Factory docs, home
Page navigation

Pre-release documentation. Proposed designs are labelled in the page; acceptance evidence is maintained in the build review. Local paths appear as placeholders. Current state: Build review.

2 October takeover commission: Codex resumes Claude’s work and owns PR/merge decisions; Astra xhigh independently reviews uncertain changes. Historical Opus/Fable holds and stop instructions below are superseded for this commission. Current receipts, source and verified checks decide readiness. Live credentials and runtime qualification retain their separate gates.

Written plan, not an acceptance claim. The owner commissioned bounded implementation and parallel independent reviews on 30 September 2026. On 1 October 2026 the owner had Opus 5.5 review the stacked drafts #11–#54 in place of the blocked Fable review, then authorised their merge to main. The retired Fable acceptance assignment is superseded by the 2 October takeover; independent current-source review and named acceptance evidence remain required. Current run state lives in the handover and the owner's status record, not in this plan. This plan authorises no live operation or provider qualification.

This index covers every scope added in the adoption conversation: E28–E35, Temporal Planning E36, Operational Signal Intake/Repair E37/E38, Architecture/composition E39–E41 and Activity levels E42. It supplements the epic acceptance and estimates, which remain authoritative. It does not absorb the separate Knowledge or D4/R7 work. Inspect and coordinate with the existing delivery-d4 work at 42d2700 and Knowledge draft at 52096ff; neither may be overwritten or reimplemented from this plan.

Read and reconcile before choosing work

  1. Read working instructions, project intent, the task capsule and handover. Handover and programme #2 identify the latest source branches, exact commits, draft PRs, reviews and remaining gates. Private conversation history is not a prerequisite.
  2. Match the intended source checkout and clean/dirty state to that handoff. New drafts branch from main. Do not implement an apparent gap twice because main or the historical 23-family inventory predates a draft. Read its exact package manifest, guide, receipt and review disposition first.
  3. Read the linked Module reference and relevant source/tests. Existing names in the packets are source APIs; a source class is not automatically an exported package contract. Names marked proposed do not exist yet. First define and independently review their versioned contract; do not import an imagined API.
  4. Take one bounded packet, preserve other worktrees and record prerequisites, current authority and proof gaps before editing. Do not touch live .factory stores. A merge needs independent review and the owner's merge instruction; local checks alone never authorise one. The takeover commission provides merge authority; independent current-source review and the named acceptance proofs remain required.

Scope and order

Historical packet baseline: The scope table and draft-review paragraph below retain the original planning snapshot. Current bounded local stage and Factory concert acceptance is recorded in the joint qualification; exact-main CI, complete Workflow adoption and live-provider/customer gates remain separate. Original packet requirements, estimates and evidence pins are unchanged.

ScopeWhat exists or is plannedNext bounded action
E28 foundations/resourcesExisting source plus separate distribution/resource draftsA01: reconcile each exact draft and remaining package proof; preserve old resource histories
E29 context/projectionsGuidance, Attention/digest, Owner policy and Branch health have separate draft work; Portfolio/Dashboard adoption remains a distinct packetA02, then P01/P02 where latest handoff still shows a gap
E34 guidanceCompatible eight-value draft and migration proof are separate from initial adoptionA03: preserve old pins, verify latest reader/write limits, retain live migration gate
E30 ImplementationLocal worker candidate; author native checks passed; installed/final review and acceptance openW30
E31 VerificationComplete independently adoptable Workflow still plannedW31
E32 ResearchCollector and discovery schedule exist; complete on-demand Workflow is plannedW32
E33 MarketingFact-linked complete Workflow and publishing adapter are plannedW33
E35 FactoryComposition of the original 23 families and four complete Workflows remains unprovedC35 after their independent consumers pass
E36 Temporal PlanningT36.1 evaluator draft, separate independent review and coordinator checks passed; no forecast or dispatcherT36 packets below
E37/E38 operational workIntake and Repair contract; no production connectors or Repair Workflow implementedO37/R38 packets below

The original E28–E35 estimate remains 24–38 reviewed increments. E36 adds 5–8; E37/E38 add 8–13; E39–E41 add 13–20; E42 adds 7–10; combined 57–89, low confidence. These packets organise that scope; they add no new estimate or calendar promise. Shared generic delivery work is counted once in E36. E35 retains its original dependencies; newer modules are optional future extensions.

Branch health, Discovery/Service, Repository/Gate/Delivery and Producers have reviewed runtime drafts. D01 installed documentation and P01 Portfolio adoption now have independently reviewed implementation drafts; coordinator integrated checks passed. P02 Dashboard and T36.1 dates passed separate independent source/package reviews and 1,863 full integration checks; current draft evidence records their scope. All four complete Workflows remain unimplemented. Capacity and corrected operational-signal intake passed separate independent source/package review; their combined local validation passed 1,926 checks; separate final documentation delta review passed. Material Timing passed separate independent source/package review; 1,945 combined local checks passed; separate independent integration delta review passed. Exact draft heads and gates belong in current evidence and handover tracking; historical receipts remain unchanged.

Reconcile foundations first. Portfolio and Dashboard can be designed in parallel with pure Workflow contracts, but serialise package manifest/compiler changes and shared generated output. Implementation and Verification can proceed independently once their required Module boundaries are proved. Research and Marketing accept external inputs, so neither waits for a completed Factory planner or release pipeline. Temporal calculations and operational intake contracts can be developed independently when recommissioned; both must share one reviewed delivery protocol. Repair composes the Implementation/Verification contracts, source intake and independently supplied release/measurement facilities.

Common packet contract

For every future packet, the implementation record must contain:

  • Exact base/branch and prerequisite evidence; current exported API versus proposed API; owning files and allowed graph edges. Preserve source parser/error/class identity where compatibility requires it.
  • Typed bounded input/output, failure states, caller-owned identities/storage/clock/configuration, trust/authority scope, stable logical operation and Command IDs, canonical payloads and explicit protocol versions. Import starts no work.
  • One useful external consumer of the actual private tarball, a substitute collaborator where applicable, strict type compilation and exact compiler-derived runtime/declaration/resource closure. No private imports or unrelated Factory setup.
  • Focused normal and fault cases: invalid/cross-Product input, replay/reopen/corruption, concurrent writers, cancellation/revocation, stale evidence/pins, exhausted Budget and unknown effects. Add the packet-specific failures below; no simulated outcome can become live proof.
  • Original historical records and receipts retained byte-for-byte. New source/resource bytes legitimately invalidate old current pins; parse history through recorded protocol identity, refuse new dispatch on mismatch and never silently repin or migrate.
  • Independent review and additive evidence: exact artefact SHA-256, source hashes, environment, commands/results and limits. Store bulky logs in the archive; keep published evidence free of secrets and machine-specific private links. Root coordinates full checks, generated docs, stack integration, push and draft PR status. Do not count skipped CI as passed execution.

A packet is complete only when its useful consumer, failures, compatibility and substitution proof pass and its documentation states the remaining limits. Acceptance or a live effect needs its own gate. If a collaborator is missing, return a durable typed unavailable/wait state; do not widen authority or invent success to finish a demo.

A01 foundational adoption reconciliation

Inputs: latest handoff, inventory, package.json, tsconfig.modules.json, scripts/build-modules.ts, test/distribution.test.ts, corresponding test/helpers/*-distribution.ts and adoption guides/receipts. Output: exact per-family coverage with public entries, closure, compatibility, independent review and Fable status. This is reconciliation of existing drafts before further implementation.

Family groupExisting files/contracts to inspectRequired remaining proof if absent from the latest draft
Journal; Execution; Schedule; Intelligencejournal.ts, execution.ts, lifecycle-schedule.ts, intelligence.ts and their guidesCaller SQLite/clock, passive imports, exact replay/versions, epoch/Budget refusal, schedule admission before contact and unqualified model refusal
Executor; worker; fence/workspace; Evidence/Assurance; guarded recoveryconfined-executor.ts, local-worker.ts, fixture-manifest.ts, fence/workspace and Evidence/guarded sourcesActual launched JS closure, missing/mutated/linked resources before launch/write, old source-mode records, cancellation/ownership/restart; never use TS sidecars as executed-code integrity
Repository; collector; Repository Assurance; gate/recordsrepository.ts, repository-protocol.ts, repository-collector.ts, repository-assurance.ts, product-gate.ts, delivery-records.tsCaller-owned repositories/snapshots and genuine Evidence branding; exact Candidate export; revocation and cross-Product refusal; unknown integration outcome reconciliation; recorded old/new protocol pairings
Producersproducer.ts, claude-producer.ts, openai-producer.ts; Producer referencePassive public contract and neutral types, substitute bounded fixture transport, cancellation and unknown remote use, secret canaries and qualification refusal; packaging never qualifies a live provider
Discovery / Service accessDiscovery core/facade as present in latest draft, discovery-text.ts, Service sources and reporting typesCaller cache/transport, original journal and class identity, citation expiry/robots/byte bounds, source facade compatibility; names-only service inputs, partial/unknown coverage and no authentication effects

Run the existing source suites named in each agent reference plus the matching actual-package helper. Keep earlier consumer assertions when combining drafts. A private packed artefact and a clean import are necessary but do not close the whole epic. Any additional source refactor must be justified by an observed external-consumer barrier.

A02 and A03 project adoption and guidance

A02 reconciles the separate Guidance, Attention/digest, Owner policy and Branch health drafts before P01/P02. Inspect each public core versus Factory facade, neutral reporting/attention/branch contracts and original records. Inputs remain caller-owned identities/facts/readers/clock. Outputs preserve exact error identity, provenance, unknown/partial Counts and old replay. Re-run the specific adoption helper and its source boundaries; reject stale/future owner coverage before new digest writes and missing GitHub pages as unknown. Branch health defaults and optional Portfolio mirroring belong in the Factory facade, not mandatory public setup.

A03 is E34's compatible guidance work, not an unrelated values redesign. Read the current project-guidance core/facade, Dashboard editor and saved-reader grammar. Validate both old six-value history and new eight-value writes without changing old digests/pins, including concurrent saves and uncertain acknowledgements. Use disposable data; a live migration remains a separate authorised operation with readback. A planner consuming guidance is still future work and cannot be claimed from successful storage/editor tests.

T36 Temporal Planning sequence

The full contract defines the records, generation fence and failure matrix. Proposed implementation files are src/temporal-planning-contract.ts, src/temporal-planning.ts and a separate src/command-delivery.ts; confirm ownership/names before creation. The draft T36.1 value-only API now lives in the first two files; its independent source/package review and coordinator integration checks passed. No command-delivery API exists. Reuse neutral contracts and keep Factory/provider dependencies out.

PacketInputs → outputsCompletion and failure proof
T36.1 dates and dependenciesVersioned Date claims, constraints, caller clock/policy → immutable plan with conflicts/unknownsPassive typed external evaluator; month/quarter/range, withdrawn/conflicting sources, cyclic/cross-scope dependency and unsupported time-zone refusal
T36.2 capacityComparable reviewed work and unfinished samples; account/window/model/effort observations, reserves → forecast range or unknownShared windows all constrain admission; no percent-to-token fiction, duplicated quota or renewed ceiling; sparse/stale data, human review delay and forecast calibration
T36.3 material timingCurrent facts/readiness, audience/channel/material policy and calendar → candidate windows with basisDST gaps/folds, embargo/quiet/expiry boundaries, missing audience evidence and no unsupported best-time/availability claim
T36.4 durable proposalsPlan revisions, distinct Commitments and causal Signals → immutable Events and Command generationsLogical-operation identity, before-offer atomic supersession, offered/unknown inbox freeze, retired-generation rejection, bounded catch-up/cancellation
T36.5 delivery/compositionJournal outbox plus commissioned receiver/authority/admission ports → durable acceptance and independently tracked effectsCrash at each send/ack boundary, current revocation, duplicate/late generations, unknown-effect reconciliation and cascade caps; substitute receiver, no automatic publishing

Each packet extends a proposed test/temporal-planning.test.ts/delivery fault suite and an actual tarball helper. Split capacity/delivery increments if observed complexity requires it; keep total estimate and uncertainty explicit rather than weaken proof.

O37 and R38 operational sequence

Use the operational contract, including its source/metric/production matrix. Proposed files are src/operational-signals-contract.ts, src/operational-signals.ts, separately owned source adapters, src/repair-workflow.ts and corresponding tests/package helpers. Do not create service-specific authority inside Temporal Planning or silently grow the metadata Service module.

O37.1 now has a bounded implementation candidate, with independent source/package review passed after the recorded-admission correction; its finite ledger and immutable-retention eligibility do not close E37. It takes a commissioned source binding and bounded redacted observations, and returns durable accepted Signals/checkpoints/coverage. Prove auth/scope/privacy, atomic cursor progress, duplicate/out-of-order input and missing pages. O37.2 adds one crash and one metric adapter with grouping/revision reconciliation and explicit read bounds; substitute a context/source adapter. O37.3 validates query/distribution/denominator/sampling, fixed and trend detector policy, low-volume and slow-drift failures. Contract fixtures do not claim real Sentry/PostHog/Datadog access; later commissioned adapter qualification records the exact source behaviour.

R38.1 takes Signals and authority/Budget ports and creates case/episode state, autonomous waits and restore-first routing. R38.2 freezes a reproducible pre-change Resolution contract and independent acceptance; absent evidence stays inconclusive. R38.3 composes W30/W31 through their public inputs/results and tests stale Candidate, wrong scope and unknown worker effects. R38.4 composes a separately proved release/recovery port and production evaluator: exact artefact/cohort, adequate traffic/window/coverage, guardrails and failed/unknown rollout. R38.5 performs separately authorised source feedback with readback, unknown acknowledgement reconciliation, grouping/triage races and new regression episodes. Required Sentry feedback blocks episode closure while unresolved. Each step has a disposable external consumer before the complete substituted-ports exercise; no local fixture counts as production recovery.

Provider event integration

The optional provider event adapter has four bounded packets: real event projection, authenticated subscriptions, notification delivery and admitted tool/job composition. EV1's single Branch health poll projection contract is accepted for bounded implementation; its source/package draft passed separate independent verification and 1,885 coordinator integration checks. The original EV2a subscription integration passed 1,976 full and 198 package/docs checks with independent combined package review. Its native-data correction passed independent source/package review, and the corrected stack at 6f4ff02 passed 2,045 full checks; fresh combined package/public acceptance remains separate. Reuse T36 delivery ownership rather than build a second dispatcher; MCP receipt cannot satisfy its accepted-command contract. These packets add no live commission, model transport or E35 prerequisite. They have no epic, estimate or consumer yet: EV1, EV2 and EV4a candidates exist, but no reachable MCP host, sender or end-to-end use. The owner decides where this work is tracked and how a live host would be reached before further EV packets; shared delivery is counted once.

C35 original Factory composition

Inputs are the exact adopted artefacts and independently proved W30–W33 contracts, plus declared Factory identities/authority/configuration. Output is Factory wiring through those same public contracts and a reconciled 23-family/four-Workflow coverage table. Likely owning files include the existing src/local-factory.ts and source Factory facades; inspect ongoing D4/R7 ownership before any edit and coordinate with its owner.

First replace one concrete dependency with its public caller-owned contract without changing runtime behaviour. Then substitute one complete Workflow and repeat its independent consumer unchanged. Exercise wrong-Product inputs, drift/revocation, unavailable collaborators, restart/cancellation and unknown effects at every composition boundary. Enforce exact graph/capability edges and source/resource closure, run focused integration tests, then let the coordinator run the serial full suite and generated output checks. Close no E01–E27 outcome from this local proof. Knowledge and D4/R7 keep their own review and authority boundaries.

AL Activity levels sequence

Registered E42 issue #82 and the independently reviewed design define proposed scope. All packets are unimplemented; no level, schedule, grant or budget ledger is activated. Preserve E39–E41 and existing historical receipts.

PacketBounded workRequired proof
AL1aNeutral contract, pinned policy and pure resolverPacked external consumer; complete week history and per-class evidence
AL1bJournal register and saved-request set/showIn-lock clock, rollback refusal, replay, isolation and corruption
AL5Bounded digest and read-only dashboardCurrent pins, unknown coverage and no writes
AL4Versioned schedule and commissioned Source selectionExplicit full rhythm, selection pin, Capture novelty and original history
AL7Shared admission protocol and authoritative budget ledgerConcurrent Products, cap changes, crashes, unknown charges and no redispatch
AL2a and AL3First bounded proposals and independent adoptionTyped floors, deduplicated owner asks and substituted consumer
AL6Owner queueAttributed counts, one urgent Product item and silence inert
AL8Branch/renewal/delivery/Marketing integrationsConsumer-owned class/purpose/evidence and grant boundaries
AL2bSunset and remaining proposals when neededPinned notice/data plan and complete source windows

AL9 guarded dashboard writes are optional and outside the estimate. Live defaults, floors, shares and reserve sizing need owner decisions before commission; static prototype checks do not prove runtime enforcement.

Source: docs/implementation-plan.md