pacphi · pacphi · Aug 5, 2026 · Aug 5, 2026
diff --git a/docs/DASHBOARD.md b/docs/DASHBOARD.md
@@ -23,7 +23,7 @@ Its position does not move when the primary area changes; only the choices insid
 | Overview | Hosts & Routing | `#overview/hosts` | Hosts & routing | Enabled execution hosts, activity assignments, primary-host policy, and escalation paths |
 | Overview | Providers | `#overview/providers` | Inference providers | Provider bindings, availability, provenance, and configuration health |
 | Overview | Runtime | `#overview/runtime` | Runtime health | Local services, MCP connections, processes, and operational readiness |
-| Overview | Intelligence | `#overview/intelligence` | Intelligence & learning | Memory, learned patterns, quality feedback, and improvement signals |
+| Overview | Intelligence | `#overview/intelligence` | Intelligence & learning | Memory, learned patterns, reasoning-graph growth, and improvement signals, updated near-live while the view is open |
 | Usage | Scorecard | `#usage/score` | Usage scorecard | Token consumption, API-equivalent cost, efficiency, and trends |
 | Usage | Limits | `#usage/limits` | Provider limits | Current provider windows, reset timing, and available capacity |
 | Usage | Findings | `#usage/findings` | Usage findings | Actionable anomalies, efficiency opportunities, and evidence-backed recommendations |
@@ -61,7 +61,13 @@ Overview keeps status and routing in one health-first area:
   of which inference provider served a particular session.
 - **Providers** presents inference-provider bindings and their configuration provenance.
 - **Runtime** presents operational services, processes, and MCP readiness.
-- **Intelligence** presents memory, learning, and quality-improvement signals.
+- **Intelligence** presents memory, learning, and quality-improvement signals: the neural pattern
+  store's current size, its separate lifetime patterns-learned counter, reasoning-graph growth, and
+  the route-learner's improvement delta. It reads files ruflo/agentic-qe already write under
+  `.claude-flow/` and updates near-live over its own SSE stream while the view is open, falling back
+  to the general status poll otherwise. See [Project intelligence](ddd/project-intelligence.md) and
+  [ADR-0024](adr/0024-project-intelligence-telemetry.md) for the full model and the two learning
+  metrics' load-bearing distinction.
 
 ## Usage
 

diff --git a/docs/adr/0024-project-intelligence-telemetry.md b/docs/adr/0024-project-intelligence-telemetry.md
@@ -0,0 +1,164 @@
+# ADR-0024 — Project intelligence: live learning telemetry from ruflo/agentic-qe's own state
+
+- **Status:** Implemented
+- **Date:** 2026-08-05
+- **Deciders:** agentic-kit maintainers
+- **Related:** [ADR-0005](0005-dashboard-in-page-routing-reveal.md),
+  [ADR-0009](0009-usage-scorecard-local-transcript-analytics.md),
+  [ADR-0012](0012-observability.md)
+
+## Context
+
+Overview's **Intelligence** destination (`#overview/intelligence`, added by
+[ADR-0005](0005-dashboard-in-page-routing-reveal.md)'s 2026-08-04 information-architecture
+amendment) has always advertised "memory, learned patterns, quality feedback, and improvement
+signals." Until now its "learning over time" strip rendered only two sparklines:
+
+- **patterns learned**, sourced from `.claude-flow/health-history.json` — a ring `dashboard-server.mjs`
+  itself appended to, and only while a dashboard happened to be running to observe
+  `.claude-flow/neural/stats.json`. On a machine (or CI checkout) where the dashboard had never
+  polled long enough to accumulate a ring, this file simply did not exist and the panel showed
+  `no data`;
+- **improvement Δpp**, sourced from the route-learner's existing `.claude-flow/improvement.json`
+  (unchanged by this decision).
+
+Meanwhile ruflo and agentic-qe were already writing two richer, always-present sources that the
+dashboard never read: `.claude-flow/neural/patterns.json` (the neural pattern store — a live JSON
+array of pattern entries, each carrying `createdAt`/`type`) and
+`.claude-flow/data/intelligence-snapshot.json` (point-in-time samples of the reasoning/knowledge
+graph's size: `nodes`, `edges`, `pageRankSum`). `src/commands/status.mjs`'s CLI `learning` row
+already reads `.claude-flow/neural/stats.json` for the same lifetime `patternsLearned` counter the
+health ring's samples happened to carry — but the *store's own current inventory* (how many
+patterns are on disk right now, which shrinks under pruning/compaction even as the lifetime
+counter only ever climbs) had no reader at all. The panel's "no data" state was therefore
+avoidable, not fundamental: real trend data already existed on disk.
+
+Separately, every value on this panel only ever refreshed on the dashboard's general ~30s
+`/api/status` poll (the same cadence used for unrelated subsystem-health cards), so a user
+actively watching learning happen — a pattern being stored, the graph growing — waited up to that
+long to see it.
+
+This telemetry is a different kind of fact than anything [ADR-0012](0012-observability.md)/
+[Observability](../ddd/observability.md) models. It carries no session, actor, host, provider, or
+model identity; no lifecycle (`queued → running → completed`); no per-field evidence confidence
+(`observed`/`correlated`/`inferred`/`assumed`/`planned`); and no transcript content requiring
+redaction. It is four scalar/array-shaped reads over this project's own `.claude-flow/` state —
+the same local trust boundary `ak status` already reads directly, with no host-specific
+anti-corruption adapter needed because there is only ever one shape, ak's own. The panel that
+surfaces it also lives under **Overview**, not under **Observability**'s Live/History scope. The
+implementation deliberately did not construct an `ObservedSession`, did not flow through the
+canonical event normalizer, and did not extend the replay/snapshot cursor — it reuses only
+source-agnostic transport plumbing (`JsonlTailer`, `sseChannel`, `reserveClientSlot`/`clientGone`,
+`transcriptSseFrame`) that Dashboard delivery already shares across contexts.
+
+## Decision
+
+### 1. A new bounded context: Project intelligence
+
+Project intelligence is its own bounded context (see the updated
+[context map](../ddd/context-map.md) and [Project intelligence](../ddd/project-intelligence.md)),
+not an Observability extension and not folded into ADR-0005's navigation-shell amendments. Its
+model, invariants, and the reasoning for keeping it separate from `ObservedSession` are specified
+in that document; this ADR records the decision and its consequences.
+
+### 2. One read-only history module composes four existing sources
+
+`src/lib/dashboard/intel-history.mjs` adds:
+
+- `readNeuralPatternStoreHistory(cwd)` — every entry currently on disk in
+  `.claude-flow/neural/patterns.json`, as `{ createdAt, type }`;
+- `readGraphHistory(cwd)` — every sample in `.claude-flow/data/intelligence-snapshot.json`, as
+  `{ timestamp, nodes, edges, pageRankSum }`;
+- `readGlobalStats(cwd)` — the current cumulative counters in `.claude-flow/neural/stats.json`
+  (`patternsLearned`, `trajectoriesRecorded`, `signalsProcessed`, `lastAdaptation`), reading via the
+  same `readJson` helper and `?? 0` defaulting `status.mjs`'s `learning` row already uses, so the two
+  call sites cannot drift apart;
+- `readHealthRing(cwd)` and `appendHealthSnapshot(cwd, snapshot)` — moved (not duplicated) from
+  `dashboard-server.mjs`, unchanged behavior, now capped at 500 samples with field-level dedup
+  (a repeated poll of unchanged stats writes nothing);
+- `readIntelHistory(cwd)` — the combinator `collectData()` and the SSE route both call, returning
+  `{ patternStore, graph, healthRing, globalStats }`.
+
+The **patterns-learned counter** (`globalStats.patternsLearned`, a lifetime total) and the
+**pattern-store size** (`patternStore.length`, entries actually present right now) are
+independent metrics that may legitimately diverge as the store is pruned or compacted. This
+project's own repository demonstrates the divergence today: 28 live pattern-store entries against
+a 1,337 lifetime counter. Every reader, doc comment, and rendered label keeps the two separate;
+none averages, sums, or substitutes one for the other.
+
+### 3. Push updates over a new SSE route, additive to the existing poll
+
+`src/lib/live/intelligence-watch.mjs` adds `IntelligenceWatch`, which polls the three source files'
+`mtime` every second (default), corroborated by a change-only tail of
+`.claude-flow/data/pending-insights.jsonl` (via the existing `JsonlTailer`, whose line *contents*
+are never read — a record arriving at all is the signal), and flushes a debounced
+(2.5s trailing-edge, measured from the most recent detected change) `onUpdate(readIntelHistory(cwd))`
+call. `dashboard-server.mjs` exposes this as `GET /api/live/intelligence`: one `event: init` frame
+with the current `readIntelHistory(cwd)` on connect, then an `event: update` frame per flush,
+fanned out to every connected client. It reuses Dashboard delivery's proven SSE discipline —
+reservation-before-await client-cap tracking, the forwarding-cleanup idiom, `sseChannel` backpressure
+and heartbeat — but sends `transcriptSseFrame` payloads, not `sseFrame`'s session-privacy-redacted
+ones, because this payload was never session or transcript content in need of that redaction
+pipeline.
+
+`GET /api/live/intelligence` shares Dashboard delivery's transport primitives with, but is not part
+of, [Observability](../ddd/observability.md)'s `/api/live`, `/api/live/events`,
+`/api/live/transcripts/:host/:id/events`, and `/api/live/playback/:host/:id` family documented in
+[OBSERVABILITY.md](../OBSERVABILITY.md); it is not covered by that document's evidence, privacy, or
+capability-coverage contract, and unlike Observability's ruflo/agentic-qe sources, it requires no
+`--live-source` registration — the four files it reads are always this project's own.
+
+### 4. `/api/status` keeps working as the fallback path
+
+The SSE route is additive. `collectData()`'s existing return gains `globalStats`, `patternStore`,
+and `graph` alongside the unchanged `health` field (still `intel.healthRing`); a client without
+`EventSource` support, or one that has not yet opened the stream, still gets the full picture on the
+next poll. The `/api/status` error-fallback payload was extended with the same three keys
+(`globalStats: null, patternStore: [], graph: null`) so its shape never diverges from the success
+path.
+
+## Consequences
+
+### Positive
+
+- The Intelligence panel shows real trend data — pattern-store growth and reasoning-graph growth —
+  sourced from files that already existed, at zero new collection cost and no new write path beyond
+  the existing, now-relocated `appendHealthSnapshot`.
+- Users watching active learning see updates within the debounce window (≤ ~3.5s) instead of waiting
+  out the general status poll.
+- The lifetime-counter-vs-store-size divergence is now visible and labeled instead of silently
+  absent or conflated.
+- A genuinely different domain got its own bounded context instead of stretching
+  `ObservedSession`'s Session/Actor/Activity model, or ADR-0005's navigation-shell decision, to cover
+  facts neither was designed to grade.
+
+### Negative
+
+- A second `/api/live/*`-prefixed SSE endpoint and client-cap surface to operate, alongside
+  `/api/live/events` and `/api/live/transcripts/...`.
+- `mtime`-based polling can in principle miss two rewrites that land on the same filesystem-reported
+  millisecond; the `pending-insights.jsonl` tail is a second, independent trigger but not a formal
+  guarantee.
+- No per-field confidence grading exists for this data, unlike Observability's evidence model —
+  accepted because every source is this project's own local file, not another host's evidence
+  requiring provenance.
+
+### Risks and mitigations
+
+| Risk | Mitigation |
+|------|------------|
+| Pattern-store size and the lifetime counter get conflated in a future edit or display | Documented domain invariant, shared header comment across both modules, and separate rendered figures with distinct captions |
+| Debounce coalesces a burst into a stale window | Trailing-edge debounce measured from the most recently detected change, not a fixed interval |
+| A reader is called on this project's own untrusted-shaped JSON (a partially written file, a schema drift) | Malformed or non-array data degrades to `[]`/`null`, matching `readJsonSafe`'s existing null-on-absent convention; individual malformed entries are skipped rather than throwing |
+| New SSE route reintroduces a client-cap race | Reuses the same `reserveClientSlot`/forwarding-cleanup idiom already hardened for `/api/live/events` |
+| Readers drift from `status.mjs`'s CLI `learning` row over time | `readGlobalStats` calls the same `readJson` helper and `?? 0` default `status.mjs` uses, not a hand-copied reimplementation |
+
+## References
+
+- `src/lib/dashboard/intel-history.mjs`, `tests/kit/intel-history.test.mjs`
+- `src/lib/live/intelligence-watch.mjs`, `tests/kit/intelligence-watch.test.mjs`
+- `src/lib/dashboard-server.mjs` (`collectData`, `GET /api/live/intelligence`, `lazyIntelWatch`)
+- `src/lib/dashboard/client.mjs`, `src/lib/dashboard/page.mjs` (Intelligence panel rendering, SSE
+  subscription)
+- [Project intelligence domain](../ddd/project-intelligence.md)
+- [Dashboard guide](../DASHBOARD.md)
diff --git a/docs/adr/README.md b/docs/adr/README.md
@@ -32,6 +32,7 @@ Consequences**, and cites the grounded source it rests on where relevant.
 | [0021](0021-inference-provider-provenance.md) | Inference-provider provenance for live sessions | Accepted |
 | [0022](0022-metaharness-as-optional-assurance-companion.md) | MetaHarness as an optional assurance companion | Proposed |
 | [0023](0023-fail-closed-operations-and-explicit-degradation.md) | Fail-closed mutations and explicit degraded operation evidence | Implemented |
+| [0024](0024-project-intelligence-telemetry.md) | Project intelligence: live learning telemetry from ruflo/agentic-qe's own state | Implemented |
 
 Theme: ADRs **0001–0006** define **dual-host LLM routing and leadership** — how `ak` lets ruflo route
 each development activity (architecture, implementation, testing, review, …) to the right host (Claude
@@ -121,3 +122,12 @@ managed fallbacks report degradation, SQLite retains classified failure evidence
 usage, promised backups fail closed before atomic replacement, status-line failures gain redacted
 opt-in diagnostics, process discovery is current-user and argv-minimized, setup discloses and
 verifies its project auto-approve manifest, and clean-machine tests isolate every mutable path.
+
+**0024** gives Overview's Intelligence view real trend data instead of a permanently-empty strip:
+a new `intel-history.mjs` module reads the neural pattern store, its lifetime learned-pattern
+counter, and reasoning-graph size samples that ruflo/agentic-qe already write under
+`.claude-flow/`, while a debounced `IntelligenceWatch` pushes near-instant updates over a new
+`GET /api/live/intelligence` SSE route additive to the existing status poll. It establishes Project
+intelligence as its own bounded context rather than an Observability extension — this telemetry
+carries no session, actor, host, provider, or lifecycle identity and needs no per-field evidence
+confidence, and the panel it feeds lives under Overview, not Observability's Live/History scope.
diff --git a/docs/ddd/README.md b/docs/ddd/README.md
@@ -12,6 +12,7 @@ describe the current system unless a section is explicitly marked as future work
 | [Integration management](integration-management.md) | Hosts, inference providers, bindings, capabilities, lifecycle, facts, and ownership |
 | [Routing and orchestration](routing-and-orchestration.md) | Activities, routes, leadership, escalation, projections, and canonical `ak run` execution |
 | [Observability](observability.md) | Evidence acquisition, observed-session aggregates, replay, and dashboard delivery |
+| [Project intelligence](project-intelligence.md) | Pattern store, learning counters, reasoning-graph size, and live delivery for Overview's Intelligence view |
 
 ## Relationship to other documentation
 

diff --git a/docs/ddd/context-map.md b/docs/ddd/context-map.md
@@ -20,6 +20,8 @@ Native Evidence ----> Evidence Acquisition ----> Canonical Evidence
                     |         +----> Workspace Snapshot Cache             |
                     |                       |                              |
                     +-----------------------+----> Dashboard Delivery <----+
+                                                          ^
+Project State (.claude-flow/*) ----> Project Intelligence ┘
 
 Maintainer Administration is a separate, deliberately-egressing context.
 ```
@@ -74,11 +76,22 @@ an advisory read-model cache, not the append-only Evidence Archive and not a sou
 Restoration supplies inert History context using the original capture time; it cannot query a
 current checkout and present that state as historical.
 
+### Project intelligence
+
+Owns read-only trend projections over ruflo/agentic-qe's own project-level learning state: the
+neural pattern store, its lifetime learned-pattern counter, reasoning-graph size samples, and the
+machine-health sample ring. It reads `.claude-flow/*` files directly — there is only ever one
+shape, ak's own, so no anti-corruption adapter is required — and is independent of Evidence
+Acquisition and Observability's canonical event model. It carries no session, actor, host,
+provider, or lifecycle identity and grades no per-field evidence confidence.
+
+See [Project intelligence](project-intelligence.md).
+
 ### Dashboard delivery
 
 Owns protected HTTP/SSE delivery, browser DTOs, filters, presentation, and interaction state. It
-may combine read models from Observability, Historical Usage, routing, and integration facts. It
-cannot manufacture or strengthen domain facts.
+may combine read models from Observability, Historical Usage, Project Intelligence, routing, and
+integration facts. It cannot manufacture or strengthen domain facts.
 
 ### Maintainer administration
 
@@ -99,6 +112,8 @@ and credential policy is distinct from the offline-first dashboard and integrati
 | Observability | Workspace snapshot cache | Last safe metadata-only session workspace capture |
 | Workspace snapshot cache | Dashboard delivery | Inert last-recorded History context after restart |
 | Historical usage | Dashboard delivery | Historical aggregates and findings |
+| Project state (`.claude-flow/*`) | Project intelligence | Direct local file reads; no anti-corruption adapter needed |
+| Project intelligence | Dashboard delivery | Read-model projection, delivered by poll (`/api/status`) and SSE push (`/api/live/intelligence`) |
 
 ## Boundary rules
 
@@ -108,3 +123,6 @@ and credential policy is distinct from the offline-first dashboard and integrati
 - Dashboard presentation cannot upgrade provenance.
 - Historical usage and live topology share identifiers, not aggregate ownership.
 - Network egress occurs only in commands and contexts whose contract explicitly permits it.
+- Project intelligence reads local project state directly; it never enters Evidence Acquisition's
+  anti-corruption layer or Observability's canonical event model, and it establishes no session,
+  actor, host, provider, or lifecycle identity.