E36 design; the bounded T36.1 source/package draft passed independent review and coordinator integration checks. Temporal Planning relates dated evidence, feature dependencies, delivery capacity and communication windows. It returns inspectable plans, accepted plan Events and typed Command proposals for an external caller. The name describes its responsibility; it does not select the Temporal workflow engine.
The owner commissioned this design and epic on 30 September 2026. A future implementation must pass the independent adoption gate. E36 adds scope beyond the fixed 23-family audit and E28–E35 programme. It does not change the original customer-value contracts or authorise a live schedule, provider contact or publication.
Responsibility and neighbours
Planning owns value, priority, Slice selection and commitments. Temporal Planning supplies temporal constraints and forecasts to those decisions. A competitor announcement can make a review timely; it cannot establish customer need, select a Slice or grant Authority.
| Existing part | Current responsibility | E36 boundary |
|---|---|---|
| Calendar | Pure civil-date, ISO-week and zone arithmetic | Reuse compatible arithmetic; no capacity, forecast or marketing policy exists there |
| Lifecycle schedule | Discovery occurrence identity, admission ledger, deferral and replay | Keep its current grammar, IDs and Budgets; an optional adapter consumes planning proposals |
| E14 | Evidenced priority, rejected alternatives and the next valuable Slice | Combine timing with customer/business Signals; retain selection ownership and active acceptance |
| E26 | Retirement, notice, migration and recovery obligations | Represent those windows as constraints; never shorten an obligation to improve a forecast |
| E32 Research | Attributed findings and proposals | Optional source of date claims and competitor Signals; external callers may supply other conforming evidence |
| E33 Marketing | Fact-linked material and optional authorised publishing | Consume suggested windows and confirmed readiness; keep claim validation and publishing Authority |
| E35 Factory composition | Existing Module and four Workflow adoption proofs | Stays independent of E36; later Factory wiring uses the same public planning contract |
A T36.1 draft implements pure attributed date/dependency evaluation with caller values and immutable results; separate independent review and current integration checks passed. The bounded capacity candidate passed separate independent source/package review; 1,926 combined local checks passed; separate final documentation delta review passed. The bounded Material Timing candidate passed independent source/package review; 1,945 combined local checks passed; separate independent integration delta review passed. The bounded proposal ownership kernel passed 2,020 combined local checks in its original integration. Its returned-data correction passed independent source/package review, and the corrected stack at 6f4ff02 passed 2,045 full checks; Root's earlier 2,121-check combined full run and independent same-package Journal/EV2/notification-ownership physical proof passed; final publication review, CI/Fable and live qualification remain open; broader durable planning and reliable command delivery remain unimplemented. calendar.ts is arithmetic; lifecycle-schedule.ts names discovery rhythms and has no general delivery forecast. The journal records pending outbox Events but supplies no delivery or acknowledgement protocol. Those are implementation gaps, not implied facilities.
Proposed public contract
The draft entry point software-factory/temporal-planning now exports only the T36.1 evaluator. The broader operations below remain proposed. Keep the deterministic planning calculation passive and usable without Factory identities, Portfolio, providers, a scheduler or a model. A durable facade may require the public CommandJournal, with caller-owned storage and clock. Declare that dependency and test the same typed API in a real packed external consumer. Use neutral identifiers and versioned collaborators for evidence validation, capacity observations and current permissions; source text and arbitrary caller claims cannot authenticate Evidence or Authority.
All records bind a consumer namespace and Product/scope. They carry a protocol version, stable ID, revision or immutable digest, source/causal references, observation time and applicable validity. The implementation must publish strict parsers, bounded collections/text, typed refusal reasons and compatibility readers before any adoption claim. Unsupported versions, missing identities, dependency cycles, cross-Product data and invalid intervals refuse without partial writes. Unknown inputs remain explicit.
| Record | Required meaning |
|---|---|
DateClaim | Attributed announcement, externally stated expectation or observed occurrence; subject, source identity/digest, quoted or derived basis, published/observed/checked times, precision and interval, source zone, stated confidence and validation status, expiry and retraction/supersession |
TemporalConstraint | A dependency or allowed/forbidden interval, relationship to a dated claim or feature, lead time, owner and whether hard or advisory; includes notice and recovery obligations |
CapacityObservation | Account/window identity, unit, observation and reset times, known range or unknown reason, coverage, concurrent-use scope, admission ceiling and policy revision |
ThroughputSample | Comparable work class, scope and environment, start/end or unfinished age, measured usage/window/model/effort, human/independent review and dependency/rework delay, quality/result and coverage |
CapacityForecast | As-of time, horizon, feasible increment/date range or explicit unknown, sample set and method, assumptions, dependencies, reserves, existing work and expiry |
TimingRecommendation | Material type, audience/channel and zone, candidate intervals, readiness/dependency conditions, policy/evidence basis, trade-offs and confidence or unknown reason |
PlanRevision | Immutable inputs and policy digests, forecast and recommendations, dependency order, unresolved blockers and prior revision; a current pointer never alters old revisions |
Commitment | Separately approved scope/date window, approving authority and limits, supporting plan revision, expiry and withdrawal; no forecast becomes a Commitment automatically |
PlanEvent | A fact accepted by the planning domain, with immutable Event/revision identity and a stable logical occurrence, recorded and effective times, plan revision, correlation/causation and provenance |
CommandProposal | A typed request for another domain to consider, with target contract/version, stable logical operation ID, immutable Command ID and generation, evidence/plan pins, conditions, validity, Budget and causal references |
These are contract obligations, not implemented TypeScript declarations. Proposed operations are evaluate(inputs, policy, asOf) for pure calculation; record, revise, cancel and expire through version-checked journal commands; and bounded evaluateDue(asOf, cursor) for current temporal occurrences. Reading or rebuilding projections never evaluates due work, writes Events or issues Commands.
Dates, uncertainty and capacity
An external announcement is a fact about what its source stated. Its expected release date is still a claim; an observed release needs separate evidence. Preserve imprecise months, quarters and ranges, disagreements, source withdrawals and expiry. Never turn a date-only claim into a precise UTC appointment. A forecast names its assumptions and uncertainty; a Commitment records a separate authorised decision. Revising a forecast cannot rewrite or silently move a Commitment: record the conflict and propose a decision.
Forecast capacity from attributable completed, independently reviewed increments and unfinished comparable work. Include independent review, repair, queue delay, dependencies and interruption; excluding failed or unfinished work would overstate throughput. Record the sample window, size, coverage and estimation method. Sparse or incompatible samples produce scenarios or unknowns, not invented confidence percentages. Compare predicted ranges with later observed completion, usage and quality; revise the next policy/model explicitly without editing old predictions or weakening acceptance.
Subscription capacity is shared across Products and the owner's other sessions. Keep account identity and each overlapping quota/reset window separate. An observed percentage is not a token count or a reserved share. Never add unlike units, assume one reset clears every limit, extrapolate a fixed number of increments from a percentage alone, or treat local allocation as provider reservation. Fresh usable observations can constrain an estimate; missing, stale or conflicting headroom stays unknown and cannot grant dispatch. Current admission remains a separate check immediately before work starts.
Forecasts respect already admitted work, unknown effects debited at their commissioned bound, WIP/dependency limits and explicit incident/product/owner reserves. Deduct reserved capacity once within its declared account/window; do not count the same shared headroom for each Product. Scope or model changes can invalidate comparability. Limits, reserves and stop thresholds are caller policy inputs, never hard-coded Module defaults. They must be commissioned before dependent execution, and exhaustion proposes deferral/replanning without raising a ceiling, switching providers or purchasing capacity. Forecasts must name any scenario assumptions about future resets or replenishment.
Material timing and local time
Suggest days and times per material type, audience, channel and purpose. Release notes need confirmed availability; an explanatory article needs reviewed content and current claims; a launch announcement needs its declared availability/embargo conditions; a retirement notice must leave the approved notice and recovery periods. A planning reminder may request a draft before those conditions hold, but it must not assert a release happened.
Use supplied policy and attributable channel/audience observations to rank windows. Without those observations, return labelled policy-based options and a measurement proposal. There is no universal best weekday or time in this contract. Timing results retain their population/window and observational or causal qualification; track attribution and adverse effects under a frozen measurement contract. Competitor timing can create a conflict or opportunity for consideration, alongside research and business-value Signals.
Represent an instant separately from a local date or wall time. Persist the IANA zone, declared gap/fold policy, resolved instants and resolution provenance with each revision. Existing calendar arithmetic uses compatible gap/fold resolution; E36 must explicitly adopt that policy or reject ambiguous/nonexistent times, rather than infer it. Check DST, date-line/ISO-week boundaries, quiet hours, holidays supplied by the caller, embargoes and expired windows. A changed zone rule or clock does not reinterpret recorded occurrences; a new revision records any changed future resolution. Refuse unsupported zones or absent rules required by the policy.
Events and Command proposals
A plan.recorded, plan.revised, plan.cancelled, plan.expired or window.eligible Event records the domain decision that actually occurred. A forecast date passing does not prove delivery, a competitor release, available capacity or permission. Preserve both effective and recorded times; a late observation is never backdated as an earlier dispatch.
An explicit, versioned rule combines temporal Events with research findings, business-value Signals and current dependencies. It can propose a research job, priority review, approved-Slice execution, marketing draft, assessment read-out or publication request. Operational Signals may also request investigation or reprioritisation; their service connectors, incident/reproduction/repair ownership and source closure remain outside E36. Each intent is a declared discriminated contract with bounded fields and a permitted target; never accept arbitrary executable code, shell strings or an unvalidated destination. Public type identity and semantic compatibility must be independently proved with a substitute receiver.
The proposal contains the full causal set and policy revision, not only whichever input arrived last. A missing required Signal remains unknown; a time window cannot bypass the evidence requirement. Rules may request a decision within current delegation; a proposal to publish still requires the separate current publishing grant, fact checks and Action/Receipt protocol. It grants no provider, spend, credential, release or customer-contact permission.
Keep three identities distinct. A stable logicalOperationId names the intended business operation, independent of polling time, restart, transport attempt or changed plan pins. Duplicate causes converge on that operation. An immutable Event ID names one recorded fact/revision, so successive plan.revised facts remain distinct without requesting the business effect again. A Command ID names one immutable canonical payload and generation of that logical operation. Persist the bounded identity mapping, scoped by namespace and target. Same Command ID and payload replay; changed payload under that ID conflicts.
A revision may replace a proposal only before any delivery has been offered. Supersession and the first delivery-offered marker must use one versioned atomic state transition or an independently proved equivalent fence. Supersession retires the old generation without altering its Event, payload or receipts; the new payload receives a new Command ID/generation under the same logical operation. Commit the offered marker before any send. A crash or uncertain inbox acknowledgement therefore freezes that generation even when the sender cannot tell whether the receiver accepted it.
After offering, never replace a command merely because its effect has not been observed. Reconcile the original identity. A later generation requires durable receiver cancellation or terminal reconciliation proving the old generation can no longer produce an effect, including delayed messages. Keep a durable retired-generation fence. Receiver admission and effect ownership bind the logical operation and authorised generation, as well as the Command ID, so a new ID cannot duplicate the same operation. A completed operation is not repeated by revising its plan; a genuinely new business occurrence must be explicitly identified. Inbox acceptance already crosses the replacement boundary, even when downstream execution has not started.
Delivery and acknowledgement protocol
The optional MCP Events adapter can carry notification projections through shared delivery facilities. Its webhook receipt is a separate transport fact; it cannot satisfy the durable accepted-command protocol below. Do not cast a callback success into receiver acceptance, task completion or a new grant.
The proposed adapter has an at-least-once message boundary. Exactly-once external effects are not claimed. The existing journal can atomically store a plan transition, its receipt and Events; its pending outbox has no acknowledgement API. Implement delivery state in an explicit adapter-owned journal aggregate or another proved atomic store, preserving existing journal schema and historical Events. A raw SQLite sequence is only a cursor within a bound journal identity, never a globally unique Event ID.
- The producer commits an immutable Event/proposal before delivery. The adapter records a bounded per-consumer delivery entry keyed by producer/store identity, Event ID and consumer contract; it binds the source digest, logical operation, immutable Command ID and generation. Re-reading an outbox page finds the same entry. Atomically freeze the generation with the delivery-offered marker before any send, fenced against concurrent supersession. Delivery selection has a bounded lease/ownership fence; time alone never permits conflicting side effects.
- The receiver validates namespace, target contract, payload and the authorised generation, refusing retired generations. It atomically records its inbox entry, logical-operation ownership and accepted domain Command/result where they share a store. If stores differ, declare the boundary: durable inbox acceptance precedes recoverable domain dispatch with the same Command ID. A crash between stores resumes from that identity; no distributed transaction is assumed.
- The receiver acknowledges durable acceptance with its receipt identity and payload digest. That acknowledgement does not claim downstream work completed. Duplicate delivery returns the same acceptance or rejects conflicting bytes. Only a verified acknowledgement advances the producer's per-consumer delivery state; checkpoints cannot skip an unaccounted entry. Rejected, deferred, expired and parked entries keep a durable reason and disposition.
- Immediately before every new downstream dispatch, re-read the current plan/proposal disposition, authority/revocation, ownership, evidence/claim freshness, dependencies, readiness, admission Budget and shared-account limits. Historical replay grants no new contact. Use the target's admitted Command/Action identity and fence; never infer permission from a stored proposal or an old success receipt. The target must define its dispatch/authority linearisation boundary and revocation behaviour; reads across stores do not create an atomic guarantee.
- A lost acknowledgement may cause message redelivery; a lost external result requires reconciliation. Unknown effects stay owned, charged at their bound and undispatched again until the existing target protocol proves a permitted next step. Cancellation/revision/expiry stops new dispatch but does not delete or declare success for in-flight/unknown work. Record supersession and reconcile the original operation.
- Consumer removal and poison messages require a scoped durable disposition. Bound pages, attempts, bytes, pending age, fan-out, causal depth and transitions per tick/root occurrence. Repeated rule/occurrence combinations refuse or park with a reason; rate limits and a maximum causal depth prevent self-triggering cascades. Capacity exhaustion retains a wake condition without issuing a larger Budget.
Catch-up policy is commissioned per intent: skip expired work, coalesce equivalent missed windows into one current occurrence, or enumerate a bounded set where each occurrence is genuinely required. Preserve the missed occurrence references and consumed watermark. Never replay every missed marketing slot by default, renew a deadline by restart, or coalesce distinct committed operations silently. Paused/cancelled plans cannot create new eligible proposals merely because the caller advances time.
Acceptance and increments
A real external tarball consumer must complete a useful dated feature plan, capacity forecast, material-window proposal and bounded stage-trigger exchange using caller-owned evidence, clocks, storage and a substitute receiver. It must not start Factory, a provider or a live publisher. Independent probes must cover conflicting/withdrawn external dates, stale and unknown quotas, shared-account contention, dependency cycles and infeasible commitments, throughput uncertainty, DST/expiry, provenance and cross-Product refusal, crash before/after each durable boundary, replay and concurrent delivery, cancellation/revocation during uncertainty, supersession racing the first offer, accepted-but-unexecuted inbox entries, delayed retired generations, catch-up and recursive trigger refusal. Reconstruct old revisions and predictions exactly. Demonstrate that a combined temporal/research/value cause yields one logical operation without granting publication.
| Increment | Complete bounded proof |
|---|---|
| 1. Neutral plan | Versioned date/constraint/plan contracts, deterministic dependency checks and a useful external consumer with attributed dates and honest unknowns |
| 2. Capacity | Comparable throughput ranges, shared windows/reserves/unfinished work, expiry and feedback, plus independent overcommitment challenges |
| 3. Material windows | Material/readiness policy, local-time resolution and evidence-based ranking, with DST and unsupported-claim cases |
| 4. Durable proposals | Immutable revisions, separate Commitments, stable occurrences and causal Events/Commands; replay, revision/cancellation and bounded catch-up |
| 5. Delivery and composition | Commissioned inbox/outbox adapter, receiver substitution and current dispatch gates; crash/unknown-effect/cascade proof and optional Research/Marketing/Factory parity |
Estimate 5–8 reviewed implementation increments, low confidence; the durable delivery and capacity boundaries may need separate increments. This includes each increment's consumer, faults, review and documentation. It is additional to E28–E35's 24–38, giving an E28–E36 subtotal of 29–46; no calendar duration follows from these counts. Re-estimate after the first neutral consumer and a disposable inbox/acknowledgement crash experiment. E28 supplies foundational adoption; E32/E33 are optional integration dependencies, not prerequisites for a host supplying its own evidence and material contracts. Public effects require separately proved adapters and current authority.
The initial E36 change is this design, ubiquitous language and epic integration only. No planning Module, scheduling service, outbox dispatcher, model qualification or publishing feature is implemented by this document.
The later E37/E38 operational design adds service intake and Repair separately. Its additional estimate keeps this Module and its shared delivery work counted once.
T36.5a receiver source is a bounded candidate: Guidance acceptance applies one exact already offered Guidance command under current authority and reconciles its historical revision. It leaves the permanent T36.4 offer intact. Corrected component/package passed independent review; it was merged on 1 October 2026 after an Opus 5.5 review, Fable acceptance is pending and EV3 transport is separate.
