Skip to content

Commit 8723db4

Browse files
committed
docs: refresh roadmaps to match current delivery state
- Condense the LifeOS roadmap Delivered section to outcomes and SSOT ownership instead of duplicating implementation history - Record the temporarily skipped Android emulator integration gate (flutter/flutter#191095) that blocks on-device exit evidence - Fix the Android encryption initiative inconsistency and update the Platform Quality Track now that scroll caching and transition work have landed - Reflect delivered voice interaction, Android self-update, and the cross-domain simplification wave in the product snapshot - Route personal memory work to current SSOTs now that Memory V1 is delivered, and add app/interaction and core/speech to the agent task map AI code authority
1 parent 405bc89 commit 8723db4

4 files changed

Lines changed: 60 additions & 178 deletions

File tree

‎docs/agent-map.md‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -13,7 +13,7 @@ router, not a second source of product or architecture truth.
1313
| Add or change a domain | [LifeOS Shell](architecture/lifeos-shell.md), owning domain SSOT | Owning `apps/mobile/lib/app/domain_packs/*_pack.dart`, `apps/mobile/lib/features/<domain>/` | Domain composition, router, and opt-in tests |
1414
| Change FinanceOS behavior | [FinanceOS](domains/financeos-domain.md), then the relevant Finance topic SSOT | `apps/mobile/lib/features/finance/` | Focused repository/tool/widget tests |
1515
| Change a domain Agent | [Agent Experience](ai/agent-experience.md), owning domain SSOT | Owning `apps/mobile/lib/features/<domain>/agents/`, `DomainPack.agentBuilder` | Agent unit tests, outcome corpus, composition test |
16-
| Change AI runtime or interaction | [AI Architecture](ai/ai-architecture.md), [AI Protocol](ai/ai-protocol.md), [Agent Runtime](architecture/agent-runtime-current.md) | `apps/mobile/lib/core/ai/`, `apps/mobile/lib/app/agent_runtime/`, `apps/mobile/native/lifeos_native/` | AI wire-enum and FRB-entrypoint gates |
16+
| Change AI runtime or interaction | [AI Architecture](ai/ai-architecture.md), [AI Protocol](ai/ai-protocol.md), [Agent Runtime](architecture/agent-runtime-current.md) | `apps/mobile/lib/core/ai/`, `apps/mobile/lib/app/agent_runtime/`, `apps/mobile/lib/app/interaction/`, `apps/mobile/lib/core/speech/`, `apps/mobile/native/lifeos_native/` | AI wire-enum and FRB-entrypoint gates |
1717
| Add or change an AI tool | [AI Architecture](ai/ai-architecture.md), owning domain SSOT | Domain tool barrel and `DomainPack` descriptors | Tool tests and device degradation contract |
1818
| Change persistence or a synced table | [Sync v3](sync/sync-v3.md), owning domain SSOT | `apps/mobile/lib/core/persistence/`, `apps/mobile/lib/core/sync/sync_table_registry.dart` | Sync fixtures, repository tests, architecture lints |
1919
| Change backup or destructive data management | [LifeOS Shell](architecture/lifeos-shell.md), owning domain SSOT | `apps/mobile/lib/core/backup/`, `apps/mobile/lib/core/data_management/` | Restore atomicity and data-management tests |

‎docs/index.md‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -23,7 +23,7 @@ This documentation uses a docs-as-code structure:
2323
| Change architecture or cross-domain code | [Architecture Northstar](architecture/lifeos-architecture-northstar.md), [LifeOS Shell](architecture/lifeos-shell.md) |
2424
| Add or modify a domain | [LifeOS Shell](architecture/lifeos-shell.md), then the owning domain SSOT |
2525
| Work on device AI | [Device AI Architecture](ai/ai-architecture.md), [Runtime Event Contract](ai/ai-protocol.md), [Agent Experience](ai/agent-experience.md) |
26-
| Plan Personal Memory V1 | [Personal Memory V1 Implementation Plan](ai/personal-memory-implementation-plan.md), then confirm sequencing in the [LifeOS Roadmap](roadmap/roadmap-lifeos.md) |
26+
| Work on personal memory beyond V1 | [AI Architecture](ai/ai-architecture.md) for current Memory behavior; the [Personal Memory Implementation Plan](ai/personal-memory-implementation-plan.md) retains only the deferred Phases 4–5 |
2727
| Work on Rust agent runtime | [Agent Runtime Current Architecture](architecture/agent-runtime-current.md) |
2828
| Work on sync | [Sync v3 Protocol](sync/sync-v3.md), [Protocol Test Catalogue](sync/sync-protocol-tests.md) |
2929
| Add business diagnostics | [Structured Logging](development/logging.md) |

‎docs/roadmap/roadmap-finance.md‎

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2,7 +2,7 @@
22

33
Status: active FinanceOS sequencing SSOT.
44

5-
Last reviewed: 2026-08-01.
5+
Last reviewed: 2026-08-29.
66

77
FinanceOS is the always-on seed domain. This document contains only
88
Finance-specific product sequencing. Cross-domain shell, Memory Runtime,
@@ -99,6 +99,9 @@ Current evidence:
9999
Exit evidence:
100100

101101
- Prove interruption recovery on the production Android file-backed database.
102+
The shared Android emulator integration gate is temporarily skipped on CI
103+
pending an upstream Flutter fix; see the shared delivery-risk note in
104+
`roadmap-lifeos.md` Next.
102105
- Expand the representative dataset only when measured production exports
103106
exceed the current 1,000-row boundary.
104107

‎docs/roadmap/roadmap-lifeos.md‎

Lines changed: 54 additions & 175 deletions
Original file line numberDiff line numberDiff line change
@@ -2,7 +2,7 @@
22

33
Status: active cross-domain sequencing SSOT.
44

5-
Last reviewed: 2026-08-23.
5+
Last reviewed: 2026-08-29.
66

77
This roadmap contains only work that changes cross-domain product outcomes or
88
shared delivery risk. Current architecture belongs in the architecture SSOTs,
@@ -27,7 +27,16 @@ remains a schema-agnostic Sync v3 row store.
2727
The first production cross-domain loop is complete: Life signals expose local
2828
evidence, confirmed proposals create source-preserving Execution actions, and
2929
Execution Review shows whether the current source signal is cleared or still
30-
active without claiming causality.
30+
active without claiming causality. The Personal Intelligence Loop and Personal
31+
Memory V1 are delivered; their current behavior is owned by the AI and
32+
architecture SSOTs.
33+
34+
Recent releases simplified Finance planning and cashflow, Execution plans and
35+
review, Knowledge review, and Settings without changing domain boundaries.
36+
Mobile AI chat now supports hands-free voice turns through on-device ASR/TTS
37+
inside the existing interaction session, and Android builds self-update
38+
through GitHub releases. These are current capabilities owned by the domain
39+
SSOTs and [AI Architecture](../ai/ai-architecture.md), not roadmap items.
3140

3241
## Now
3342

@@ -101,181 +110,53 @@ changes sequencing for all downstream roadmap work.
101110

102111
## Delivered
103112

104-
These cross-domain initiatives have repository exit evidence. Follow-up work
105-
must enter `Now` or `Next` with a new measurable outcome instead of reopening
106-
the removed legacy architecture.
113+
These cross-domain initiatives have repository exit evidence and are complete.
114+
This section records outcomes and ownership only; current guarantees belong in
115+
the SSOTs and executable tests named below. Follow-up work must enter `Now` or
116+
`Next` with a new measurable outcome instead of reopening the removed legacy
117+
architecture.
107118

108119
### Personal Intelligence Loop
109120

110-
Outcome: turn the existing domain Agents, Memory Runtime, Life signals, and
111-
proposal safety seams into one personal intelligence loop that correlates
112-
active-domain evidence, stays silent when nothing material changed, and gives
113-
the user one traceable next-step judgment on the Life surface.
114-
115-
Delivered evidence:
116-
117-
- The Life hub already composes bounded deterministic signals contributed by
118-
active `DomainPack`s, and source-preserving Life actions can be evaluated
119-
against later domain state.
120-
- `AgentArtifact`, stable findings, trace links, evidence navigation,
121-
dismiss/snooze state, proposal confirmation, and outcome corpus evaluation
122-
already provide most of the trust and presentation substrate.
123-
- `EventRecord` now requires typed domain/kind/time/source identity, confidence,
124-
facts, entities, and evidence anchors. Schema v73 rebuilt the derived event
125-
index directly; no legacy free-form event contract remains.
126-
- `PersonalProfileSnapshot` materializes explicitly confirmed structured
127-
profile facts with authority, provenance, scope, and validity metadata.
128-
`LifeContextSnapshot` combines that stable profile with fingerprinted
129-
active-domain state, freshness, changes, goals, constraints, and relevant
130-
history.
131-
- The app-owned Daily Navigator consumes domain-owned `LifeSignal`s, gates on
132-
material change and evidence freshness before any model call, and owns the
133-
Life-surface cross-domain judgment. The retired Health-owned briefing code,
134-
preferences, UI, schedules, notifications, and tests were removed.
135-
- `AgentTrigger` supports schedule, event, threshold, state transition,
136-
freshness, and manual triggers. Knowledge write debounce uses the shared
137-
coordinator, while Daily Navigator combines signal triggers with a schedule
138-
fallback.
139-
- `AttentionArbiter` is the only notification policy owner. It persists
140-
`silent`, `surface`, or `interrupt` decisions under a global interrupt
141-
budget; the background-safe callback reads only a precomputed primitive
142-
snapshot and never invokes an LLM or writes business data.
143-
- Durable Memory has explicit retrieval role, evidence authority, provenance,
144-
and supersede lineage. Domain Agents no longer write periodic summaries into
145-
Memory or directly send notifications; Agent Artifacts remain temporary
146-
unless the user explicitly confirms a Memory proposal.
147-
- Structured Agent feedback links accepted, dismissed, snoozed, completed, and
148-
undone outcomes to Life context, finding, and attention fingerprints. The
149-
policy corpus evaluates silence, duplication, stale data, evidence,
150-
actionability, severity, domain isolation, and interrupt budget behavior.
151-
- Advanced Settings now exposes local-only Developer Issue capture. It retains
152-
the last domain route, build identity, latest trace id, at most five tool
153-
error codes, and optional screenshot identity; export is explicit, excludes
154-
user identity and local paths, and never creates a remote issue.
155-
156-
Delivered vertical slices:
157-
158-
1. **Typed event and evidence identity.** Added explicit domain, namespaced kind,
159-
occurred/observed times, source row family/id/fingerprint, confidence, and
160-
stable evidence identity to the cross-domain event contract. Migrated
161-
storage and all four current domain indexers directly. Unknown domains fail
162-
closed rather than defaulting to Finance; source routes resolve through
163-
registered domain contributions.
164-
2. **Personal profile and Life context.** Materialized a structured
165-
`PersonalProfileSnapshot` for goals, preferences, constraints, and rules,
166-
with authority, provenance, confirmation, validity, and update metadata.
167-
Domain baselines and routines remain deterministic current-state inputs or
168-
explicitly confirmed profile rules rather than model-authored profile data.
169-
Composed it with deterministic active-domain state, freshness, recent
170-
changes, goals, and relevant history into a fingerprinted
171-
`LifeContextSnapshot`. Domain business calculations remain domain-owned;
172-
app composition assembles only domain-neutral slices.
173-
3. **Daily Navigator vertical slice.** Added one app-owned Life synthesis Agent
174-
and a Life-surface presentation owner without creating a new opt-in domain.
175-
It starts with available Health and Finance findings, with Execution context
176-
included only when enabled and relevant. The Agent correlates, prioritizes,
177-
explains, and proposes a next step; it does not recalculate domain metrics
178-
or write business rows directly. It invokes the device LLM only after a
179-
deterministic material-change gate, retains a deterministic fallback, and
180-
replaces the removed duplicate domain-owned briefing implementation.
181-
4. **Signal-driven triggers and attention arbitration.** Introduced event,
182-
threshold, state-transition, freshness, schedule, and manual trigger specs
183-
without conflating them with persisted run provenance. Migrated the existing
184-
Knowledge debounce path first, then connect real Finance import and Health
185-
refresh transitions. Routed every candidate through one global attention
186-
decision (`silent`, `surface`, or `interrupt`) that considers novelty,
187-
severity, confidence, actionability, recent notifications, and current
188-
dismiss/snooze state. A bounded background-safe path may inspect only a
189-
precomputed snapshot and record pending attention; it must not invoke an
190-
LLM or mutate business data.
191-
5. **Memory hygiene, feedback, and policy evaluation.** Distinguished source
192-
facts, user-confirmed records, deterministic derived records, model-derived
193-
candidates, and temporary Agent artifacts through evidence authority,
194-
provenance, retrieval role, and access policy. Agent
195-
artifacts do not become durable Memory by default; periodic reviews remain
196-
artifacts, and only explicit confirmation can promote an Agent inference
197-
into durable Memory or the personal profile. Linked accepted,
198-
dismissed, snoozed, completed, and undone suggestions back to their
199-
snapshot/finding fingerprints. Extended the Agent corpus to test whether a
200-
result should appear at all, and added Developer Mode issue capture with
201-
explicit local export of route, build, trace, bounded tool errors, and an
202-
optional screenshot.
203-
204-
Repository verification:
205-
206-
- Every event written by Finance, Health, Knowledge, and Execution has an
207-
explicit domain and source identity. Regression fixtures prove that unknown,
208-
Knowledge, and Execution events cannot enter Finance synthesis by exclusion.
209-
- Every evidence-bearing ready artifact in the vertical slice resolves to the
210-
correct active-domain source route. Missing or inactive-domain evidence
211-
fails closed instead of routing to a generic or wrong domain page.
212-
- A deterministic fixture produces the same personal profile and Life context
213-
fingerprint from the same inputs. Stale inputs are marked explicitly, and
214-
unconfirmed profile candidates cannot act as hard constraints or authorize
215-
a proposal.
216-
- A checked-in Health/Finance fixture produces one Daily Navigator artifact
217-
with conclusion-level evidence and at most one confirmed action path. A
218-
no-material-change fixture produces no visible artifact and records zero LLM
219-
calls; inactive-domain and stale-data fixtures make no cross-domain claim.
220-
- Trigger tests prove debounce/idempotency, threshold and state-transition
221-
behavior, and schedule fallback. Background tests prove that the
222-
background-safe path performs no LLM call, tool effect, proposal apply, or
223-
business write.
224-
- Every notification is backed by a persisted attention decision. Duplicate,
225-
unchanged, dismissed, and actively snoozed findings cannot consume the
226-
interrupt budget, while silent decisions remain available to local quality
227-
evaluation.
228-
- Retrieval tests prove that temporary Agent artifacts are excluded from
229-
durable personal context unless explicitly promoted; periodic Agent reviews
230-
no longer write durable Memory rows.
231-
- The Agent policy corpus covers `should_surface`, `should_stay_silent`,
232-
severity, evidence completeness, actionability, duplicate suppression,
233-
stale data, wrong-domain combination, and no-model-call gating. Architecture
234-
composition and lint tests continue to enforce active-domain isolation and
235-
the boundaries owned by the LifeOS Northstar and Agent Experience SSOT.
236-
- The local dogfood surface records export state and Agent feedback records
237-
real outcome transitions without promoting either record into Memory.
238-
239-
Owner: cross-domain app composition. Domains own deterministic facts,
240-
findings, routes, and proposal application; shared contracts and lifecycle
241-
policy remain in their existing core seams.
121+
Outcome: correlate active-domain evidence into one traceable next-step
122+
judgment on the Life surface while staying silent when nothing material
123+
changed.
124+
125+
Delivered in five vertical slices: typed event and evidence identity across
126+
all four domains; the confirmed `PersonalProfileSnapshot` and fingerprinted
127+
`LifeContextSnapshot`; the app-owned Daily Navigator gated on material change
128+
and evidence freshness; signal-driven triggers with one global attention
129+
arbiter (`silent` / `surface` / `interrupt`); and memory hygiene with Agent
130+
feedback and policy evaluation. Personal Memory V1 (Phases 0–3) shipped in the
131+
same loop; Phases 4–5 remain deferred in
132+
[its implementation plan](../ai/personal-memory-implementation-plan.md).
133+
134+
Owning SSOTs: [LifeOS Shell](../architecture/lifeos-shell.md) for composition,
135+
trigger, and attention seams; [AI Architecture](../ai/ai-architecture.md) for
136+
the agent runtime and memory; [Agent Experience](../ai/agent-experience.md)
137+
for findings, artifacts, and feedback. Coverage lives in the Agent policy
138+
corpus and the attention, memory, event, and composition test suites.
242139

243140
### Agent Runtime Capability Contract
244141

245-
Outcome: keep NaviWealth's runtime integration documentation aligned with the
246-
exact pinned standalone runtime instead of manually copying a fast-moving
247-
capability inventory.
248-
249-
Delivered evidence:
250-
251-
- NaviWealth pins `third_party/agent-runtime` at
252-
`a89e11b0924107b637e741aeb8d75950b6430ac6`.
253-
- That commit contains the dependency-light `bindings/ts` package and lists it
254-
in the runtime README.
255-
- Exact runtime APIs are already owned by the submodule crates, schemas,
256-
fixtures, and contract tests; the host documentation should describe the
257-
integration boundary rather than duplicate their feature list.
258-
259-
Repository verification:
260-
261-
- `tool/generate_agent_runtime_capabilities.dart` derives a machine-readable
262-
manifest from the pinned submodule's Cargo metadata, package metadata,
263-
schemas, and README.
264-
- `docs/architecture/agent-runtime-capabilities.json` records the pinned commit,
265-
bindings, transports, stores, providers, and protocol/snapshot versions.
266-
- `tool/check-mobile-static.sh` runs the generator in `--check` mode, and CI
267-
checks out the submodule before that gate.
268-
- `agent-runtime-current.md` references the generated manifest and retains only
269-
NaviWealth-owned integration boundaries and host limitations.
270-
271-
Owner: native Agent Runtime integration. This track may proceed independently
272-
and does not block the Personal Intelligence Loop vertical slice.
142+
Outcome: keep runtime integration documentation aligned with the pinned
143+
standalone runtime instead of manually copying a fast-moving capability
144+
inventory. `third_party/agent-runtime` is pinned; the generated manifest
145+
`docs/architecture/agent-runtime-capabilities.json` is checked in CI by
146+
`tool/check-mobile-static.sh`; `agent-runtime-current.md` documents only
147+
NaviWealth-owned integration boundaries and host limitations.
273148

274149
## Next
275150

276151
These are accepted follow-ups but are not allowed to displace `Now` work
277152
without an explicit reorder.
278153

154+
Shared delivery risk: the Android emulator integration workflow is
155+
temporarily skipped on CI while an upstream Flutter Semantics assertion blocks
156+
the integration flow (flutter/flutter#191095). The on-device exit evidence of
157+
the first two initiatives below is pending the restored gate
158+
(`RUN_FLUTTER_TESTS` in `.github/workflows/integration-device.yml`).
159+
279160
### Data Portability Recovery Correctness
280161

281162
Outcome: make encrypted recovery trustworthy under destructive failure modes,
@@ -344,8 +225,7 @@ Current evidence:
344225
and successful reopen after the original key is restored. Its CI wrapper
345226
rejects a nominally green Flutter run when any encryption, migration,
346227
Keystore, or backup/restore evidence marker is missing, and emits a separate
347-
privacy-safe JSON summary. A green emulator artifact is still required
348-
before this initiative can leave `Now`.
228+
privacy-safe JSON summary.
349229

350230
Exit evidence:
351231

@@ -372,23 +252,22 @@ feature parity on Web without weakening device-only guarantees.
372252

373253
Current evidence:
374254

375-
- Scroll perf analysis identified repeated `List<_FeedItem>` and
376-
`List<_TimelineItem>` allocations in Activity Feed and AI Chat that were
377-
causing GC pressure on every scroll tick. Caching these lists removes the
378-
per-frame allocation spike.
255+
- The identified scroll hotspots are fixed: Activity Feed and AI Chat cache
256+
their flattened feed/timeline lists instead of rebuilding them per frame,
257+
and AI navigation transitions were smoothed. Remaining hotspots must be
258+
found by measurement, not by re-deriving the fixed list.
379259
- Web builds exclude AI runtime and Health platform integration by design; the
380260
remaining parity gap is UI completeness, not architecture.
381261
- The project has lint gates for cross-feature imports, domain-neutral
382262
contracts, and FRB entrypoints.
383263

384264
Exit evidence:
385265

386-
- Activity Feed and AI Chat maintain 8ms frame budget on 120Hz test devices
387-
during sustained scroll, measured in Flutter DevTools performance traces.
266+
- Sustained-scroll frame budgets are validated by measurement on 120Hz test
267+
devices in Flutter DevTools performance traces, with a targeted widget or
268+
integration test guarding each identified hotspot.
388269
- Web parity checklist covers every cross-domain shell route, settings panel,
389270
Finance view, and Knowledge view currently available on mobile.
390-
- Platform quality regressions are caught by a targeted widget or integration
391-
test added for each identified hotspot.
392271

393272
## Triggered Bets
394273

0 commit comments

Comments
 (0)