Skip to content

Improve Wiki/Project quality and lazy concept-first topic curation - #3225

Merged
Hillary Mutisya (hillary-mutisya) merged 2 commits into
mainfrom
dev/hillarym/memory-pr2-quality
Oct 26, 2026
Merged

Hillary Mutisya (hillary-mutisya) merged 2 commits into
mainfrom
dev/hillarym/memory-pr2-quality

Conversation

@hillary-mutisya

Copy link
Copy Markdown
Collaborator

Scope

Standalone extraction from the immutable integration source at 886f3a3, based on main 8b92819. This branch does not require the schema Questions/acquisition PR.

  • Model-authored substantive Wiki/Project summaries and completeness; grounded project assignments, milestones, decisions, risks and open questions.
  • Exact queried-subject Wiki admission, semantic contribution/consolidation checks, and bounded summary/section-body-only repairs preserving citations, anchors, protected structures and human edits (including adapter receiver binding).
  • Complete raw topic incidence alongside semantic concept admission/reconciliation, source-local evidence-bounded packing, strict membership/provenance accounting, omission repair and durable partial/completed curation checkpoints.
  • Explicit, independently audited lazy concept summaries with request coalescing, durable caching, fingerprint/eligibility invalidation and cancellation/privacy guards. Existing topic browsing exposes concepts, raw members and evidence without a new graph explorer.
  • Quality-relevant source-synopsis contracts and deterministic fixture alignment; maintenance quiet periods/coalescing adapted to main without blocking its unpublished-index segment collection.

Deliberate exclusions

No schema question Views/acquisition, knowledge coordinator/discovery/checkpoint APIs, integrated source-plus-view answering, new graph projection/query APIs or graph explorer UI, unrelated Views/Inbox redesign, or benchmark orchestration changes. Native curation latency and source-synopsis model failures remain follow-up work; this is not live-model, LongMemEval quality, or performance qualification.

Local validation

  • Full pnpm run build: passed.
  • Requested offline quality/topic suites exercised; the full memory sweep exercised 926 tests (921 initially passed). All five initial failures in three suites were corrected (prompt/mock contract alignment and main-specific ingestion/segment-GC scheduling); the final eight-suite lifecycle/retrieval/topic recheck passed 84/84, including every originally failing test. Other quality suites passed in the sweep.
  • Complete memory client tests: 37/37; actual in-process/RPC/MCP topic contract: 2/2.
  • Complete browser client suite: 304/304; affected gateway/lifecycle/query tests: 44/44.
  • Focused offline source-synopsis identity/composition fixture: 1/1 (no replay or live model).
  • Pinned formatting and all four ratchets against origin/main: lint, complexity, circular dependencies, and debt passed. Circular validation ran only after Git-fixture tests stopped.

Integration notes

Expected hunk-level overlap with the independent schema Questions PR: fileMemoryService.ts, viewStore.ts, viewTypes.ts, viewSynthesis.ts, viewBuilds.ts, and client viewProtocol.ts. No schema/coordinator infrastructure is needed for this branch. Existing main storage/import behavior is retained and regression-tested.

Extract source-grounded Wiki/Project summaries, scoped admission and bounded canonical repairs. Preserve raw topic incidence, durable semantic curation and explicit audited lazy summaries without schema questions, the knowledge coordinator or the graph explorer. Adapt maintenance coalescing to preserve main's import and segment-collection lifecycle.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
@hillary-mutisya

Copy link
Copy Markdown
Collaborator Author

Final verification

Verified exact latest HEAD 4a8dd2dba8e00e304ce293789391d696aefe453e after CI completed:

  • All 13 required checks passed: nine Windows/Linux/macOS TypeScript and test matrix jobs across Node 22/24/26, .NET Debug/Release, repository policy, and Azure TypeAgent Smoke Tests (Linux and Windows).
  • All eight latest-head GitHub workflow runs completed successfully: build-ts, build-dotnet, build-package-shell, repo-policy-check, CodeQL Advanced, format-pr, docs-pr-amend and Prepare Generated File Healing. CLA also passed. The fork-only formatting check was intentionally skipped by its workflow; the applicable formatting workflow passed.
  • No remote CI remediation, reruns, gate weakening or deadline changes were needed. Local extraction fixes and the curated Overview request-fixture follow-up are documented above; this does not recast the initial local 921/926 sweep as an all-green run.
  • Local full pnpm run build, pinned formatting, and all four committed gates passed: pnpm run code-lint -- --ratchet --base origin/main; pnpm run code-complexity:ci; pnpm run code-circular -- --ratchet --base origin/main --exceptions-file tools/scripts/code/circular-baseline-exception.json; pnpm run code-debt -- --gate --base origin/main.
  • Final local rechecks passed: lifecycle/retrieval/topic/prompt 84/84, memory client 37/37, in-process/RPC/MCP topic contract 2/2, browser client 304/304, gateway/lifecycle/query 44/44, deterministic source-synopsis fixture 1/1.
  • Own worktree is clean; initial main 8b9281905389d39dbc568f389393b665745cb739 is an ancestor. Cumulative extraction is 64 files, +6,627/-571. No dependency on PR1; expected shared-file overlap is listed in the description.

No known required CI checks remain unresolved. The PR remains open for human review; no merge or auto-merge was requested. Native curation latency/source-synopsis failures remain follow-up; no live-model, LongMemEval or performance qualification is claimed.

@hillary-mutisya
Hillary Mutisya (hillary-mutisya) merged commit e24bd0d into main Oct 26, 2026
30 checks passed
Hillary Mutisya (hillary-mutisya) added a commit that referenced this pull request Oct 26, 2026
…ed answering (#3226)

## Summary

PR3: automatic knowledge discovery, readiness checkpoints and integrated
answering, based on main after #3224 and #3225.

- Adds bounded corpus-level question proposal, utility/admission and
routing with persisted receipts, fingerprints, existing-contract
suppression, hourly call budgets and allowance-reserving repair.
Existing corpora remain off until enabled; new corpora default to
`sourceObserved`. Policy admission does not confirm answers or authorize
tools.
- Adds fixed-source/family readiness checkpoints, producer-complete
bounded continuation and explicit
partial/blocked/failed/superseded/cancelled states. Coordinates
ingestion, maintenance, retry, invalidation, privacy removal and
shutdown. Lazy curated-topic summaries do not require prose for
readiness; complete curation coverage is still required.
- Adds integrated answering over originals and eligible published Views:
exact retained-source expansion, separate derived/user-assertion
evidence, authority/review/freshness/conflict/scope/eligibility checks,
original-source citations, evidence/request bounds, bounded follow-up
retrieval/refinement and opt-in diagnostics.
- Wires service/client/protocol/RPC/MCP and browser operation
inventories, browser answering, Questions knowledge status/policy
controls and source-plus-View answer rendering.
- Adds deterministic qualification-harness checkpoint orchestration and
metadata-only profiling fidelity tests.

## Boundaries

Extracted feature commits and later scoped repairs from immutable
integration source, reconciled against current main instead of replacing
shared files with older versions. Preserves merged question inventories,
answer-quality repairs and lazy topic behavior.

Excludes PR4's scoped graph APIs/projections/entity-event/hierarchy
work, graph UI/RPC/MCP, generic Views/Inbox/navigation redesign and lazy
read-projection optimizations. No PR4 branch dependency.

No live inference, pilot, replay, LongMemEval evaluation or performance
qualification was run. The previous stopped pilot is not a completed
qualification or scoring result.

## Validation

Commands run from `ts` unless a package directory is specified:

- `pnpm run build` — full workspace passed.
- Memory service: `pnpm run jest-esm --runInBand
--testPathPattern='(knowledgeCoordinator|knowledgeDiscovery|knowledgeLifecycle|knowledgeAnswering|integratedAnswer|schemaViews|topicService|viewBuilds|viewDrafts|informationalWiki|viewSynthesisEvidence|wikiContribution)[.]spec[.]js'`
— 301 tests passed.
- Memory service: `pnpm run jest-esm --runInBand
--testPathPattern='(answerKnowPro|fileMemoryService|markdownPersistence|maintenanceScheduler|topicGeneration|topicFamily|corpusIngestionIsolation)[.]spec[.]js'`
— 86 tests passed.
- `pnpm --filter @typeagent/memory-client test:local` — 49 passed.
- `pnpm --filter @typeagent/memory-mcp-server test:local` — 38 passed;
aligned timeout/controlled-model fixtures additionally rechecked with
focused knowledge/inventory tests.
- Browser: `pnpm run test:views:gateway` — 102 passed; final timeout RPC
test additionally rechecked.
- Browser: `pnpm run test:views:client --
--testPathPattern='memoryHub(KnowledgeStatus|Questions|Search)[.]test[.]ts'`
— 53 passed; `pnpm run test:views:client` — all 339 tests passed after
adding malformed-status isolation.
- `node --test
tools\memory-maintenance-qualification\longmemeval-runner.test.mjs
tools\memory-maintenance-qualification\ingestion-profile.test.mjs` — 43
deterministic tests passed.
- `pnpm run prettier:changed` — passed, pinned repository formatter.
- `pnpm run code-lint -- --ratchet --base origin/main` — passed, 281 →
281 existing violations.
- `pnpm run code-complexity:ci` — passed, no new over-cap functions.
- `pnpm run code-circular -- --ratchet --base origin/main
--exceptions-file tools/scripts/code/circular-baseline-exception.json` —
passed, 210 → 210 cycles (11 baseline exceptions), run without
overlapping own Git-fixture tests.
- `pnpm run code-debt -- --gate --base origin/main` — passed, no
focused/new skipped tests.

## Remediation

Preserved current-main cancellation/scheduler and answer-quality fixes
during extraction. Updated complete MCP operation inventory, replaced
legacy KnowPro answer-generation fixtures with deterministic
integrated-reader fixtures, isolated existing controlled View models
from unrelated automatic discovery, and aligned transport timeout
validation with the service's bounded 1–3,600,000ms contract. Complete
Hub client testing also identified a malformed knowledge-response
failure escaping into Questions; preparation now retains source progress
and reports the failure locally.

## Latest-head CI

Verified head `b7dcb1065b5f77fffbf6cf5f9be52dc5816ed89d`: all required
checks pass. All nine TypeScript OS/Node lanes, both .NET
configurations, all three shell packaging lanes, formatting, repository
policy, CodeQL, generated-document regeneration and external
Linux/Windows smoke tests pass. The fork-only formatting check is
intentionally skipped; no actionable failures or pending checks remain.
No merge or auto-merge was requested.

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Hillary Mutisya (hillary-mutisya) added a commit that referenced this pull request Oct 26, 2026
## Summary

- Add bounded, corpus-qualified document/event entity, topic and
maintained-hierarchy graphs. Eligibility, suppression, event authority,
exact provenance/counts, generation-bound node/edge/support paging,
budgets and explicit building/stale/error/unsupported coverage remain
visible.
- Separate index preparation from reads; invalidate graph caches across
document/event lifecycle changes. Reuse merged curated topics/lazy
summaries, preserve current source fingerprints and historical conflict
fallback, and revalidate detached view/event preparation before
installing results.
- Expose graph operations through service, in-process/MCP clients,
RPC/browser gateway and strict protocols. Advertise graph support
truthfully, including host-specific MCP tool inventories.
- Add scoped Explore/overview, graph controls/visualizers/details and
safe evidence drilldown. Finish generic Views card/list-detail/build
flyout, grouped activity/Inbox/navigation and lazy summary inventories
with responsive detail readers.

## Independent base and exclusions

Targets **main**, not PR3. Prerequisites #3224 (Questions/acquisition)
and #3225 (quality/curated lazy topics) are merged. Extracted from the
immutable integration source using two-tip comparisons while retaining
main's later fixes. No PR3 coordinator, automatic corpus question
discovery, knowledge checkpoint/policy/readiness API/UI, integrated
source-plus-view answering, or benchmark harness integration is
included. Existing Questions functionality is retained.

## Validation

All final local checks passed in the isolated PR4 worktree:

- `pnpm run build` from `ts` (full workspace).
- Complete memory-service offline unit suite: **80 suites, 1,030
tests**. From `ts/packages/memory/service`: `pnpm run jest-esm
--maxWorkers=2 --workerIdleMemoryLimit=512MB
--testPathPattern=".*[.]spec[.]js$"` (local
`NODE_OPTIONS=--max-old-space-size=8192`). Includes
graphService/Projection/Paging, viewIndexConcurrency, snapshot
reuse/freshness/historical conflict parity, topic/event/privacy
lifecycle, all view/editorial/consolidation and other existing service
regressions.
- `pnpm --filter @typeagent/memory-client run test:local`: **48 tests**.
- `pnpm --filter @typeagent/memory-mcp-server test`: **12 suites, 39
tests** (real bounded MCP protocol round trips and host inventory
included).
- `pnpm --filter @typeagent/browser run test:views:gateway`: **111
tests**, including scoped graph gateway and strict validation/lifecycle.
- `pnpm --filter @typeagent/browser run test:views:client`: **41 suites,
371 tests**, including
graph/Explore/Views/activity/Questions/management.
- Pinned `pnpm run prettier:changed`.
- `pnpm run code-lint -- --ratchet --base origin/main`.
- `pnpm run code-complexity:ci`.
- `pnpm run code-circular -- --ratchet --base origin/main
--exceptions-file tools/scripts/code/circular-baseline-exception.json`
(not concurrent with Git-fixture tests).
- `pnpm run code-debt -- --gate --base origin/main`.

Coverage is deterministic/offline; no live inference, qualification
pilot, recording or long-lived server was used. These checks do not
establish generated knowledge/answer quality. Read-only graph responses
may report separately scheduled preparation or unsupported provider
projections rather than claim completeness.

## CI remediation

The first matrix run exposed two pre-projection hydration-count
expectations in snapshotReuse. The follow-up ordinary commit asserts one
fresh projection read and zero full-state hydrations, while retaining
expired-scope freshness, privacy checks and historical consolidation
parity. The complete 80-suite service run passes after this fix; no test
or gate was disabled.

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant