01 / 08 — CASE STUDY
CONCEPT VISUAL · GENERATED

StatusActive · source snapshot

Year2026

RoleAuthor · Architect · Sole Engineer

StackTypeScript · Node · SQLite · MCP

Duration2025 → ongoing

TypePersonal · open source

Handoff note Source evidence

IGRIS

An engineering OS for AI coding agents. Durable work state, explicit ownership, and one brain beyond the conversation.

THE WORK

// 01 · PROBLEM · APPROACH · OUTCOME

01// PROBLEM

The session ends. The work does not.

Usage limits, crashes and tool switches leave the next session reconstructing decisions, failed checks and ownership. A transcript is not a handoff.

IGRIS keeps the brief, owner and stopping point outside the model. The repository documents a same-machine Claude → Codex handoff; it is a demonstration, not a measured recovery rate.

02// APPROACH

State first. Constraints at the boundary.

A local SQLite brain stores briefs, claims, sessions and knowledge. MCP and harness adapters project that state into the tools around the model. Project standards and the working tree remain separate authorities.

Hooks enforce workflow checks on supported harnesses; Codex, Gemini CLI and Cursor use advisory bridges. Local atomic claims prevent duplicate local ownership, not cross-machine conflicts.

03// OUTCOME

An inspectable system, not a black box.

Seven cognition extractors turn transcripts and brain records into candidate knowledge. Review is the default for semantic changes; optional automatic actions and deterministic hygiene are explicit exceptions.

The dashboard exposes graph, search, triage and diagnostics. This source snapshot includes working-tree hardening, not a published release or a fresh test run. Cross-machine handoff and phase resume still need stronger evidence.

ARCHITECTURE

// 02 · DURABLE STATE · REVIEW PATH

IGRIS default cognition review path. local brain records and transcripts feed a shared host, then pending learnings and suggestions reach operator review through the local dashboard. a chosen mapped apply action returns through the MCP gateway and component handlers to the same store. handlers also associate supported maintenance with pre-state, undo and audit. notes preserve automatic-action and hygiene exceptions and fail-soft undo capture. — generated default semantic-change path, not a product screenshot or a map of all mutations. brain records and transcripts feed one shared cognition host; validated candidates enter pending queues, not approved knowledge. an operator reviews through the local dashboard and may reject, dismiss or choose a supported mapped apply action. the action reaches the MCP gateway and component handlers before memory or graph changes persist. gateway, handler and store labels refer to the same boundaries in the durable architecture figure. pure database readers separately support dashboard inspection and opening record detail; the queue-to-review arrow describes a human interaction, not a requirement to read through the apply gateway. optional automatic actions and deterministic hygiene are exceptions to default review, not hidden bypass arrows in this slice. supported maintenance carries pre-state, undo and audit accompaniment, but capture can fail softly and recovery is not guaranteed. enablement, budget, timeout and process-isolation limits remain in the accompanying prose; isolation is not a kernel security sandbox. source snapshot: 2026-09-06; inspected implementation, not execution or accuracy evidence. [C01–C07, K01, K02, K06, K07, U05]

STACK

// 03 · WHAT IT'S MADE OF

CoreTypeScript · Node

PersistenceLocal SQLite · guarded disk cache

GatewayMCP · component handlers

RetrievalBM25 + vectors · rank fusion

GraphTyped entities · qualified edges

CommunitiesLouvain · cluster pass default off

CognitionShared host · isolated subprocesses

HarnessesAdapters · capability-specific gates

PortabilityChecked bundles · optional replication

PROCESS

// 04 · INSPECT THE WORK

IGRIS cognition diagnostics list perception, subconscious, synapse, janitor, arbiter, curator and cartographer. the cartographer row has a yellow NO SIGNAL warning; other rows show terminal event and queue details. the hostname and database-path footer are outside the crop. — genuine local diagnostics, captured 2026-09-06. the registered extractor roster shows run evidence, schedules and queue populations. cartographer remains `NO SIGNAL`: the retained local event window is silent, while an other-host observation is separate. these states are a snapshot, not extractor accuracy or proof that every job ran here.

the IGRIS whole-brain graph displays a large circular field of pale nodes with dense central clusters and connecting lines. the scope is whole brain, the refresh control is visible, and the inspector says nothing selected. individual record labels are not visible. — genuine whole-brain graph, captured 2026-09-06 with all projects and entity types unfiltered. native pan and zoom expose the dense topology with nothing selected. counts belong to this snapshot, not a benchmark. the query freshness footer is outside the viewport; this image does not certify an exact query time or record-level readability.

three IGRIS overview cards show brief states, priorities and instances, with the igris-ai scope label above. the private project registry and graph scale card are cropped out. — genuine overview detail for the igris-ai scope, captured 2026-09-06. brief states, priorities and instance records describe this local snapshot. these counts do not establish delivery speed, recovery quality, active-process liveness or a permanent project total.

NUMBERS

// 05 · AUDITED SCOPE · NOT PERFORMANCE

// Feature evidence

source-mapped rows · not 57 shipped claims

// Cognition extractors

bounded jobs · one execution host

// Dashboard routes

overview · graph · layers · triage · search · diagnostics

// Frontier entries

working-tree, experimental or deferred scope

LESSONS

// 06 · WHAT TO TAKE FORWARD

State survives a change of model.

Persist the job card, ownership and last inspection result. A new context window should resume work, not reconstruct a conversation.

Name the enforcement boundary.

Hooks and advisory bridges are not equivalent. Workflow checks have exemptions and bypasses; they are not a security perimeter or distributed consensus.

Show inference before calling it knowledge.

Review defaults, provenance and maintenance history make proposals inspectable. Optional auto-apply needs separate judgment; process isolation is not a kernel sandbox.

Earn the next claim.

The audit inspected source and test definitions, not fresh test results. Dashboard captures came from an existing bundle. Repeatable phase resume and cross-machine handoff remain evidence to earn.

Implementation & source evidence 57 features · qualifications · source notes

the matrix is the companion reference, not headline copy. I means inspected implementation; I/T adds relevant inspected test source; WIP/T identifies current uncommitted or untracked implementation with test source. none means freshly passed tests or released behavior. experimental, frontier and deferred states are explicit. evidence paths are relative to the source repository. test witnesses are inspection receipts, not execution results.

Delivery, persistence and portability18 features

Delivery, persistence and portability — implementation and source receipts
Key	Feature / status	Mechanism	Source receipt	Qualification
D01	local-first engineI/T	current component factories register memory, errors, projects, context, sessions, briefs, edges, goals, instances, sync, cache, schedules, cognition, monitoring and catalog.	`brain-mcp-server/src/engine/index.ts:67-141`; `brain-mcp-server/src/engine/__tests__/legacy-bridge.test.ts`	retired worker, coordination and metrics components are not current factories; source/test presence is not a running registration result.
D02	strict tool inputI/T	one gateway rejects missing required keys and extra keys for strict tools.	`brain-mcp-server/src/engine/gateway.ts:126-196`; `brain-mcp-server/src/engine/__tests__/gateway-strict-input.test.ts`	this is not complete schema validation or universal authorization.
D03	bounded delivery/recoveryI	architect plans, forger builds, sentinel verifies, warden reviews; mender assists diagnosis before re-test and review; orchestrator commits.	`core/skills/hunt/SKILL.md`; `core/agents/sentinel.md:1-19`; `core/agents/warden.md:1-19`; `README.md:74-100`	sentinel has Bash; warden declares read-only tools. role separation does not make all writes impossible or establish a recovery rate.
D04	brief-first covered writesI/T	a pre-tool hook can deny covered edits without qualifying brief state.	`core/hooks/shared/pre_tool_use.sh:291-310`; `core/hooks/shared/pre_tool_use.sh:348-501`; `test/phase_guard.test.bash` is a witness for the separate phase gate	exemptions, unknown-project no-op, filesystem fallback, explicit bypass and final fail-soft behavior exist. arbitrary shell writes are not contained.
D05	phase and closing commitI/T	pre-commit phase checks and closing-footer checks cover acceptance criteria, named-role agent events, commit format and secret scanning.	`scripts/git-hooks/pre-commit`; `scripts/git-hooks/commit-msg:131-332`; `test/phase_guard.test.bash`; `test/agent_event_gate.test.bash`	one-shot bypasses exist. the configured committing sequence is not an unbypassable review guarantee.
D06	brief ownershipI/T	conditional SQL update grants a local claim, permits owner re-entry and refuses a different owner.	`brain-mcp-server/src/engine/components/briefs/handlers.ts:32-90`; `brain-mcp-server/src/engine/components/briefs/__tests__/claim.test.ts:111-165`	atomicity is local to the database, not global consensus across replicas.
D07	per-instance handoffI/T	session gathering restores resume fields and distinguishes rested work, live siblings and abandoned live state.	`cli/src/verbs/session.ts:105-121`; `cli/src/verbs/session.ts:312-338`; `cli/src/verbs/session.ts:408-562`; `cli/src/__tests__/session-gather.test.ts:192-325`; `core/skills/boot/SKILL.md:115-142`	a live sibling is not free work; there is no single shared CURRENT_SESSION.md authority.
D08	crash awarenessI/T	process id plus start time checks local liveness without confusing reused ids.	`cli/src/lib/process-liveness.ts:17-65`; `core/skills/boot/SKILL.md:241-248`; `cli/src/__tests__/session-gather.test.ts:279-325`	remote or missing metadata remains unknown. staleness does not authorize claim theft or instance deletion.
D09	stable machine identityWIP/T	minted machine id takes precedence; null-id legacy rows can fall back to hostname aliases.	`brain-mcp-server/src/machine-identity.ts:18-65`; `cli/src/lib/machine-identity.ts`; `brain-mcp-server/src/__tests__/machine-identity-parity.test.ts`; `cli/src/__tests__/machine-identity-sweep.test.ts`	untracked hardening in this snapshot, not verified release content.
D10	personal replicationI/T	asynchronous merges use applicable-table last-write-wins, natural/composite keys and append/dedup variants; local JSONL and brain SQL queues retain pending work.	`brain-mcp-server/src/tools/sync.ts:120-164`; `brain-mcp-server/src/tools/sync.ts:745-852`; `brain-mcp-server/src/tools/sync.ts:948-988`; `cli/src/lib/sync/data.ts`; `cli/src/__tests__/sync-queue.test.ts`	not a CRDT, consensus protocol or cross-machine atomic ownership mechanism; full cross-machine evidence remains frontier.
D11	cross-owner bundlesI/T	tiered project slice → checked archive → ancestry/conflict preview → explicit import policy → receiving brain; checksums, executable/unknown-store rejection and idempotency ledger constrain import.	`cli/src/verbs/export.ts:1-25`; `cli/src/verbs/import.ts:1-22`; `cli/src/verbs/import.ts:79-82`; `cli/src/__tests__/import.test.ts:386-405`	distinct from personal replication. free text can retain sensitive material despite structural checks.
D12	guarded projectionI/T	canonical brain state is projected to disk; rebuild protects newer local brief text unless forced.	`brain-mcp-server/src/engine/components/cache/index.ts`; `brain-mcp-server/src/engine/components/cache/__tests__/projection-guard.test.ts`; `docs/architecture/SYSTEM.md:54-55`; `docs/architecture/SYSTEM.md:70-71`	a guarded cache, not a two-way text CRDT; working tree and standards remain separate authorities.
D13	context and catalogI	context types and a reusable-asset catalog separate experience, project standards and assets; reuse, harvest, ground and promote supply procedural access.	`core/context-doc-types/INDEX.md`; `core/os/knowledge-map.md`; `core/skills/ground/SKILL.md`; `core/skills/harvest/SKILL.md`; `brain-mcp-server/src/engine/components/catalog/index.ts`	these surfaces do not establish automatic architectural understanding.
D14	promotion lineageI	promotion retains a document pointer on the learning.	`brain-mcp-server/src/tools/memory.ts:880-894`; `brain-mcp-server/src/tools/memory.ts:934-944`; `core/skills/promote/SKILL.md`	traceability does not automatically remove every duplicate fact.
D15	descriptor extensionsI/T	declared skills, agents, MCP, hooks and context types feed adapter projections, add-core and generated module indexes.	`cli/src/lib/harness-descriptor.ts:1-30`; `cli/src/lib/harness-descriptor.ts:46-60`; `cli/src/verbs/add-core.ts:10-25`; `core/scripts/gen_os_index.sh:1-36`; `core/scripts/gen_os_index.sh:193-198`; `cli/src/__tests__/harness-descriptor.test.ts`	new emitters or delegation shapes may require code; cognition registration remains explicit.
D16	harness tiersI/T	Claude Code, OpenCode and Antigravity have supported hook projections; Codex, Gemini CLI and Cursor have advisory bridges.	`README.md:160-170`; `cli/src/lib/harness-descriptor.ts:89-100`; `core/hooks/bridges/`; `cli/src/__tests__/antigravity-hooks.test.ts`	static-agent projection, hook gating and cognition backend support are separate capability sets, not parity across harnesses.
D17	doctor and driftI/T + WIP	checks missing/stale core, broken bridges and channel mismatches; working-tree changes add machine-identity diagnostics.	`cli/src/verbs/doctor.ts:1-95`; `cli/src/verbs/doctor.ts:135-141`; `cli/src/lib/drift/`; `cli/src/__tests__/doctor.test.ts`	detection is not universal autonomous repair; current identity hardening is not a release claim.
D18	agent and ceremony evidenceI/T + WIP	brain-timed agent_events and ceremony_events support derived cost/workflow diagnostics; closing checks inspect named-role presence.	`cli/src/lib/kpi-read.ts:20-30`; `cli/src/lib/kpi-read.ts:96-103`; `brain-mcp-server/src/engine/components/instances/index.ts`; `cli/src/verbs/ceremony.ts`; `scripts/git-hooks/commit-msg:222-332`	emissions can fail softly. no measured token, cost, recovery or success outcome follows from event infrastructure alone.

Seven cognition extractors7 features

all rows are implemented with test source, not freshly run. enablement and default/optional behavior follow N02. default review of semantic changes is qualified per row, not a blanket approval promise.

Seven cognition extractors — implementation and source receipts
Key	Feature / status	Mechanism	Source receipt	Qualification
C01	perceptionI/T	transcript → parsed events → bounded extraction → deduplicated learning and embedding in the review channel.	`brain-mcp-server/src/engine/components/cognition/extractors/perception.ts:11-23`; `brain-mcp-server/src/engine/components/cognition/extractors/perception.ts:44-65`; `brain-mcp-server/src/engine/components/cognition/__tests__/perception-instance.test.ts`; `brain-mcp-server/src/engine/components/perception/__tests__/integration.test.ts`	pending and approved retrieval populations differ. dedup uses deterministic/vector dependencies; this is not training. missing configuration resolves on despite fresh-install disabled settings.
C02	subconsciousI/T	deterministic brain digest → bounded inference with identifier validation/confidence cap → open-typed suggestions and optional apply payload.	`brain-mcp-server/src/engine/components/cognition/extractors/subconscious.ts:10-22`; `brain-mcp-server/src/engine/components/subconscious/__tests__/validator.test.ts`; `brain-mcp-server/src/engine/components/subconscious/__tests__/engine-integration.test.ts`	suggestions are reviewable; an apply payload is not proof an action ran. identifier checks reduce risk, not certify truth. current architecture is not the obsolete rule detector.
C03	synapseI/T	learning pairs → relation judgment → pending inferred-edge suggestion or opt-in direct edge.	`brain-mcp-server/src/engine/components/cognition/extractors/synapse.ts:19-26`; `brain-mcp-server/src/engine/components/cognition/extractors/synapse.ts:200-270`; `brain-mcp-server/src/engine/components/synapse/__tests__/apply-roundtrip.test.ts:90-136`	auto-approve defaults false; direct edges are possible when opted in, so proposals-only is not universal.
C04	janitorI/T	normalized-embedding near duplicates → keep/merge/false-positive judgment → reviewed merge.	`brain-mcp-server/src/engine/components/cognition/extractors/janitor.ts:6-32`; `brain-mcp-server/src/engine/components/janitor/__tests__/apply-roundtrip.test.ts`	thresholds are tuning settings, not accuracy results. deterministic confidence and expiry sweeps are separate mutations.
C05	arbiterI/T	same-topic nearest neighbors plus opposition cues → newer-wins/scoped coexistence/evolved merge/non-contradiction → conflict-resolution proposal and application path.	`brain-mcp-server/src/engine/components/cognition/extractors/arbiter.ts:6-31`; `brain-mcp-server/src/engine/components/arbiter/__tests__/td439-merge-guard.test.ts:318-326`	review is normal; automatic resolution is optional and gated. stale-synthesis/hash guards do not make semantic reasoning infallible. shares janitor's flag.
C06	curatorI/T	stale/deprecated candidates → keep/lower-confidence/prune judgment → maintenance proposal.	`brain-mcp-server/src/engine/components/cognition/extractors/curator.ts:7-32`; `brain-mcp-server/src/engine/components/curator/__tests__/undo.test.ts:73-116`	optional auto-prune has an anomaly ceiling; age is not proof of falsehood. shares janitor's flag.
C07	cartographerI/T	deterministic Louvain communities → cluster-level synthesis proposal → meta-learning and member-to-meta edges through application.	`brain-mcp-server/src/engine/components/cognition/extractors/cartographer.ts:7-30`; `brain-mcp-server/src/engine/components/cartographer/__tests__/apply-roundtrip.test.ts:90-147`; `brain-mcp-server/src/engine/components/edges/community.ts:4-29`	uses the proposal/apply maintenance path; cluster sub-toggle defaults off and cadence limits runs. actual primitive is Louvain, never Leiden.

Advanced knowledge machinery16 features

Advanced knowledge machinery — implementation and source receipts
Key	Feature / status	Mechanism	Source receipt	Qualification
K01	shared execution hostI/T	disabled, cold-start, daily-call-budget, minimum-input and empty-context gates precede backend fallback, timeout, prompt wrapping, parse validation and per-candidate persistence.	`brain-mcp-server/src/engine/components/cognition/engine/index.ts:184-307`; `brain-mcp-server/src/engine/components/cognition/__tests__/budget.test.ts`; `brain-mcp-server/src/engine/components/cognition/__tests__/engine.test.ts`	force does not bypass disabled/budget/no-candidate gates. budget counts calls, not calibrated dollars.
K02	subprocess isolationI/T	per-run HOME, empty MCP configuration, forwarded authentication and write-path containment isolate extraction configuration.	`brain-mcp-server/src/engine/components/cognition/backend/isolation.ts:4-27`; `brain-mcp-server/src/engine/components/cognition/backend/isolation.ts:74-125`; `brain-mcp-server/src/engine/components/cognition/__tests__/isolation.test.ts`	not kernel, filesystem or network security isolation; authentication symlinks are read-only by usage, not a universal enforced boundary.
K03	extractor extensionI/T	new instances implement the shared input/prompt/output contract and join an explicit import/registry array.	`brain-mcp-server/src/engine/components/cognition/extractors/index.ts:27-49`	host lifecycle is reusable; arbitrary file drop is not live discovery. the audit reports component extensibility test suites, not a fresh execution.
K04	derived health/yieldI/T + WIP	roster derives instances and classifies disabled, wedged, upstream-blocked, failing and no-signal; separates host, production/review and expiry/judgment populations.	`brain-mcp-server/src/engine/components/cognition/roster.ts:82-93`; `docs/COGNITION.md:76-123`; `docs/COGNITION.md:125-198`; `cli/src/__tests__/dashboard-cognition-endpoint.test.ts:148-172`; `cli/src/__tests__/dashboard-cognition-endpoint.test.ts:305-318`	current health hardening includes working-tree changes. aggregate yield is not unlabeled accuracy or productivity evidence.
K05	dismissal/recurrence feedbackI/T	finding keys block repeated identifier candidates while discriminating different claims about the same entity.	`brain-mcp-server/src/engine/components/subconscious/finding-key.ts:4-25`; `brain-mcp-server/src/engine/components/subconscious/finding-key.ts:43-65`; `brain-mcp-server/src/engine/components/subconscious/__tests__/finding-key.test.ts`; `brain-mcp-server/src/engine/components/subconscious/__tests__/dismiss-loop.test.ts`; `brain-mcp-server/src/engine/components/subconscious/__tests__/recurrence.test.ts`	corpus and tuning affect recurrence; this is not a never-repeats guarantee.
K06	deterministic hygieneI/T	rediscovery confidence bumps, stale-pending expiry, rejection re-evaluation and stale-candidate detection operate without LLM judgment.	`brain-mcp-server/src/engine/components/janitor/hygiene.ts:7-35`; `brain-mcp-server/src/engine/components/janitor/__tests__/hygiene.test.ts`	explicit exception to universal human approval; these are not extra cognition extractors.
K07	maintenance undo/provenanceI/T	captured pre-state supports a transactional inverse, edge removal and invalidation for re-embedding.	`brain-mcp-server/src/engine/components/janitor/undo.ts:4-25`; `brain-mcp-server/src/engine/components/janitor/undo.ts:34-44`; `brain-mcp-server/src/engine/components/curator/__tests__/undo.test.ts:73-206`; `brain-mcp-server/src/engine/components/cartographer/__tests__/apply-roundtrip.test.ts:122-147`	capture can fail softly or lack its table; not every mutation is guaranteed recoverable.
K08	edge-type emergenceI/T	deterministic signature-frequency statistics produce candidate relationship-type proposals.	`brain-mcp-server/src/engine/components/janitor/emergence.ts:4-25`; `brain-mcp-server/src/engine/components/janitor/emergence.ts:56-72`; `brain-mcp-server/src/engine/components/janitor/__tests__/emergence.test.ts:100-132`	not an eighth LLM; pass is default off and runtime vocabulary mutation is deferred.
K09	deterministic clusteringI/T	TypeScript Louvain uses fixed ordering and tie-breaks to form graph communities.	`brain-mcp-server/src/engine/components/edges/community.ts:4-29`; `brain-mcp-server/src/engine/components/edges/community.ts:37-52`; `brain-mcp-server/src/engine/components/edges/__tests__/community.test.ts`	no randomness in the current implementation; reserved seed is ignored. not Leiden.
K10	hybrid memory retrievalI/T	FTS5 BM25 and vector rankings combine through weighted reciprocal-rank fusion; both arms and hydration apply approved-review gates.	`brain-mcp-server/src/tools/memory-read.ts:479-664`; `brain-mcp-server/src/utils/hybrid-search.ts:31-102`; `brain-mcp-server/src/utils/__tests__/hybrid-search.test.ts:18-116`	reports hybrid/bm25_only/vector_only/none. vectors are selected before project filtering, which can underfill scoped recall.
K11	local embeddings/fallbackI/T	all-MiniLM-L6-v2 produces 384-dimensional vectors; optional sqlite-vec supports local vector storage.	`brain-mcp-server/src/utils/embeddings.ts:1-29`; `brain-mcp-server/src/utils/embeddings.ts:93-156`; `brain-mcp-server/src/engine/storage/__tests__/sqlite-vec.test.ts`	missing model/extension or offline cold-cache failure can fall back to lexical retrieval; not always hybrid or offline-ready.
K12	error memoryI	volatile error details are normalized into lookup fingerprints with stored solutions; mender uses these during diagnosis.	`brain-mcp-server/src/tools/errors.ts:30-79`; `brain-mcp-server/src/tools/errors.ts:101-171`; `core/agents/mender.md`	fingerprint is neither cryptographic nor causal proof. a retrieved fix still needs re-test.
K13	typed provenance/traversalI/T	briefs, learnings, errors, sessions, goals, concepts and decisions connect through typed dependency, lineage, recurrence and cluster edges; neighbors/path/subgraph queries traverse them.	`brain-mcp-server/src/engine/components/edges/handlers.ts:39-81`; `brain-mcp-server/src/engine/components/edges/traversal.ts`; `brain-mcp-server/src/engine/components/edges/__tests__/traversal.test.ts`; `brain-mcp-server/src/engine/__tests__/graph-traversal.integration.test.ts`	not ontology reasoning or guaranteed GraphRAG answer synthesis.
K14	whole-brain identityI/T	composite type/project/id keys avoid brief-id collisions; project drill uses depth-one edge closure with boundary/phantom nodes and ambiguity reporting.	`brain-mcp-server/src/engine/components/edges/whole-graph.ts:4-36`; `brain-mcp-server/src/engine/components/edges/whole-graph.ts:60-93`; `brain-mcp-server/src/engine/components/edges/whole-graph.ts:116-145`; `brain-mcp-server/src/engine/components/edges/__tests__/whole-graph.test.ts`; `brain-mcp-server/src/engine/components/edges/__tests__/project-qualification.test.ts`	completeness needs scope and boundary interpretation; hard caps are safeguards, not benchmarks.
K15	schedule claimsI/T	in-process smart sleep, BEGIN IMMEDIATE slot claim, overlap guard and retries manage due dispatch.	`brain-mcp-server/src/engine/components/schedules/daemon.ts:100-171`; `brain-mcp-server/src/engine/components/schedules/__tests__/daemon-fireloop.test.ts`	requires a running host; a crashed running row can wedge a schedule. not always-on distributed execution.
K16	cron/tool/shell dispatchI	schedule handlers dispatch MCP, noop and trusted arbitrary shell work.	`brain-mcp-server/src/engine/components/schedules/handlers.ts:65-127`; `brain-mcp-server/src/engine/components/schedules/cron.ts:162-230`	shell is not sandboxed; cron day-field semantics deviate as documented. MCP error envelopes can coexist with dispatch success, so failed tools are not universally retried.

Dashboard surfaces9 features

UI status here means inspected implementation. test-source presence is called out where the supplied audit identifies a witness; no route or accessibility certification was executed by WRITER.

Dashboard surfaces — implementation and source receipts
Key	Feature / status	Mechanism	Source receipt	Qualification
U01	dashboard shellI	React/Vite application implements Overview, Graph, Layers, Triage, Search and Diagnostics; pending-route map is empty.	`cli/dashboard/src/router.tsx:71-100`; `cli/package.json:54-67`	an existing compiled capture is authentic product evidence, not proof of a rebuilt dirty source tree or independent health verification.
U02	graph canvasI	force-graph canvas has custom paint/hit regions, type-coded graph elements, label placement, density tiers, ordered degradation and selected-record inspector.	`cli/dashboard/src/pages/Graph.tsx:304-391`; `cli/dashboard/src/graph/instance.ts:309`; `cli/dashboard/src/graph/tier.ts:25-78`; `cli/dashboard/src/graph/labels.ts:149`; `cli/dashboard/src/graph/shapes.ts:406-514`	scope, boundary, truncation and freshness affect interpretation. the zoomed image has nothing selected and does not expose record detail or its query footer.
U03	graph freshnessI	scope changes fetch a graph; explicit refresh updates it; other pages use polling.	`docs/dashboard.md:1414-1446`; `cli/dashboard/src/pages/Graph.tsx:1-20`	animated layout or shell live state does not imply continuous graph-data refresh.
U04	record layersI	briefs, goals, context and memory layers offer boards and record inspection with project-qualified identity.	`cli/dashboard/src/pages/Layers.tsx`; `cli/dashboard/src/pages/layers/Briefs.tsx:159`; `cli/dashboard/src/pages/layers/Briefs.tsx:416-563`; `cli/dashboard/src/pages/layers/Goals.tsx:321`; `cli/dashboard/src/pages/layers/ContextDocs.tsx:210`	private project/client content needs review; internal brief screenshot remains excluded pending consent.
U05	triage writesI	a narrow POST door maps supported dashboard actions through the brain write bridge/gateway; triage exposes pending knowledge.	`cli/src/lib/dashboard/server.ts:18-50`; `cli/src/lib/dashboard/server.ts:87-130`; `cli/src/lib/brain-write-bridge.ts:355`; `cli/dashboard/src/pages/Triage.tsx:141-147`	read handles coexist with a controlled write path; the dashboard is not entirely read-only. no triage action was taken for these images.
U06	layer searchI/T	brief/learning retrieval joins substring results for goals, suggestions and context documents; disjoint uniformly weighted lists use deterministic interleave.	`cli/src/lib/dashboard/search-fuse.ts:14-37`; `cli/src/lib/dashboard/search-fuse.ts:57-85`; `cli/src/__tests__/dashboard-search-fused.test.ts:182-247`	not every layer is semantic retrieval; an empty search screenshot does not prove returned results.
U07	cognition diagnosticsI/T	registered-instance health, terminal-event evidence and yield populations are read into the diagnostics route.	`cli/dashboard/src/pages/Diagnostics.tsx`; `cli/src/lib/dashboard/cognition-read.ts`; `cli/src/__tests__/dashboard-cognition-endpoint.test.ts:148-172`	actual NO SIGNAL is meaningful evidence, not a field to replace with fictional health. current health hardening includes WIP noted in K04.
U08	local protection/accessibilityI	loopback bind, Host/Origin fences, JSON-only writes/body cap, reduced-motion handling and semantic project-scope controls.	`cli/src/lib/dashboard/server.ts:4-45`; `cli/dashboard/src/lib/usePalette.ts:75-92`; `cli/dashboard/src/components/chrome/ProjectScope.tsx:30-53`	not a multi-user authenticated service, universal sandbox or complete accessibility certification.
U09	standalone HTML exportI	a separate brief visualization tool emits an older standalone HTML view.	`docs/visualization.md:172-190`; `brain-mcp-server/src/engine/components/edges/visualization-tool.ts`	CDN dependencies and no live updates belong to this export surface, not the newer bundled whole-brain canvas.

Frontier and deferred work7 features

Frontier and deferred work — implementation and source receipts
Key	Feature / status	Mechanism	Source receipt	Qualification
F01	agent teamsexperimental	`/team hunt` is an experimental parallel-agent workflow.	`README.md:249`	not promoted here to a verified general-purpose delivery guarantee.
F02	phase resumehardening	deterministic cross-harness phase-entry/resume behavior is still being tightened.	`README.md:250`	the documented same-machine task handoff does not prove every phase resumes deterministically.
F03	cross-machine evidencefrontier	full cross-machine handoff evidence remains to be earned.	`README.md:251`	implemented asynchronous sync and local claims do not establish end-to-end distributed ownership correctness.
F04	dynamic edge registrydeferred	statistical proposals can describe edge types; runtime vocabulary mutation is not applied by this path.	`brain-mcp-server/src/engine/components/janitor/emergence.ts:20-25`	no autonomous ontology evolution claim.
F05	optional cluster/emergenceI/T, default off	configured sub-passes provide community synthesis and statistical emergence proposals.	`docs/COGNITION.md:45-74`; `brain-mcp-server/src/engine/components/cognition/extractors/cartographer.ts:7-30`; `brain-mcp-server/src/engine/components/janitor/emergence.ts:4-25`; `brain-mcp-server/src/engine/components/janitor/__tests__/emergence.test.ts:100-132`	implemented is not enabled; cartographer is cadence-limited and synthesis remains distinct from statistical emergence.
F06	stochastic clusteringreserved, not implemented here	a seed is reserved for a potential future stochastic community variant.	`brain-mcp-server/src/engine/components/edges/community.ts:22-24`; `brain-mcp-server/src/engine/components/edges/community.ts:47-48`	current deterministic Louvain ignores the seed; no randomized behavior is claimed.
F07	HTML inline/liteplanned	older standalone renderer documents inline/lite enhancements as plans.	`docs/visualization.md:167-180`	these are not limitations of the current bundled dashboard and are not presented as implemented.

Reference notesSource context

N01: documented two-harness same-machine storyboard, README.md:47-55 and docs/images/launch/fr175-cross-harness-storyboard.gif. The audit did not reproduce the demonstration; no cross-machine outcome is inferred.
N02: enablement behavior and shared flags, docs/COGNITION.md:6-32 and docs/COGNITION.md:45-74. Fresh-install disabled configuration and missing-perception default-on behavior are different cases.
N03: supplied implementation audit, snapshot at dirty HEAD a060f677cd017cf7c776c2b2393b8c51a754ea9a. cli/package.json:1-4 and brain-mcp-server/package.json:1-4 contain source version 7.2.1, not registry or installed-release proof. Audit test execution was blocked by absent local runners; cited tests were inspected, not freshly passed. Screenshot provenance records existing cli/dist/dashboard capture without source rebuild; HTTP readiness was reported, not independent exact-candidate /api/health verification by WRITER.
N04: architecture backbone, brain-mcp-server/src/engine/index.ts:67-141, brain-mcp-server/src/index.ts:1-27, brain-mcp-server/src/engine/gateway.ts:126-196, cli/src/lib/harness-descriptor.ts:1-30, cli/src/lib/dashboard/server.ts:1-50. The detailed feature matrix supplies each subsystem's mechanism and boundary.

BUILDING AGENT SYSTEMS?

Inspect how the work survives the session. Read the handoff note, explore the evidence, or compare notes on engineering systems.

Build with me ↗explore the projects →