primary spans or source units within the must-review tier; spot-check and context-only tiers are reported separately and never count toward must-review completion. Historical-context review contributes to reviewed coverage, not approval of current requirements. Inherited unchanged coverage retains the original decision identity/date.
Resolving a finding is a distinct event from the decision that raised it. Clarification and dismissal close the finding; acceptance of a revised requirement is a new decision on the new revision's span, and the original changes-requested or cut-from-scope decision on the old revision stays as recorded. Correcting a decision appends a replacement event and recomputes affected current projections; old sign-off evidence remains readable but no longer advertises current complete approval. No database delete or silent history rewrite is used as Undo; the client-side undo window cancels an uncommitted submission only.
4. Source anchors and chunk generation#
Implement a versioned canonical review projection for block and rich-text formats. Retain original source/revision/hash and renderer/projection version. Within that immutable projection, anchors include document ID, revision, source block/node or section occurrence, UTF-16 range where needed, exact quote/hash and surrounding-context digest. Validate surrogate boundaries. The API's existing text offsets use UTF-16; do not mix bytes, normalized Markdown offsets and rendered DOM offsets.
Source block IDs and section:N indexes are only anchors within their known revision. Renames, reimports and duplicate headings can change or collide; mapping across revisions requires explicit evidence. Markdown parser segmentation respects code, diagrams, tables, list parents, footnotes and reference definitions. Rich-text segmentation preserves node ancestry and semantics. Media/context references include immutable IDs and hashes when available; unknown mutable external context is marked unverified and cannot support an automatic equivalence claim.
Generate structure-based candidate units deterministically, then group them by coherent decision. Each chunk receives a deterministic default review question by document kind (PRD, ADR, architecture/specification) and a deterministic claim label per span from versioned heuristics (modal verbs, tense, references to code paths, bound tasks and their states); suggested semantic grouping, titles, specific questions and label refinements may come from admitted assistant tooling but must pass exact source-reference, coverage and size validation and are labelled as assistance. A model does not decide truth, approve content, exclude requirements or certify unchanged meaning. Deterministic fallback remains usable if assistance fails. Large units receive an over-budget label and staged-reading suggestion.
The estimator is a versioned formula over words, tables, code fences, diagrams and cross-references with per-kind multipliers. Calibration derives new parameters from the owner's measured chunk durations once a minimum sample exists, records the derivation, and applies to future plans only; existing plans keep the estimator version they were generated with.
Plan edits validate nonempty primary coverage, no unintended primary overlap/gaps, dependency validity, cycle reporting, lineage and estimate bounds. Shared context may overlap without duplicate credit. Split/merge creates new chunk identities and preserves prior decisions on their original subjects. Do not automatically promote all children or a merged chunk to approved; derive provably equivalent coverage separately and require explicit decisions where the approval subject changed.
5. Inventory, triage, backfill and sequencing#
Use canonical content reads or a governed service-side indexed census; do not scrape retired paths or query an old SQLite store. Ordinary search pagination sorted by update time is not a stable full-corpus backfill. Implement a stable snapshot/cursor and cutoff strategy with reconciliation for concurrent updates, counts and per-source dispositions. Corpus observed during scoping on 2026-10-03 (not the census): prd 8; adr and architecture more than 100 each; spec 300 to 450; guide more than 1,000 with technical specifications among them; 11,450 documents in total, most eligible ones revision-1 drafts imported 2026-09-20.
First produce a dry-run census: observed documents by kind/format/release, misclassification candidates, duplicate aliases, unknown version associations, missing source/revision/media and historical lineage. Preserve original metadata; review-specific mapping fixes do not rewrite operational content. For each version membership, select the revision by provenance class in this order: linked publication snapshot, release association recorded on the revision, preserved source lineage; only then fall back to the latest approved revision at cutoff or the latest draft/in-review revision, and record the class. Pin selected revisions and manifest hashes before creating personal work.
The census proposes dispositions; the owner's triage pass confirms or changes them through triage commands, in a document-level queue that shows title, kind, memberships, size, provenance, operational state and bound work. Tiers derive from triage: must review, spot check (deterministic sample recorded in the manifest), context only. V2 through V10 product roots default to historical context. The activation read model reports estimated owner hours per tier from chunk estimates before the manifest freezes.
Resume bounded batches using checkpoints, uniqueness constraints and deterministic source identity. Re-running a batch produces no duplicate campaign documents/chunks. Source read failure stays pending/unavailable; partial completion is never a successful complete migration. An owner can correct uncertain associations and explicitly exclude irrelevant material with reasons.
Sequence by explicit product-version order, tier, required foundations and decision dependencies, then type/order heuristics, owner adjustments and stable IDs. Handle cycle groups and missing prerequisites visibly. The time selector can propose the next suitable unit without mutating global order or worker priorities. Fairness must eventually surface deferred/long items through explicit backlog views.
6. Revision evolution, the content watermark and event processing#
A new source revision emits or is discovered by a durable, replayable event path in the content service. Every relevant change to an existing subject also advances that subject's content watermark in the same transaction: new revision, archival or restore, eligibility metadata, relation changes and referenced-asset changes, for the document and its dependency edges. Document creation, archival or restore, and kind/domain/release/eligibility metadata changes additionally advance the inventory generation of every campaign scope in the same transaction, so material that becomes eligible outside a checked membership is detected at current sign-off (§7). Enqueue bounded work in a transaction/outbox or reconcile content events by stable cursor. Do not perform model calls or scan the whole corpus in source-edit transactions. Background jobs use a separate bounded workload budget, not task-board claims or the automatic worker switch.
Compute source correspondence and current coverage for the affected document. Exact hashes plus unchanged pinned context preserve equivalence for the compared material only; the claim is scoped and labelled. A missing dependency mapping, unpinned mutable context or unverified media yields “unchanged span, dependencies uncertain”, never universal current approval. Semantic alterations, missing mappings, changed media/links or added/removed scope require review. A formatting rule must be narrow, versioned and tested; code whitespace and normative punctuation remain material by default. Cross-document dependency traversal is bounded, deduplicated and cycle-aware, with explanation paths and an explicit unknown outcome for incomplete processing. Catch-up items carry the revision's provenance (task or agent, stated reason).
Batch source edits coalesce by latest target revision without losing historical event evidence. Do not replace the source revision in an active session. Publish a current-coverage projection only when its complete generation matches the watermark and inventory generation; otherwise show the last known coverage plus “updating” or “stale.” A periodic reconciliation repairs missed/out-of-order events without advancing either generation out of order. Baseline manifests change only through explicit extensions/successors.
7. API and authorization contract#
Proposed namespaced endpoints: /v1/eve-content/granular-review/query and /commands, with same-origin authenticated Workspace proxies under /api/content/. Final route names are fixed in EGR.02 and tested through both transports. Existing content/review commands keep their contracts; the bridge invokes the existing content review command with a sign-off evidence reference and adds no new approval command. The one additive extension to existing commands, an optional provenance object on the content edit and successor commands for dispatched revisions (§8), is defined in EGR.02.
Queries cover campaigns, inventory/dispositions, triage queue, plan/chunk/source/context, implementation status per span (from existing task bindings and links), reviewer findings and revision provenance passthrough for the pinned revision, next recommendation, progress per tier, drafts, findings, history, correspondence/catch-up, synthesis, watermark and inventory-generation state and receipts. A bounded decision-visibility read returns submitted owner decisions and rationale for the spans a task is bound to and never returns drafts, notes, bookmarks or undispatched concerns and questions; it is the read task context uses. Pagination is bounded and stable; filters are validated and unsupported fields refused. Responses include subject IDs, source revision, plan/manifest generation, watermark generation, inventory generation, capabilities and freshness.
Commands cover prepare-campaign/plan proposals, triage, activate/revise-plan, adjust inclusion/order, save-session/draft, annotate, decide (approve, request changes, cut from scope, historical-context, defer), correct-decision, defer/resume, resolve/dismiss finding, record-standing-decision (delegating to the board's existing mechanism), prepare/confirm dispatch, acknowledge synthesis, sign-off (baseline or current, with its outcome) and archive/start-successor-campaign. Mutation preconditions include expected subject version, plan version, revision/hash and manifest generation appropriate to the operation. A stale sign-off is always refused server-side. Idempotency keys bind actor + exact action payload; altered reuse is refused.
Revision preconditions identify the immutable revision being reviewed, not an obligation that it remain the document's current head. A pinned baseline decision remains possible after a worker saves a newer revision, with a visible newer-version notice. Sign-off explicitly selects baseline or current-manifest scope. Baseline sign-off checks its frozen subjects/decisions. Current-manifest sign-off additionally verifies, in the same transaction, two things: the content watermark of every document and dependency in the checked membership against the manifest's checked generations, refusing on any movement and naming the moved subject; and the inventory generation of the campaign scope against the manifest's checked inventory generation. If the inventory generation moved, the sign-off evaluates the documents created or changed since the checked generation with the census classifier version pinned in the manifest: with no eligible candidate it records the recheck, advances the manifest's checked inventory generation and proceeds; with eligible candidates it is refused and lists them for triage, after which an extension manifest (include) or recorded dispositions (exclude, historical) advance the checked generation and a current sign-off can proceed. “Current” therefore means the fixed membership of the checked manifest, verified against the latest eligible inventory. The receipt names the manifest hash, member watermark generations, inventory generation, classifier version and any pending candidates. Two matching background checkpoints are not sufficient. Creation and eligibility-change races are integration-tested in EGR.36. The UI never conflates the two receipts.
Only the stable owner in an authenticated Dev session may submit triage dispositions, personal decisions, exclusions, resolution and sign-offs. An owner-authenticated session proves account authority, not instruction: an assistant-executed command must carry execution provenance and an instruction reference bound to the exact action and subject, and is refused without it. Worker/reviewer/delegate tokens are denied even if they carry operational review authority. Preparation and read capabilities are narrowly mapped to existing admitted scopes where sufficient; any new scope or owner configuration requires the owner's existing administration process, not self-granting. Assistance receives selected authorized context rather than a blanket private-note read capability.
Use existing session, CSRF/origin, no-store, domain-command access and error envelope conventions. Unknown result after a timeout retains its request receipt for reconciliation; a definitive refusal does not freeze navigation. UAT is read-only; personal data is excluded from ordinary publication export by default. If a deliberately selected read-only personal snapshot is later wanted, it needs explicit permission and a separate tested contract.
8. Concern dispatch and cross-service consistency#
Prepare a reviewable dispatch payload. Confirming it creates one durable operation bound to a payload hash and source finding. Use a deterministic request identity for each downstream action and retain it through crashes. Board/content retries cannot create duplicate tasks, revisions, successors, proposals or standing decisions. Partial completion is visible and recoverable before another dispatch is permitted.
What the existing commands allow today (apps/oshun/eve-content/src/service.ts; tools/eve-task-board/lib/document-proposals.mjs; recheck current main): a new draft revision can be saved on any document that is not archived, not an approved ADR and not a preserved task source; the successor command is accepted only for an archived document or an approved ADR and requires a note; the board's document-proposal API accepts new prd, spec and adr proposals only and creates a new document, not a revision. Dispatch maps onto these by subject kind and lifecycle state; no source is archived merely to satisfy successor eligibility.
| Subject | Proposed revision, including cut-from-scope | Alternative |
|---|---|---|
| Draft or in-review document of any kind | A new draft revision of the same document through the existing edit command, authored by the dispatching identity, carrying the predecessor revision, hash, spans and finding id in the provenance object | A later task in the owning family |
| Approved prd, spec, architecture or other non-ADR kind | Default: a later task in the owning family carrying the predecessor binding and the proposed change. A direct new draft revision is offered only when its state effect (the document's current revision becomes a draft pending review) is shown in the preview and explicitly confirmed | — |
| Approved ADR or archived document | The existing successor command, with the predecessor binding in the required note and in the provenance object, plus a references relation to the predecessor revision | A later task |
| Preserved task-source document | A later task only; the source is immutable | — |
| Concern asks for a new document | The existing document-proposal API for prd, spec and adr; a later task for other kinds | — |
One governed, additive extension is required and is defined in EGR.02: an optional provenance object on the content edit and successor commands (finding id, predecessor revision/hash, spans, dispatch receipt id), validated and stored on the revision event, under the same authority and idempotency rules as the edit and refused when the subject's lifecycle state does not permit that command. It is integration-tested per mapping row in EGR.15 and EGR.27. Standing decisions use the board's existing mechanism unchanged. For cut-from-scope, the preview lists the bound tasks the superseded span would affect, from existing bindings and links; existing task linking is read/link behavior, not a hidden task edit or lease takeover. Follow-up text carries exact evidence, desired outcome, current-vs-historical applicability and personal-review backlink. Dispatch never invokes ask-question/park merely to record a local question.
Task/review completion feeds status into the finding but never resolves it automatically. Owner resolution references the implementation/revision evidence inspected and records its kind; accepting a revised requirement is a new decision on the new revision and never rewrites the original decision. Any separately requested operational block uses existing permission, review and lease rules outside the personal decision transaction.
9. Workspace design contract#
Visual thesis: a quiet reading workspace using Eve's established typography, spacing and one action accent, with source material as the dominant surface. Content plan: triage queue; review queue and “Continue”; focused source and decision controls; contextual inspector; progress/catch-up and synthesis. Interaction thesis: a restrained queue-to-reader transition with stable focus, fast context disclosure, a brief visible undo window before a decision commits, and a clear saved/decision receipt transition, all respecting reduced motion.
Suggested navigation entry “Elikem Granular Review,” distinct from “Review queue.” Code identifiers stay person-neutral (the namespace is granular-review); the display label carries the owner's name. A namespaced Workspace view holds campaign, version, document, chunk and selected revision in the URL. If implementation adds page.tsx routes, update the applicable code-owned walkthrough matrix in the same commit. The existing review queue shows, for PRD, ADR and architecture revisions, a “review in chunks” link and the personal coverage summary with exact-source links, and after a completed current-manifest sign-off offers its own existing approve action with the sign-off as evidence. That queue remains the only approval surface; the personal layer adds none.
Queue uses a readable list with version and tier grouping and counts; the triage mode is a distinct list with per-document dispositions. Avoid a dashboard-card mosaic. Reader reuses source rendering, accessible diagrams, code/table support and the planned continuous-document/diff work. The inspector shows implementation status per span (bound and linked tasks with states), claim labels, the independent reviewer's findings and operational review history for the pinned revision, revision provenance, and related requirements or suspected contradictions side by side before the decision. Desktop context is collapsible; mobile uses disclosures and retains keyboard/focus semantics. Review plans use explicit split/merge/reorder controls with keyboard alternatives. Action wording always identifies the personal layer; the bridge's operational action is labelled as the existing operational approval.
Reuse shared UX work: DC.02 revision diff; DC.05 long-document reader; DC.15 accessible diagrams; M0.04 draft protection; M0.05 response/subject binding; M0.07 revision URLs; M0.08 error/session recovery. The new feature adds domain behavior, not duplicate implementations of those prerequisites. Optional integration with a future Decisions landing page must not force completion of that whole initiative.
10. Reliability, resource isolation and measurements#
Personal review is available without a model call. Backfill and correspondence jobs are bounded, cancellable, restartable and rate limited; no full-corpus reads on queue load. Index owner/campaign, version membership, tier, pending outcomes, document/revision, dependency edges, watermark generations and source-event cursors. Render long sources lazily without hiding their existence or omitting search results.
Initial test targets, to be measured and ratified in EGR.01/EGR.27: a dataset at least as large as the live eligible corpus and a synthetic case at five to ten times the measured live corpus, including expensive documents, event storms and concurrent operational load; queue read p95 at most 750 ms server-side, decision write p95 at most 1 second excluding network; meaningful reader content within 2.5 seconds in the agreed throttled browser profile; background work adds no more than 10% p95 latency to representative operational content/board requests under the same controlled load. The earlier 30,000-document/300,000-chunk/1,000,000-event case is retained as a later capacity exercise, not a release gate. Record hardware, cold/warm conditions, concurrency, query plans and errors. Targets are proposed acceptance budgets, not measured current capabilities; revise only with evidence and owner agreement.
Measure save failures, mutation conflicts, duplicate suppression, source gaps, event lag, watermark lag versus projections, unknown correspondence, dispatch uncertainty, queue latency, job resource use, and estimator error (estimate versus measured chunk duration) for calibration. Do not log full private notes, prompts or secrets. Use the existing observability system; no new external analytics dependency. Retain durable decisions/sign-offs according to existing backup policy and disclose any unresolved retention requirement before rollout.
11. Verification contract#
Core unit/property tests: lossless anchors and Unicode, total/nonoverlapping primary coverage, duplicate headings, nested lists, large tables, fences, rich text, version ordering, per-membership revision selection and provenance classes, tier denominators, cycle handling, plan lineage, state reducers including the three synthesis outcomes and finding-resolution kinds, claim-label heuristics, estimator calibration, progress denominators and material-change detection. Include CRLF/BOM, reordered sections, meaningful punctuation/code whitespace, missing media and unsupported formats.
Real PostgreSQL service/integration tests: migration on fresh/existing databases, repeat/restart backfill, owner access matrix including the decision-visibility read (decisions readable, notes never), execution-provenance refusal for assistant-executed commands without an instruction reference, direct-write refusal, idempotency, stale preconditions, concurrent tabs, transaction rollback, immutable event/sign-off history, watermark advancement for every covered change type and current sign-off races against relation, archival and asset changes, outbox crash recovery, versioned projections, missed-event replay, findings/dispatch against the real task bridge including cut-from-scope successor proposals and standing-decision capture, and UAT read-only behavior.
Non-interference suite snapshots existing document states/revisions, operational review state, task readiness/leases, automatic switches and publication selection before/after every personal action, including triage and the bridge's link and coverage display; prove equivalence and test failure of the entire granular domain while operational flows continue. Only an explicitly confirmed handoff, or the bridge's explicitly clicked existing approve action, may produce its documented downstream action.
Browser tests in apps/oshun/eve-content/tests: all PRD scenarios P01–P15 through real service fixtures, plus delayed responses, dropped replies, stale manifest, session expiry, reload/back navigation, cross-device resume, mobile viewport, keyboard, screen-reader semantics, axe, long-source search, failed assistance, undo window, and current-vs-baseline and approved-vs-changes-required labels. Extend relevant existing Workspace, review, task, relationship, publication and read-only regression specs.
Owner pilots (EGR.34, EGR.35) have their own acceptance criteria recorded as evidence: orientation, chunk coherence, decision effort versus estimate, interruption recovery. Run affected lint/typecheck, core/client/MCP contract tests, isolated PostgreSQL suites, production build, production-backed Playwright and security/performance/recovery checks. Use existing verify-content.py and verify-content-browser.py where appropriate; local-only component tests are insufficient. Provision missing test setup when feasible. Follow host resource gates, serial expensive targets, disposable databases, and log exact commits/commands/results. Do not run these suites against production records.
12. Rollout, recovery and ownership#
Product workers own core/server/UI/test implementation and Eve-authored guides/evidence. Platform-ops owns host resources, deployment, release-controller integration, migrations/backfill execution, backups and recovery drills, and prepares the owner-only flagged exposure the vertical-slice pilot needs. Reviewer/Elikem approval and publication permissions remain separate. Operational tasks are filed as platform-ops, not assigned to product workers by implication.
Sequence: reviewed contracts; isolated feature-disabled schema and owner authority as one milestone (EGR.03 is never exposed without EGR.04); deterministic anchors and plans; owner preview of generated chunk plans over about ten real documents (EGR.34) before any UI investment; the core personal loop, drafts and recovery; platform preparation (OPS.01) and the vertical-slice owner pilot on a small frozen campaign with baseline synthesis (EGR.35), with current coverage labelled unavailable; evolution, the content watermark, current sign-off and the operational-approval bridge; dispatch, synthesis views and assistance; fixture and full-corpus dry run; independent verification; owner pilot from v0 across one v1.x update; platform-controlled Dev exposure; explicit tested UAT promotion if requested. Enabling a feature flag does not start the historical backfill or activate the task worker switch implicitly. Pilots are feedback gates, not releases.
Rollback disables entry points and background processing while retaining immutable personal records. Do not drop data to downgrade. Rehearse backup/restore and projection rebuilding on disposable environments, including source-revision access after restore, watermark rebuild from content history and pending dispatch reconciliation. Quotas/rollout thresholds and operator actions live in an Eve runbook. Completion requires implementation commits on origin/main, passing applicable checks, reviewed required content, recorded pilot outcomes and deployment evidence; a planned rollout is not a deployed feature.
Related Eve records#
- Product scope and acceptance.
- Delivery sequence and coverage.
- Existing Workspace UX audit, revision 3 read on 2026-10-03; existing tasks are reused without rewriting that audit.
- Owner pilots: chunk-plan preview and vertical-slice pilot.
- Final feature acceptance task.