m/Assistant

Fork 0

Files

T

m 939a5152e1

CI / test (push) Successful in 48s

Details

ARCH-GEN2-000: adopt Gen2 architecture governance

2026-08-29 01:53:27 +03:00

58 KiB

Raw Permalink Blame History

Домовой — GEN2 Migration Map

Дата: 2026-08-29 Статус: migration design / input to GEN2 Master Plan Основание:

DOMOVOY_GEN2_GAP_MAP.md;
GEN2_TARGET_ARCHITECTURE.md;
GEN2_CANONICAL_DATA_MODEL.md;
architecture analysis packages v2 + v3;
фактические V1 contracts/modules/PostgreSQL migrations 001–045.

0. Назначение документа

Этот документ отвечает не на вопрос «каким должен быть Домовой», а на вопрос:

как перевести существующий Assistant V1 в Gen2, не потеряв уже работающие инварианты, данные, recovery и накопленный опыт.

Для каждого V1 owner/runtime/table/package используется одно из решений:

KEEP — оставить смысл и основной путь;
EXTEND — расширить существующего owner без параллельного механизма;
REWORK — сохранить owner, но существенно изменить контракт/модель;
REPLACE — создать Gen2 replacement и переключить на него writer/read path;
RETIRE — прекратить новые записи, оставить history/read compatibility, затем удалить;
ADD — новый owner/runtime, которого в V1 нет;
COMPAT — временный compatibility/read model в период миграции.

Gen2 не является greenfield rewrite. Gen2 не является добавлением второго Assistant рядом с первым.

1. Неподлежащие потере V1-инварианты

Следующие смыслы считаются migration constraints, а не деталями реализации:

LLM != Assistant.
Capability != Authority.
Observation != Claim != Belief.
Intent != Goal != Task.
ActionProposal != AuthorizedAction != ActionAttempt != Outcome.
External mutation проходит только через Authority -> Commit -> Executor.
UNKNOWN_OUTCOME запрещает blind retry возможного повторного эффекта.
Только canonical owner пишет canonical state.
Cross-owner direct DB write запрещён.
Prospective future relevance остаётся canonical в Continuity.
Scheduler/wake mechanism не становится владельцем обязательства.
Human correction не переписывает прошлое молча.
Raw evidence отделено от derived interpretation.
Search/vector/cache/context являются rebuildable projections.
Provider/model types не проникают в domain owners.
Work/Agent execution переживает процесс через durable state/checkpoints.
Producer и evaluator/verifier могут быть независимыми.
Self-development проходит evidence/proposal/isolation/verification/canary/promotion/rollback.
Secret value не является обычной памятью/настройкой/context excerpt.
При неполных данных состояние остаётся UNKNOWN/PARTIAL, а не угадывается.

Если Gen2 task требует нарушить один из этих пунктов, это не implementation detail: нужен новый архитектурный ADR с явным изменением Конституции.

2. PRE-GEN2 — обязательная фиксация текущей разработки

Это выполняется до первой строки Gen2 implementation.

2.1 Цель

Получить единственную, воспроизводимую точку GEN2_BASE_COMMIT, в которой:

все актуальные открытые изменения сохранены;
ничего не потеряно через reset/discard;
текущие ветки сведены в основной поток;
main отражает фактическое состояние проекта перед архитектурной миграцией;
есть tag для полного возврата к V1 baseline.

2.2 Последовательность

git status
-> inventory modified/untracked work
-> отделить secrets/build garbage/generated temp
-> FAST baseline gate
-> commit актуальных изменений в их ветках
-> push рабочих веток where needed
-> controlled merge в main
-> resolve conflicts preserving both valid changes
-> FAST gate on main
-> push main
-> record GEN2_BASE_COMMIT
-> tag domovoy-pre-gen2-<date>
-> push tag

Нельзя

git reset --hard неизвестной работы;
discard чужих/старых незакоммиченных изменений;
force-push main;
«почистить» файлы без понимания их происхождения;
включать secrets/tokens/passwords/build artifacts в baseline commit.

Если FAST baseline не проходит:

сохранить текущую работу commit'ом/backup branch;
исправить baseline;
только после зелёного FAST gate делать merge/tag.

2.3 PRE-GEN2 FAST gate

Не запускать многосуточный full suite.

Минимум:

compile затронутых packages;
targeted baseline tests для текущего diff;
devctl preflight;
devctl check-diff/эквивалент текущей задачи;
быстрый architecture check;
git diff --check.

Full V1 suite используется только если baseline имеет неизвестное состояние или merge затронул критические owner boundaries.

3. Закон миграции writer/read path

3.1 Один canonical writer

В каждый момент у semantic concept есть ровно один canonical writer.

Запрещён переход:

V1 canonical writer -> dual write V1 + Gen2 as two truths

Разрешены:

A. In-place extension

Один owner/table family расширяется новыми revisioned полями/таблицами.

B. Shadow projection

Gen2 projection строится из V1 canonical state, но не является writer.

C. Controlled cutover

V1 writer
-> backfill Gen2
-> shadow read/equivalence
-> freeze V1 writer for this concept
-> switch canonical writer to Gen2
-> V1 becomes read-only compatibility projection
-> observe
-> retire compatibility path

3.2 Rollback

Rollback должен возвращать:

executable/runtime version;
active writer pointer;
schema compatibility;
read path.

Rollback не должен требовать «удалить уже записанные Gen2 данные». Новые Gen2 rows остаются historical/candidate evidence и могут быть проигнорированы старым V1.

3.3 No destructive migration first

На первом этапе запрещать:

DROP TABLE;
destructive column rewrite;
reuse старого поля с новым смыслом;
ID remapping без lineage;
переписывание raw evidence;
массовое удаление old state после cutover.

Retirement — отдельная поздняя задача после observation window.

4. Owner-level migration matrix

V1 area	Gen2 decision	V1 owner	Gen2 owner	Основной способ
Identity	KEEP	Identity	Identity	in-place, минимальные расширения
Authority/Scope	KEEP + CLARIFY	Authority	Authority	старый Scope становится AuthorityScope
Interaction	KEEP + EXTEND	Interaction	Interaction	сохранить append-only ingress
PersonalMemory	REPLACE/RETIRE	Interaction projection	no Memory owner	shadow retrieval -> compatibility retire
Situation project focus	REWORK/REPLACE	Situation	Situation	Focus V2 parallel tables + cutover
Project registry/aliases	REPLACE	Situation	World	backfill referents/aliases
World referents/schema	ADD	—	World	new owner, low-risk first
Epistemic truth bundles	REWORK/REPLACE	Epistemics	Epistemics	normalize evidence/claim/assessment
Experience/Reflection	EXTEND	Epistemics	Epistemics	in-place/versioned extension
Goals/Tasks/Plans	EXTEND	Work	Work	retain IDs + add context/freshness/risk refs
Long-running Work	EXTEND	Work	Work	reuse AgentRun/Checkpoint/Graph
Continuity	KEEP + EXTEND	Continuity	Continuity	preserve prospective semantics
Capability registry	KEEP + EXTEND	Capability	Capability	preserve provider/implementation registry
Skill	REWORK/EXTEND	Capability	Capability	migrate SkillProfile -> Skill revisions
Reflex	ADD	FastPath prototype	Capability	explicit new lifecycle
Dynamic Configuration	ADD + ENV migration	runtime ENV	Configuration	import -> shadow resolve -> cutover
Secrets	KEEP	platform secrets	platform secrets	only SecretRef in Config/Memory
SELF	ADD	distributed evidence	SELF	evidence-only first
Commit	KEEP + EXTEND	Commit	Commit	transition ledger + execution receipt
World outcome verification	ADD/EXPAND	partial Experience	Epistemics	separate receipt from verified outcome
Artifact/raw memory	KEEP + EXTEND	Artifact	Artifact	preserve lineage, add segmentation
Context Compiler	REWORK	cognitive contract/worker	no owner	V2 compiler over retrieval lanes
Memory Retrieval	ADD projection fabric	ad-hoc	no owner	rebuildable projection schemas
Core runtime	REWORK composition	runtime	runtime	shrink giant orchestration function
Cognitive worker	KEEP + EXTEND	no canonical owner	no owner	replaceable reasoning
Research runtime	REWORK/EXTEND	Epistemics adapter	no new owner	freshness/revalidation integration
Work graph runtime	KEEP + EXTEND	Work runtime	Work runtime	long-job semantics
Supervisor/Lifeboat	ADD	—	runtime only	independent release/recovery plane

5. Identity — KEEP

5.1 Current V1

Paths:

contracts/identity/**
modules/identity/domain/**
adapters/postgres/identity_repository.go
migration 001_bootstrap_creator_authority.sql

Current tables:

identity.actors
identity.principals
identity.principal_bindings

5.2 Gen2 target

Identity сохраняет:

Actor identity;
Principal identity;
PrincipalBinding;
authentication/session binding references.

World objects не заменяют Principal/Actor.

5.3 Migration

KEEP in-place.

Не переносить Identity в World.

Допустимые будущие расширения:

identifier history;
aliases/locators linked to identity;
person/world-object link through typed reference, но ownership Identity остаётся.

5.4 Cutover / rollback

Cutover не требуется.

5.5 Tests

FAST:

existing identity binding/session tests.

MILESTONE:

principal isolation при появлении World/ContextScope.

6. Authority — KEEP, Scope semantic split

6.1 Current V1

Paths:

contracts/authority/**
modules/authority/domain/**
adapters/postgres/authority_repository.go
migration 001_bootstrap_creator_authority.sql

Tables:

authority.scopes
authority.authority_grants

scope_id также используется во многих V1 rows.

6.2 Problem

V1 Core слово Scope охватывает слишком много: authority + project/context.

Gen2 разделяет:

AuthorityScope — где действует grant;
ContextScope — где/над чем сейчас работаем.

6.3 Migration decision

KEEP Authority tables. REINTERPRET explicitly as AuthorityScope; do not repurpose their rows.

DB column scope_id может физически остаться именем scope_id в V1 tables, но Go contracts нового кода должны использовать distinct AuthorityScopeID.

ContextScope создаётся в situation, не в authority.

6.4 Compatibility

Existing:

work.tasks.scope_id;
work.goals.scope_id;
continuity.prospective_items.scope_id;
commit.*scope_id/task_scope_id;
situation.focus_bindings.scope_id

считаются references на AuthorityScope, пока явно не мигрированы.

ContextScope добавляется отдельной ссылкой, не заменяет эти значения.

6.5 Risk

Самая опасная ошибка Gen2 — начать использовать old authority.scope_id для Company/DB/Project filtering.

Запретить architecture lint'ом после ADR-G2-001.

7. Interaction — KEEP append-only ingress, EXPAND channels

7.1 Current V1

Paths:

contracts/interaction/types.go
modules/interaction/domain/chat.go
semantics.go
travel.go
task_decomposition.go
personal_memory.go
adapters/postgres/interaction_repository.go

Tables:

interaction.conversations
interaction.utterances
interaction.original_human_intents
interaction.relations
interaction.utterance_interpretations
interaction.project_references
interaction.semantic_candidates
interaction.semantic_candidate_capability_hints

7.2 Keep

Keep canonical:

Conversation;
HumanUtterance;
OriginalHumanIntent;
correction/supersede/clarify relations;
interpretation candidates.

utterances remain append-only source evidence.

7.3 Extend

channel currently DB-checks WEB_CHAT. Gen2 should allow typed ingress:

web chat;
speech transcript;
image/document input;
device/sensor-derived human interaction where appropriate.

Raw audio/image stays Artifact, not large inline Interaction text.

7.4 ProjectReference migration

interaction.project_references is legacy project-centric semantic output.

Target:

contextual referents become typed World refs / Focus signals;
table retained read-only for V1 historical data;
new Gen2 writes use generic typed context-reference contract, not arbitrary project_id.

Do not delete historical project references.

7.5 PersonalMemory migration

See next section.

8. PersonalMemory V1 — REPLACE then RETIRE

8.1 Current V1

modules/interaction/domain/personal_memory.go

Important positive property:

PersonalMemory is explicitly a rebuildable projection over HumanUtterance history, not canonical truth.

Current limitations:

rescans utterances;
extracts only current location and preferences;
truncates preferences;
emits ContextText;
no ContextScope/World refs/freshness/contradiction/negative experience.

8.2 Gen2 destination

There is no new Memory owner.

Replacement:

canonical owners
-> retrieval projections
-> lane retrieval
-> Context Compiler V2

Human preferences/statements remain source evidence in Interaction; structured epistemic interpretations can live in Epistemics.

8.3 Migration stages

Stage PM-1 — freeze semantics

Do not add new permanent fact kinds to PersonalMemory.

Only bug fixes allowed.

Stage PM-2 — shadow retrieval

Build retrieval documents from existing:

HumanUtterances;
Epistemics evidence;
Experience;
Work/Continuity;
Capability.

Compare produced context with current ContextText on existing household scenarios.

Stage PM-3 — read cutover

Core chat uses Context Compiler V2 bundle. PersonalMemory.ContextText() remains compatibility fallback.

Stage PM-4 — retire

After MILESTONE memory validation:

no new call sites to CompilePersonalMemory;
function kept one release only for rollback;
then remove code, not utterance history.

8.4 Rollback

Switch Context Compiler feature flag/read strategy to V1 compatibility projection. No data rollback required.

9. Situation — REWORK to Focus V2

9.1 Current V1

Paths:

contracts/situation/types.go
modules/situation/domain/focus.go
adapters/postgres/situation_repository.go

Tables:

situation.focus_bindings
situation.projects
situation.project_aliases
situation.active_project_focus

Current concepts are strongly project/workspace-oriented.

9.2 Gen2 target

Situation owns:

ContextScope;
FocusFrame;
FocusTransition;
FocusStack/head;
confidence/provenance;
active goal/task refs;
typed World refs;
runtime/session binding.

It does not own World object definitions.

9.3 Data migration

Create new table family:

situation.context_scopes
revisions where required;
situation.focus_frames
situation.focus_frame_revisions
situation.focus_transitions
situation.focus_heads
optional focus_stack_view as projection.

Backfill:

each active_project_focus -> FocusFrame with project WorldRef after World project import;
focus_bindings -> runtime binding + active goal/task/AuthorityScope references;
provenance/confidence preserved exactly.

9.4 Project registry move

situation.projects and situation.project_aliases stop being canonical after World cutover.

They become:

WorldObject(kind=PROJECT/ACTIVITY as applicable);
WorldAlias.

Important: do not assume every historical project is a long-lived Project. Migration records old semantic class as legacy_project metadata until explicit reclassification.

9.5 Writer cutover

Create World project referents.
Create Focus V2 schema.
Backfill FocusFrames.
Shadow Focus V2 resolver.
Compare current ActiveProjectFocus decisions.
Switch Situation canonical writer to Focus V2.
Build active_project_focus compatibility projection if old endpoints still need it.
Stop V1 project-focus writes.
Later retire old project tables only after no callers.

9.6 FAST

explicit focus selection;
no cross-company/database leak;
same V1 project command resolves equivalently;
focus stack push/pop;
no use of request text by low-level runtime resolver where prohibited.

9.7 MILESTONE

Company A/B same UNF names;
temporary diversion then return;
restart persistence;
old project endpoints compatibility.

10. World — ADD narrow bounded context

10.1 V1 source

No general World owner.

Inputs for initial backfill exist in:

Situation project registry;
explicit Interaction references;
selected configured systems/devices later;
Identity links only as references.

10.2 First Gen2 responsibility

World initially owns only:

WorldObject;
kind definitions;
aliases;
typed FacetDefinition;
typed RelationDefinition;
identity-resolution decisions/lineage.

World does not own factual truth of facet values.

Actual observations/claims about objects remain Epistemics.

10.3 Migration

Add schema/role world.

Initial seed:

current Situation projects -> World objects + aliases;
only deterministic, reversible import.

Do not migrate arbitrary chat nouns into World automatically.

10.4 Safety

No SaveEntity(map[string]any). No universal values table. No cross-owner FK.

10.5 Cutover

World object creation can be enabled before Focus writer cutover. Situation consumes WorldRef contracts, never SQL joins World tables.

11. Epistemics truth data — REWORK

11.1 Current V1

Paths:

contracts/epistemics/**
modules/epistemics/domain/truth.go
modules/epistemics/domain/research.go
modules/epistemics/application/**
adapters/postgres/epistemics_repository.go

Important V1 table:

epistemics.truth_bundles

It stores Source/Snapshot/Excerpt/Observation/Claim/Belief/Citation in one denormalized canonical-looking row.

11.2 Problem

Gen2 needs:

independent immutable evidence;
structured Proposition;
multiple Claims against same proposition;
multiple versioned BeliefAssessments;
contradiction/historical applicability;
general freshness/volatility/version binding.

A single truth bundle cannot be long-term canonical representation.

11.3 Target

New table families:

sources
source_snapshots
source_excerpts
observations
propositions
claims
belief_assessments
belief_heads
source reliability records
world outcome assessments.

11.4 Backfill

Deterministically split every truth_bundles row:

source_id -> Source
snapshot_id -> SourceSnapshot
excerpt_id -> SourceExcerpt
observation_id -> Observation
claim_id/statement -> Proposition + Claim
belief_id -> BeliefAssessment
citation fields -> evidence links

Preserve existing IDs where their semantic meaning matches. Where V1 ID combines semantics, create new stable ID and record legacy_ref.

No old truth row is deleted.

11.5 Cutover

Gen2 normalizer/backfill.
Rebuild count/digest verification.
New research writes Gen2 normalized records only.
truth_bundles becomes compatibility view/projection or read-only legacy.
Old read API served from Gen2 projection.
Retire direct bundle writer.

11.6 Rollback

Keep V1 reader available during observation. If writer rollback required, convert only new Gen2 research results through explicit compatibility writer if semantics are lossless; otherwise stop writes and return typed migration block rather than silently dropping information.

12. Freshness / volatility — ADD inside Epistemics, not permanent stale flag

12.1 Existing V1

Research snapshots already have captured_at and coarse freshness.

12.2 Gen2

Add:

TemporalApplicability;
VolatilitySpec;
VersionBinding;
last verification evidence;
TaskFreshnessRequirement;
computed FreshnessAssessment.

12.3 Migration

Existing snapshot_freshness is imported as historical assessment evidence, not as eternal truth.

Do not add a universal mutable column is_stale.

Freshness is calculated at retrieval/decision time relative to current task.

13. Episode — ADD derived episodic memory

Episode is a derived/versioned Epistemics construct, not raw source replacement.

Initial Episode builder can group:

utterances;
work goals/tasks;
relevant actions/outcomes;
artifacts;
temporal proximity.

No V1 data deletion.

Episode projections are rebuildable; stronger future extractors can supersede old episode revisions.

14. Experience/Reflection — EXTEND, preserve V1 records

14.1 Current

Tables:

epistemics.system_expectations
epistemics.experience_records
epistemics.reflection_candidates

Current Experience already references:

original intent;
source utterance;
context refs;
help method;
capability;
skill;
expected/actual outcome;
user correction;
verification evidence.

This is a strong Gen2 foundation.

14.2 Migration

Do not replace V1 experience IDs.

Add V2 expressiveness through versioned extension/linked records:

problem pattern;
attempt;
why attempted;
applicability/conditions;
environment/version bindings;
success/failure explanation;
reusable lesson candidate;
positive/negative classification.

Avoid rewriting old rows to invent unavailable details.

Unknown migrated fields remain empty/UNKNOWN.

14.3 Reflection

Keep ReflectionCandidate lifecycle and leases. Extend triggers/evidence links for:

retrieval miss;
stale knowledge;
source trust failure;
reflex drift;
self weakness.

15. Source trust — ADD evidence-backed assessments

Do not write source.trust = 0.9.

Add:

SourceReliabilityObservation
SourceReliabilityAssessment

Key dimensions:

Source;
Domain/ContextScope;
freshness;
observed history;
evidence basis.

Existing research source metadata seeds Source identity only; no synthetic trust scores during migration.

16. Work goals/tasks/plans — EXTEND, not replace

16.1 Current V1

Paths:

contracts/work/**
modules/work/domain/**
modules/work/application/**
adapters/postgres/work_repository.go

Tables include:

work.goals
work.tasks
plan candidates/plans/steps;
capability requirements;
delegation candidates;
delegations;
agent runs/checkpoints/results;
dispatch queue/leases/callback heads;
graphs/nodes/edges;
result candidates/conflicts/synthesis;
development gates/proposals/scaffolds.

16.2 Keep

Preserve stable IDs and lifecycle history.

Do not introduce a second Job/Agent framework.

16.3 Extend

Goal/Task gain typed refs where needed:

ContextScopeRef;
FocusRef;
FreshnessRequirement;
ExpectedQuality;
risk/cost-of-error/consequence refs;
budget refs.

Old project_id/project_display_name become compatibility fields. New code should prefer Focus/World references.

16.4 Migration

New columns/revision tables are additive.

Do not overwrite old scope_id; it remains AuthorityScope reference.

17. LongRunningWork / Investigation — EXTEND Work

17.1 Reuse

Reuse:

WorkGraph;
AgentRun;
AgentCheckpoint;
dispatch queue;
leases/recovery;
graph orchestration;
cancellation;
synthesis.

17.2 Add semantic work state

Add durable LongRunningWork/Investigation:

objective;
state;
hypotheses refs;
attempts;
findings;
blockers;
checkpoint;
next step;
budget;
pause/resume/preempt semantics.

A long task is not represented as one long model request.

17.3 Cutover

No cutover of old agent execution required.

New Investigation uses existing Work execution engine underneath.

18. Perspective / specialist teams — ADD contract, reuse Work Graph

Add PerspectiveSpec in Work:

objectives;
retrieval profile;
evaluation criteria;
capability requirements.

Do not create fixed “DirectorAgent/DesignerAgent” services.

Work Graph nodes may bind a Perspective dynamically.

Team is selected only if Work planner/evaluation predicts value.

19. Consequence/Risk — ADD Work-owned assessment input

Authority is not risk.

Add ConsequenceAssessment owned by Work/planning:

potential effects;
reversibility;
cost of error;
mitigation;
simulation/dry-run requirement;
recommendation for confirmation strength.

Authority consumes approved concrete action context but does not accept Work's risk score as permission.

No hard capability blacklist is introduced as product philosophy.

20. Continuity — KEEP + additive evolution

20.1 Current

Paths:

contracts/continuity/**
modules/continuity/**
cmd/continuity-runtime/**
adapters/postgres/continuity_repository.go

Tables:

prospective_items
reminder metadata additions;
attention_items
human_responses
action_receipts

20.2 Keep

Continuity remains only canonical owner of:

future relevance;
commitments/prospective semantics;
attention;
human response continuation.

Do not create another scheduler owner in Work/SELF/Memory.

20.3 Extend

Current trigger_at_unix_millis evolves toward typed TriggerSpec:

absolute time;
local-time recurrence;
after-event;
condition;
external signal.

Migration rule:

Existing timestamp reminder -> TriggerSpec.AbsoluteInstant.

Old column may remain compatibility until all runtime paths read TriggerSpec.

20.4 Commitment

If V1 lacks full canonical commitment table, Gen2 adds it under Continuity; existing commitment_id references are preserved even when metadata is initially incomplete/UNKNOWN.

21. Capability registry/model routing — KEEP + EXTEND

21.1 Current strength

Current Capability already contains:

descriptors;
implementations and revision heads;
health;
gaps/acquisition strategies;
usage observations;
provider instances/models/bindings;
qualification/evaluated capabilities;
performance evidence;
task evaluations;
admission leases;
cost/latency/reliability fields.

Do not replace this with a new “AI router”.

21.2 Migration

Extend task strategy selection to consume:

quality requirement;
freshness;
cost of error;
budget;
ContextScope;
Perspective;
Skill/Reflex availability.

Provider/model records remain Capability-owned.

22. Skill V1 -> Skill V2 — REWORK in Capability

22.1 Current

Tables:

capability.skill_profiles
capability.skill_profile_heads

Current lifecycle:

Already stores supporting and known-failure Experience refs.

22.2 Target

Stable Skill identity + explicit Skill revisions:

applicability spec;
dependency versions;
implementation binding;
evidence;
revalidation policy;
promotion history;
supersedes/merge/split lineage where required.

22.3 Migration

Prefer preserving existing skill_profile_id as Skill stable ID.

Existing (skill_profile_id, revision) rows can be transformed/backfilled into skill_revisions.

Do not duplicate active writer:

create Skill V2 tables/contracts;
backfill current profiles;
shadow compare resolver/fast-path decisions;
switch Capability skill writer;
expose old SkillProfile API as compatibility mapping;
retire skill_profiles writer;
later rename/retire physical tables if valuable.

23. Reflex — ADD explicit lifecycle

23.1 Current precursor

FastPathDecision is the correct prototype but not a persistent Reflex identity.

23.2 Gen2

Add:

Reflex stable ID;
revision;
derived_from_skill;
applicability;
implementation;
dependency/version bindings;
evidence;
ShadowPolicy;
RevalidationPolicy;
state.

A Reflex may be code/rule/workflow/small-model path.

23.3 Rollback

Any Reflex must have immediate fallback to Skill/reasoning path.

Reflex failure never removes Skill.

24. Configuration — ADD new owner, migrate ENV gradually

24.1 Current V1 evidence

Runtime composition reads many ENV values:

Postgres hosts/users/passwords;
runtime URLs;
cognitive admission limits;
provider/research endpoints;
allowed hosts;
creator username;
secret backend root;
tokens/file refs.

Some secrets already correctly use CredentialRef/file resolver.

24.2 Gen2 target

New owner Configuration:

SettingDefinition;
revisions/heads;
SettingAssignment;
ScopeSelector;
deterministic resolution;
effective config projection.

Secrets remain in secrets backend and Configuration stores only SecretRef.

24.3 Bootstrap set that may remain outside Configuration

Minimal bootstrap only:

node/runtime identity;
locator for control-state DB/service;
root secret resolver locator/credential mechanism;
emergency listen/bind settings where necessary before DB is reachable.

Everything else should become runtime configuration if it can safely be read after bootstrap.

24.4 Migration phases

CFG-0 inventory

Generate machine-readable inventory of every os.Getenv/LookupEnv, its runtime, sensitivity, default and owner.

CFG-1 definitions

Create typed SettingDefinitions for non-bootstrap values.

CFG-2 import

Import current non-secret ENV values as scoped assignments.

Map secrets to SecretRefs only.

CFG-3 shadow resolution

Each runtime computes:

effectiveConfigFromGen2
vs
effectiveConfigFromENV

and reports mismatch without changing behavior.

CFG-4 cutover by runtime

Order recommended:

research runtime;
cognitive worker;
work graph runtime;
creator web;
core runtime;
continuity runtime;
commit runtime last.

Commit/Authority mutation path is last because error cost is highest.

CFG-5 compatibility

ENV becomes read-only emergency fallback for explicitly approved bootstrap keys only.

CFG-6 retire

Architecture lint rejects new ordinary os.Getenv outside bootstrap/config adapters.

25. SELF — ADD, evidence-only first

25.1 Inputs already present

Do not duplicate evidence from:

Capability usage/performance/model qualification;
Work task evaluation;
Epistemics Experience/Reflection;
human corrections;
development gates;
runtime health;
production outcomes.

25.2 First tables

competence assessments;
weakness patterns;
improvement needs;
dependency assessments;
learning ledger;
self-composition snapshots.

25.3 First phase limitations

SELF:

does not execute actions;
does not grant authority;
does not mutate Skill automatically;
does not own Work;
does not call providers directly.

First version is read/evaluate/propose only.

25.4 Promotion

Only after evidence proves usefulness may SELF generate Work improvement proposals through existing governed development path.

26. Commit — KEEP strict core, EXTEND audit semantics

26.1 Current

Paths:

contracts/commit/**
modules/commit/**
cmd/commit-runtime/**
adapters/postgres/commit_repository.go

Tables:

commit.authorized_actions
commit.authorized_action_envelopes
commit.action_attempts

Current code already separates authorization/execution/reconciliation and returns UNKNOWN_OUTCOME.

26.2 Keep unchanged in meaning

No Gen2 cognitive/memory/world feature gets direct mutation access.

26.3 Extend

Add:

append-only ActionAttemptTransition ledger;
explicit ExecutionReceipt;
stronger idempotency correlation where missing;
link to later Epistemics WorldOutcomeAssessment.

ExecutionReceipt means:

executor reported/observed this attempt result

It does not automatically mean:

desired world state is verified.

26.4 Migration

In-place additive.

No writer cutover required.

27. Outcome verification — EXPAND Epistemics

After significant mutation:

ActionAttempt
-> ExecutionReceipt
-> read-only observation
-> WorldOutcomeAssessment

Do not move verified world truth into Commit.

For actions where independent verification is impossible, assessment remains UNKNOWN/PARTIAL/PROVISIONAL.

28. Artifact — KEEP + EXPAND raw memory substrate

Current Artifact contracts are limited but correct boundary.

Gen2 expands:

Artifact stable ID/version;
lineage;
raw external media references;
SegmentRef;
future reprocessing lineage.

Large audio/video/document content does not go into Epistemics rows or model context wholesale.

Embeddings/transcripts/summaries are projections/derived records linked to raw Artifact version.

29. Context Compiler V1 -> V2 — REWORK, no owner

29.1 Current

contracts/cognitive/context_compiler.go

Good properties:

provider-neutral;
deterministic;
DataPolicy aware;
excludes secret/unknown sensitivity;
byte budget;
explicit omissions.

29.2 V2

Inputs include:

current Focus/ContextScope;
Task/Goal;
target model context budget;
task freshness requirement;
risk/cost of error;
candidate lanes:
- current conversation;
- relevant knowledge;
- positive experience;
- negative experience;
- active commitments;
- skills/reflexes;
- contradiction evidence;
- source provenance.

Add ContextAssemblyTrace.

29.3 Migration

Keep current CompileContext semantics as low-level final packer.

Build V2 orchestrator around it rather than replacing safe data-policy checks.

Eventually old ContextCompileInput{Candidates, DataPolicy, MaxBytes} becomes internal stage.

30. Memory Retrieval Fabric — ADD projections, no canonical owner

30.1 Projection stores

Create rebuildable projection schema(s):

retrieval documents;
World alias index;
temporal index;
experience index;
current-knowledge index;
embeddings.

30.2 Builder inputs

Builders read owner APIs/events/owner-local exports; they do not directly become cross-owner canonical DB joins.

30.3 Retrieval order

hard principal/realm/scope/version filters;
exact entity/alias matching;
temporal relevance;
graph/relation candidate expansion where available;
semantic/vector candidate retrieval;
positive experience lane;
negative experience lane;
contradiction pass;
rerank against task/context budget.

30.4 Migration

Start as shadow only. No user response depends solely on new retrieval before MILESTONE validation.

31. Core runtime — REWORK composition, do not make God orchestrator

31.1 Current problem

cmd/core-runtime/main.go is a large composition root/HTTP handler that directly wires many repositories, runtime clients, semantic heuristics and response functions.

Gen2 adds more owners; adding all logic to this file would make architecture unmaintainable.

31.2 Migration

Do not rewrite entire runtime at once.

Extract by bounded responsibility:

bootstrap/composition;
authenticated interaction ingress;
focus assembly;
context retrieval/compile;
capability strategy resolution;
Work submission;
response composition.

Each extracted package depends on contracts, not other owners' repositories.

Core runtime may orchestrate APIs but must not become canonical owner.

31.3 Cutover

Handler-by-handler, preserving HTTP compatibility.

Do not change routing + data migration + UX in one commit when avoidable.

32. Cognitive worker — KEEP replaceable cognition

Current:

interpretation;
gateway;
admission;
discovery;
model qualification;
multimodal handling;
task evaluation.

Gen2:

consume ContextBundle V2;
support iterative memory tool/query;
Perspective contract;
long-running task handoff to Work rather than holding request open.

No canonical DB ownership added.

33. Research runtime — REWORK/EXTEND

Current Research already writes/returns Epistemics evidence and has connector restrictions.

Gen2:

use Source/SourceSnapshot/Observation normalized model;
explicit TaskFreshnessRequirement;
active revalidation request;
distinguish no-access vs stale-memory fallback;
preserve partial/unknown.

Research connectors remain adapters; no new Research truth owner.

34. Work graph runtime — KEEP/EXTEND

Keep durable queue/lease/recovery/graph/synthesis.

Add:

LongRunningWork binding;
budget/preemption;
Perspective per node;
evidence-linked Finding/Attempt;
strategy escalation.

Do not create parallel “background task service” with separate canonical jobs.

35. Continuity runtime — KEEP

Keep:

due processing;
attention;
human response;
internal executor transition semantics.

Extend trigger adapter model behind Continuity contracts.

No in-memory durable timer.

36. Commit runtime — KEEP, migrate configuration last

Keep strict isolation.

Gen2 changes should be minimal until:

Configuration stable;
Supervisor available;
receipt/outcome contracts tested.

Commit runtime is one of last runtimes to remove non-bootstrap ENV compatibility.

37. Creator Web — KEEP UX surface, reduce admin dependence

Creator Web remains:

chat;
attachments;
diagnostic/emergency surfaces.

Primary configuration operation should become conversational.

Do not delete diagnostic pages; they are useful when conversational layer is degraded.

UI changes are not blockers for backend Gen2 foundation.

38. PostgreSQL adapters — REWORK per owner, no generic repository

Current adapters/postgres/*_repository.go correctly separates owner repositories.

Keep this pattern.

Add:

WorldRepository;
ConfigurationRepository;
SelfModelRepository;
projection builders/stores separately.

Do not create one generic EntityRepository/MemoryRepository.

Migration-specific backfill code should live under explicit tooling/migration packages and be removable after cutover.

39. Runtime health / Supervisor — ADD independent Lifeboat

Existing runtime/health remains health primitive.

New Supervisor/Lifeboat is a separate minimal runtime/control plane:

current active release;
last-known-good release;
start candidate;
health window;
switch;
rollback;
restart recovery.

It does not own Assistant semantic state.

It must not require the broken main conversational path to initiate rollback.

40. V1 migration ledger 001–045

The historical migration files remain immutable history. Do not edit old migration files; Gen2 adds new numbered migrations.

V1 migration	Gen2 disposition
001 bootstrap creator/authority	KEEP; Scope explicitly AuthorityScope
002 continuity prospective	KEEP/EXTEND; migrate timestamp to TriggerSpec compatibility
003 situation focus bindings	REPLACE/RETIRE after Focus V2 cutover
004 work tasks	KEEP/EXTEND
005 commit actions	KEEP
006 capability health	LEGACY KEEP
007 health v2/projection	KEEP projection pattern, may be superseded by newer projection
008 authorized action envelope	KEEP
009 continuity action receipts	KEEP compatibility, clarify vs Commit ExecutionReceipt
010 notification bindings	KEEP
011 notification envelope	KEEP
012 cognitive registry	KEEP/EXTEND
013 discovery evidence	KEEP
014 binding revision heads	KEEP
015 interaction foundation	KEEP/EXTEND channel semantics
016 human intent project context	PARTIAL RETIRE: project registry/focus -> World/Focus V2; historical interaction refs retained
017 work planning	KEEP/EXTEND
018 capability resolution	KEEP/EXTEND
019 reminder metadata	KEEP
020 attention items	KEEP
021 notification assistant origin	KEEP
022 human responses	KEEP
023 experience/reflection/skill	EXTEND Experience; REWORK Skill V2
024 agent delegation	KEEP/EXTEND
025 delegation attention/verification	KEEP
026 attention response outcomes	KEEP
027 semantic candidates	KEEP/EXTEND contextual refs
028 cognitive provider profile	KEEP
029 agent execution persistence	KEEP
030 dispatch queue	KEEP
031 work graph persistence	KEEP
032 reflection leases	KEEP
033 cognitive admission leases	KEEP
034 epistemics truth bundles	REPLACE writer; RETAIN legacy read/history
035 graph dispatch binding	KEEP
036 graph synthesis	KEEP
037 development gates	KEEP
038 development gate privileges	KEEP
039 task evaluations	KEEP/EXTEND Personal Benchmark
040 task evaluation dispatch	KEEP
041 task evaluation payload	KEEP, constrain JSON to typed contract payloads
042 agent execution recovery	KEEP
043 governed development candidates	KEEP/EXTEND with expected improvement
044 graph orchestration snapshot	KEEP
045 model qualification observations	KEEP

Historical migrations are evidence. Gen2 migration starts at the next available migration number and never renumbers V1.

41. New Gen2 schema migration groups

Exact migration numbers are assigned after current repository HEAD is fixed.

Recommended groups:

G2-A — architecture/bootstrap

create roles/schemas for world, configuration, self_model, projection writer;
no user behavior switch.

G2-B — Focus/World

World object/type/alias definitions;
ContextScope/FocusFrame/transitions/heads;
project backfill tool;
compatibility read model.

G2-C — Epistemics normalized evidence

Source/Snapshot/Excerpt/Observation/Proposition/Claim/Belief;
freshness/version metadata;
truth bundle backfill;
legacy bundle compatibility.

G2-D — Retrieval/context

projection schema/index tables;
retrieval builder checkpoints;
ContextAssemblyTrace.

G2-E — Configuration

definitions/revisions/heads;
assignments/revisions/heads;
effective config projection;
ENV import audit table/tooling if needed.

G2-F — Skill/Reflex

Skill V2 revision/head;
Reflex revision/head;
shadow/revalidation evidence.

G2-G — SELF

assessments;
weakness;
learning ledger;
composition snapshots.

G2-H — Work/Continuity/Commit additive semantics

LongRunningWork;
Perspective;
ConsequenceAssessment;
TriggerSpec representation;
action transition ledger;
ExecutionReceipt;
WorldOutcomeAssessment.

G2-I — retirement markers

No DROP initially. Record deprecated writers/readers and enable architecture lint to prevent new call sites.

42. Backfill requirements

Every backfill command must be:

idempotent;
restartable;
checkpointed;
scope-limited;
dry-run capable;
counts/digests reported;
no destructive mutation of source;
able to resume after process failure.

Backfill status itself is operational migration state, not domain truth.

For large future raw memory, backfill must stream/chunk; never load entire history into one model context.

43. Feature switches/cutover controls

Gen2 needs explicit operational switches stored outside model prompts.

Examples:

Focus read strategy: V1 | SHADOW_V2 | V2;
Context strategy: PERSONAL_MEMORY_V1 | SHADOW_GEN2 | GEN2;
Epistemics write strategy: TRUTH_BUNDLE_V1 | NORMALIZED_G2;
Config read strategy per runtime: ENV | SHADOW_G2 | G2;
Skill strategy: PROFILE_V1 | SHADOW_SKILL_V2 | SKILL_V2;
Reflex: DISABLED | SHADOW | ENABLED;
SELF: OBSERVE_ONLY | PROPOSE;
Supervisor release channel.

These switches are migration controls, not long-lived user settings. After retirement they are removed.

44. Migration observation windows

Do not promote based only on unit tests.

For each cutover define an observation window appropriate to risk.

Low-risk retrieval:

shadow comparison over representative real turns.

Focus:

enough turns to include explicit/inferred switches and Company A/B ambiguity.

Configuration:

runtime restart + missing-config + secret-resolution scenarios.

Skill/Reflex:

historical replay + shadow production.

Commit:

reconciliation/unknown-outcome/restart.

No fixed universal duration is required; evidence threshold is task-specific.

45. Testing strategy during migration

45.1 FAST — default Codex loop

Run only:

changed package compile/tests;
directly affected contract tests;
one/few regression cases for the actual gap;
fast architecture lint for touched boundary;
migration SQL parse/small repository test when schema touched.

Do not run go test ./... after every edit.

45.2 MILESTONE — after architectural organ is complete

Run:

all tests of changed owner;
integration with immediate neighbors;
migration/backfill equivalence;
restart/recovery for durable state;
relevant household scenarios;
architecture boundary checks.

Suggested milestones:

A. World + Focus V2 B. Epistemics + Retrieval + Context Compiler C. Configuration + Skill/Reflex + SELF D. LongRunningWork + Consequence + Outcome + Lifeboat E. V1 compatibility retirement / Gen2 system integration

45.3 FULL — rare

Run only:

before large architecture phase merge;
after critical schema migration;
before production canary;
before retiring V1 rollback path;
final Gen2 acceptance.

FULL includes repository-wide test/vet/architecture/recovery/household/multimodal/migration checks.

45.4 Test-growth rule

If test code/execution begins consuming substantially more effort than capability implementation without protecting a new material risk:

stop test expansion and review sufficiency.

Passing tests is evidence; the task goal is the working architectural capability.

46. Runtime migration order

Recommended order is chosen by dependency/risk, not convenience.

0 PRE-GEN2 baseline
1 architecture docs + ADRs + registry generation 2
2 World vocabulary/referents
3 Situation Focus V2
4 Epistemics normalized/freshness
5 Retrieval + Context Compiler V2
6 Configuration foundation
7 Capability Skill V2 + Reflex
8 SELF observe/evaluate
9 Work long-running/perspective/consequence
10 Continuity TriggerSpec extensions
11 Commit receipt/outcome clarification
12 Supervisor/Lifeboat
13 runtime-by-runtime Configuration cutover
14 household/system integration
15 production canary
16 V1 writer/read retirement

Why World before Focus: Focus V2 needs stable typed referents.

Why Epistemics before retrieval: retrieval must index canonical evidence, not invent a new truth store.

Why Configuration before deep self-evolution: new Skills/self-written capabilities need scoped typed settings.

Why SELF after Skill/Experience/Capability evidence: SELF must derive assessments from evidence, not empty self-scores.

Why Commit changes late: highest consequence and current V1 boundary is already strong.

47. Package-level action map

KEEP mostly unchanged

modules/identity/**
modules/authority/**
modules/commit/domain/** semantic core
modules/continuity/domain/human_response.go
current Work Graph/lease/recovery semantics
platform/ids/**
platform/secrets/**
runtime/health/**
model/provider adapters behind Capability/Gateway.

EXTEND

contracts/interaction/**
contracts/epistemics/**
contracts/work/**
contracts/continuity/**
contracts/capability/**
modules/epistemics/**
modules/work/**
modules/continuity/**
modules/capability/**
corresponding owner postgres repositories.

REWORK

contracts/situation/**
modules/situation/**
contracts/cognitive/context_compiler.go
cmd/core-runtime/main.go composition/routing structure
research persistence path;
Skill profile API compatibility layer.

REPLACE/RETIRE

modules/interaction/domain/personal_memory.go as primary memory retrieval;
Situation project registry as canonical world referent registry;
ActiveProjectFocus as canonical focus;
direct truth_bundles writer;
ordinary non-bootstrap ENV reads after Config cutover.

ADD

contracts/world/**, modules/world/**;
contracts/configuration/**, modules/configuration/**;
contracts/selfmodel/**, modules/selfmodel/**;
Context retrieval contracts/projection builders;
Supervisor runtime;
Reflex contracts/domain.

Exact package names may be adjusted by ADR but semantic owners may not drift silently.

48. Architecture registry changes

Gen2 needs semantic_generation: 2, architecture_generation: 2 only after ADRs are accepted.

Registry additions:

World owner/schema;
Configuration owner/schema;
SELF owner/schema;
projection mechanism entry as non-domain owner;
Supervisor topology as runtime-only.

Registry dependency rules must prevent:

SELF -> direct owner DB writes;
World -> Epistemics truth ownership;
Configuration -> secrets raw values;
cognitive worker -> canonical writes;
projection builder -> canonical domain writes;
Work -> Authority grant mutation.

49. Known migrations that must not be combined

Avoid one task that simultaneously performs:

Focus + Memory + Configuration

Too many failure causes; impossible to diagnose.

Epistemics normalization + model provider change

Would confound truth migration with model quality.

Skill V2 + Reflex auto-generation

First make Skill V2 stable; Reflex is next layer.

SELF + autonomous self-code promotion

SELF first observes/evaluates. Existing governed development remains authority path.

Commit schema + Authority redesign

Authority semantics are not being redesigned in Gen2 foundation.

Production cutover + V1 table deletion

Retirement occurs after stable observation, never in same promotion.

50. Rollback map by subsystem

Subsystem	Rollback target
World seed	disable World consumers; V1 project registry remains
Focus V2	switch read to ActiveProjectFocus/FocusBinding
Epistemics normalized	legacy truth bundle reader; stop new writer if lossy
Retrieval V2	PersonalMemory/current deterministic context
Configuration	approved ENV bootstrap/fallback snapshot
Skill V2	SkillProfile V1 compatibility reader
Reflex	disable Reflex, escalate to Skill/reasoning
SELF	set OBSERVE_ONLY/off; no canonical dependency
LongRunningWork	existing AgentRun/Graph execution
TriggerSpec	absolute timestamp compatibility path
Outcome verification	Commit receipt remains; world outcome may be UNKNOWN
Supervisor candidate	switch last-known-good release

Rollback must not discard new evidence.

51. Data retention / retirement rules

Retirement candidates:

V1 project tables;
ActiveProjectFocus;
PersonalMemory compatibility path;
direct truth bundle writer;
SkillProfile V1 writer;
non-bootstrap ENV readers.

Retirement requires all:

no active writer;
no undocumented reader;
MILESTONE/FULL evidence;
rollback window completed;
data migrated or intentionally retained historical;
architecture scan rejects new references;
explicit retirement commit/task.

DROP TABLE is optional and can be much later. Logical retirement is more important than immediate physical deletion.

52. Production transition strategy

Production Gen2 is promoted by subsystem/phase, not by one giant day.

Recommended:

deploy code capable of V1 + shadow Gen2
-> build/backfill Gen2 data
-> observe
-> enable one Gen2 read path
-> observe
-> enable canonical writer
-> observe
-> next subsystem

For strict mutation path, prefer:

candidate release
-> health
-> controlled traffic
-> reconciliation smoke
-> promote

Supervisor/Lifeboat must exist before the first self-modifying production release that could break main conversational recovery.

53. Evidence required at each status change

KEEP

Existing regression/invariant evidence sufficient unless code touched.

EXTEND

Need targeted tests + backward-compatible read/write evidence.

REWORK

Need old vs new behavior comparison for overlapping semantics.

REPLACE

Need:

backfill;
shadow/equivalence;
writer cutover proof;
rollback proof.

RETIRE

Need:

caller inventory;
no active writer;
no required read path;
observation window;
explicit approval/task.

54. Migration blockers that must become ADR/task decisions

Before implementation, resolve/accept:

AuthorityScope vs ContextScope.
World owner and typed schema vocabulary.
Epistemics normalized truth/freshness model.
Memory retrieval with no Memory owner.
Configuration owner/bootstrap boundary.
Skill V2/Reflex lifecycle.
SELF boundaries.
Long-running Work semantics.
Consequence/Risk semantics.
Supervisor/Lifeboat.
ExecutionReceipt vs WorldOutcomeAssessment.
Gen2 compatibility/retirement policy.

If an implementation task reaches an unresolved owner conflict, result is BLOCKED_BY_ARCHITECTURE_GAP, not a local workaround.

55. What is deliberately not migrated now

Gen2 foundation leaves interfaces/open vocabulary for:

ambient microphone streams;
speaker recognition;
continuous video;
TV/speaker output choice;
home automation;
robots/vehicles;
Digital Twin;
mandatory graph DB;
automatic compilation of every Skill;
automatic global Skill rewrite by new model;
permanent fixed multi-agent staff.

No current Gen2 decision may make these require replacement of strict core semantics later.

56. Milestone definition for Master Plan

M0 — Frozen V1 baseline

Result:

clean/known main;
base commit/tag;
current work preserved.

M1 — Semantic substrate

Includes:

ADRs;
World;
ContextScope/Focus V2.

Exit:

project-centric focus no longer limits new features;
V1 compatibility works.

M2 — Memory/Knowledge foundation

Includes:

Epistemics normalization;
freshness/volatility;
Experience V2;
retrieval fabric;
Context Compiler V2.

Exit:

current task receives scoped fresh positive+negative memory;
V1 PersonalMemory is fallback only.

M3 — Evolvable behavior

Includes:

Configuration;
Skill V2;
Reflex;
SELF observe/evaluate.

Exit:

new capability can define typed scoped settings;
known solution can become Skill/Reflex;
quality is evidence-backed.

M4 — Durable autonomy and self-evolution safety

Includes:

LongRunningWork;
Perspective/team;
Consequence;
improved Outcome verification;
Supervisor/Lifeboat.

Exit:

long tasks survive restart;
self-change can be candidate-tested-promoted-rolled-back safely.

M5 — Integrated household Gen2

Includes:

migration of existing household behavior;
multi-person/scope tests;
degraded/local model paths;
production canary.

M6 — V1 retirement

Only after stable production evidence.

57. Handoff to GEN2 Master Plan

Master Plan must convert this map into executable task packets.

For every task packet it must state:

milestone;
current owner;
target owner;
exact allowed paths;
exact forbidden/protected paths;
table/contracts affected;
writer before task;
writer after task;
backfill needed/not needed;
compatibility strategy;
rollback;
FAST gate;
whether task ends a MILESTONE;
whether FULL gate is justified;
dependencies;
explicit non-goals.

No Codex task may say only:

«реализуй память Gen2»

or:

«улучши архитектуру».

Every task must finish one bounded migration step with observable exit criteria.

58. Final migration verdict

Existing V1 is worth preserving.

Do not rewrite from zero.

But the following V1 mechanisms must not remain the long-term foundation:

project-only/current ActiveProjectFocus;
ad-hoc PersonalMemory context projection;
direct denormalized truth_bundles as final epistemic model;
ENV as ordinary runtime configuration;
implicit FastPath as the only representation of Reflex;
distributed self-evaluation without explicit SELF model;
absence of independent whole-system rollback supervisor.

The following V1 mechanisms are strategic assets and should survive Gen2:

Identity/Principal separation;
Authority boundary;
Epistemic truth-first distinction;
Work durable graph/agents/checkpoints;
Continuity prospective/attention;
Capability registry/model evidence;
Skill positive/negative evidence foundation;
Commit/UNKNOWN_OUTCOME/reconciliation;
governed self-development gates;
secret references;
architecture registry/lint/governance.

Итоговая стратегия:

preserve evidence and invariants
-> add Gen2 semantic substrate
-> backfill
-> shadow
-> cut over one owner at a time
-> observe
-> keep rollback
-> retire V1 paths only after proof

Это и есть путь от текущего Assistant к развивающемуся Домовому Gen2 без потери уже сделанной работы.

58 KiB Raw Permalink Blame History Unescape Escape

Домовой — GEN2 Migration Map

0. Назначение документа

1. Неподлежащие потере V1-инварианты

2. PRE-GEN2 — обязательная фиксация текущей разработки

2.1 Цель

2.2 Последовательность

Нельзя

2.3 PRE-GEN2 FAST gate

3. Закон миграции writer/read path

3.1 Один canonical writer

A. In-place extension

B. Shadow projection

C. Controlled cutover

3.2 Rollback

3.3 No destructive migration first

4. Owner-level migration matrix

5. Identity — KEEP

5.1 Current V1

5.2 Gen2 target

5.3 Migration

5.4 Cutover / rollback

5.5 Tests

6. Authority — KEEP, Scope semantic split

6.1 Current V1

6.2 Problem

6.3 Migration decision

6.4 Compatibility

6.5 Risk

7. Interaction — KEEP append-only ingress, EXPAND channels

7.1 Current V1

7.2 Keep

7.3 Extend

7.4 ProjectReference migration

7.5 PersonalMemory migration

8. PersonalMemory V1 — REPLACE then RETIRE

8.1 Current V1

8.2 Gen2 destination

8.3 Migration stages

Stage PM-1 — freeze semantics

Stage PM-2 — shadow retrieval

Stage PM-3 — read cutover

Stage PM-4 — retire

8.4 Rollback

9. Situation — REWORK to Focus V2

9.1 Current V1

9.2 Gen2 target

9.3 Data migration

9.4 Project registry move

9.5 Writer cutover

9.6 FAST

9.7 MILESTONE

10. World — ADD narrow bounded context

10.1 V1 source

10.2 First Gen2 responsibility

10.3 Migration

10.4 Safety

10.5 Cutover

11. Epistemics truth data — REWORK

11.1 Current V1

11.2 Problem

11.3 Target

11.4 Backfill

11.5 Cutover

11.6 Rollback

12. Freshness / volatility — ADD inside Epistemics, not permanent stale flag

12.1 Existing V1

12.2 Gen2

12.3 Migration

13. Episode — ADD derived episodic memory

14. Experience/Reflection — EXTEND, preserve V1 records

14.1 Current

14.2 Migration

14.3 Reflection

15. Source trust — ADD evidence-backed assessments

16. Work goals/tasks/plans — EXTEND, not replace

16.1 Current V1

16.2 Keep

16.3 Extend

16.4 Migration

58 KiB

Raw Permalink Blame History