flowbrain_core 0.1.8

FlowBrain Core — Judgment & Knowledge core for the MakeMind ecosystem. Two cooperating subsystems: Knowledge (5 facades over a 4-layer Fact/Skill/Profile/Philosophy structure plus Ops) and Agent (work [...]

0.1.8 - 2026-07-04 - Tolerant routing / review JSON parsing (additive fix)

Fixed

• RoutingDecision.tryParse / ReviewResult.tryParse now recover the JSON object from decorated LLM output. Both did a single strict jsonDecode(raw) over the entire model response and silently returned parse_error on any failure. Short prompts got bare JSON and passed; long requests (~1.4KB) made the model wrap its answer in a Markdown code fence (```json), add a prose preamble, or append trailing reasoning — so jsonDecode threw and agent_route silently broke (targetAgentId: '', confidence 0.0, delegation lost). Parsing is now tolerant via a shared extractJsonMap: strip a code fence if present → try a direct decode → fall back to the outermost { … } substring; genuinely object-free output still returns parse_error. Both parsers share the helper (the reviewer path had the identical vulnerability). Reported from live operation (konpi field defect D5); root cause confirmed at agent_models.dart tryParse, byte-identical across 0.1.4→0.1.7. Regression: test/agent/17_unit_coverage_test.dart (fenced / bare-fence / prose-preamble+trailer / no-object cases for routing, fenced+prose for review). 155 PASS · analyze 0.

Backward compatibility

• Fully additive. No API or signature change — strictly more inputs now parse; every input that parsed before still parses identically. extractJsonMap is exported for reuse.

0.1.7 - 2026-07-02 - Agent role update seam (additive)

Added

• AgentRegistry.update / AgentFacade.updateAgent accept AgentRole? role. An agent's orchestration role (worker / manager / reviewer) can now be changed in place. Previously the only path was delete → recreate, which destroys the individual (owned axis forks, history, accumulated expertise) — contradicting the persistent-roster principle. The model was already designed for this (Agent.copyWith(role:)); only the update surfaces omitted the parameter. Safe by construction: every role consumer checks live at use time (AgentRuntime.route/review validate manager/reviewer on each dispatch), so no role-derived state exists to invalidate. Optional named parameter — existing callers unaffected. Regression: test/agent/12_agent_registry_test.dart T-AGT-REG-008 (in-place promote persists · untouched fields kept · omitting role leaves it unchanged). 153 PASS · analyze 0.

0.1.6 - 2026-06-24 - ConversationStore per-agent append atomicity (additive)

Fixed

• ConversationStore.append is now atomic per agentId. append was a non-atomic read-modify-write (load → [...existing, turn] → _kv.set); two concurrent agents.ask calls for the same agent could read the same history, each append its own turn, and the last set win — silently dropping a turn (the agent then "forgets" the lost request). A per-agentId serialization tail now chains mutating ops (append, and clear so append↔clear/TTL-sweep can't interleave either) so same-agent calls run strictly in order; different agentIds stay parallel. The tail always completes successfully (a failed op surfaces to its caller but never blocks the next) and its map entry is dropped when nothing is chained behind it (bounded growth). Regression: test/agent/14_conversation_store_test.dart (25 concurrent same-agent appends preserve all 25 · two agents stay isolated) — both fail without the serialization.

Backward compatibility

• Fully additive. No API or signature change — append / clear / remove keep their shapes; only concurrent-call ordering is now guaranteed.

0.1.5 - 2026-06-23 - FlowBrain runtime wiring (spec 12 §2·§3·§3b·§4·§4b)

Changed (behavior — additive, no API change)

• §2 — 4-axis ask composition: AgentRuntime.ask now composes all 4 assigned axes (profile · philosophy · skill · facts), not facts alone (spec platform/12-flowbrain-runtime.md §2). Order: profile (persona) → philosophy (values/prohibitions) → skill → facts. Facts keep rich _factLine; non-facts axes render defensively from their owned-fork payload (live object via toJson, or JSON Map) so no hidden Knowledge-Subsystem type is imported — 20 items/axis cap. None assigned → prompt identical (regression-safe).
• §3 enforcement proven end-to-end (real engine, not stub): test/agent/25_philosophy_prohibition_enforcement_test.dart wires a real PhilosophyEngine over a seeded ethos store and proves ask actually blocks an output hitting a hard prohibition and delivers a clean one — via Prohibition.forbiddenPatterns (mcp_bundle 0.4.4 + mcp_philosophy 0.1.2's deterministic _detectViolation). This closes the gap that TEST-24's always-block stub masked: the structural evaluator previously caught only two hardcoded NL shapes and silently fell open for every other prohibition. Runtime requirement: mcp_philosophy ≥ 0.1.2 for the deterministic path to actually block (resolved transitively via mcp_knowledge caret; semantic NL judgment remains an LLM seam — spec §3.1).
• §3 — Philosophy work-time intervention: AgentRuntime.ask now runs the assigned Philosophy's post-generation gate over the LLM output before delivery — a hard prohibition throws AgentPhilosophyBlockedException (turn not appended/delivered), soft modifications adjust the text. Opt-in: no assigned philosophy or no engine → skipped (the existing-suite agents are unaffected). Fails open on engine error. (InterventionPoint/PipelineContext/InterventionResult come from mcp_bundle — flowbrain's own direct dep; no new dependency.) detectTensions at the fork-evolution boundary is a separate follow-up.
• §4 — outcome→knowledge loop (philosophy): AgentRuntime.review now feeds the reviewer verdict back as a Philosophy FeedbackEvent → proposeFeedback (pass→positive · fail→negative · revise→mixed). The proposal is human-gated and never auto-applied (facade only emits EvolutionProposedEvent; gated by enableEvolution). Opt-in (target has assigned philosophy + engine available); fails open. AgentPhilosophyBlockedException added (agent_exception.dart). (Other axes already accumulate: facts write + profile growth counters; the FeedbackEvent wire was the missing outcome→proposal call-site. detectTensions at fork-evolution = follow-up.)
• §4b — skill refinement path: AgentRuntime.review now, on a deficient verdict (fail / revise), records a skill variation candidate per assigned skill fork via GrowthTracker.trackVariation (GrowthKind.variation → skillCandidateCount accumulator + AgentForkEvolvedEvent + FactGraph timeline). A pass verdict means the skill worked as-is → no candidate. This completes the §4 four-axis loop (philosophy reinforcement + skill refinement + facts/profile already accumulating). Reuses the existing growth mechanism — no new model. Opt-in (target has an assigned skill); fails open.
• §3b — Philosophy fork-evolution drift anchor: AgentRuntime.review now, at the fork-evolution boundary (a review verdict is the outcome that evolves the target's non-philosophy axes), calls the assigned Philosophy's detectTensions(MultiLayerContext) over the agent's evolving profile + facts-provenance, emitting AgentForkTensionDetectedEvent (count + max severity + descriptions) when tensions are found. Philosophy is the only axis that governs the other three — this surfaces drift so a host can flag/hold an evolution that would leave the constitution. Opt-in (target has assigned philosophy); advisory (emit, no auto-revert); no-op when the adapter doesn't implement detectTensions (UnsupportedError caught); fails open. (MultiLayerContext/PhilosophyEvaluationContext/Tension from mcp_bundle — direct dep, no new dependency.) AgentForkTensionDetectedEvent added (agent_event.dart).
• No public API change (additive). Tests: test/agent/24_assigned_axes_in_prompt_test.dart (profile compose · philosophy block · opt-in pass-through · fork tension emit · opt-in no-tension) + 22_* (facts) + skill-refinement candidate (§4b) + §3 prohibition enforcement (TEST-25, real engine). 150 PASS · analyze 0.

Changed (dependency floor)

• mcp_bundle ^0.4.3 → ^0.4.4 — internal-dep latest (0.4.4 hardens the Ethos object-graph fromJson with field-named validation). flowbrain_core consumes the ethos via the philosophy port; no new symbol required, constraint kept current.
• mcp_knowledge ^0.2.4 → ^0.2.5 — propagates the §3 prohibition-enforcement guarantee: mcp_knowledge 0.2.5 floors mcp_philosophy ^0.1.2 (deterministic forbiddenPatterns enforcement). Flooring it here ensures flowbrain's resolution actually carries the enforcing engine rather than relying on caret-resolves-to-latest — the §3 gate (TEST-25) is hollow without mcp_philosophy ≥ 0.1.2.

0.1.4 - 2026-06-14 - assigned facts compose into ask prompt

Changed (behavior — additive, no API change)

• AgentRuntime.ask now composes the agent's assigned facts (set via assignFacts / bk.agent.assign_facts, stored under AgentAxis.facts) into the system prompt: base systemPrompt first, then an ## Assigned knowledge (facts) section. Previously ask passed only the base systemPrompt, so assigned facts never reached the provider — a per-agent-knowledge-scoping gap (agent answered "I don't have that fact"). Facts are read from the agent's eager OwnedFork payload (handles live OwnedFork and persistent JSON Map forms), capped at 50 lines. No assigned facts → prompt identical to before (regression-safe). No public API change. Tests: test/agent/22_assigned_facts_in_prompt_test.dart (in-memory + persistent JSON-KV round-trip).

Changed (dependency floor)

• mcp_bundle ^0.4.0 → ^0.4.3 — guarantees FactRecord.toJson/fromJson. Without it, assigned facts persisted via a persistent KvStoragePort serialize to "Instance of 'FactRecord'" (toString) and the compose above yields nothing — the fix only works end-to-end with serializable FactRecord. (In-memory KV worked regardless.)

0.1.3

Added

• AgentFacade.ask / AgentRuntime.ask gain an optional resetContext flag (default false). When true, the agent's conversation history is cleared before the prompt is composed — for manager agents whose every turn should be treated fresh (bounds context growth, avoids stale prior-turn pollution that weakens the current directive). The post-ask turn is still appended; the reset is one-shot.

Changed (dependency floor)

• mcp_knowledge ^0.2.3 → ^0.2.4 — raises the floor so the re-exported OpsFacade is guaranteed to carry the behavior-execution methods (runBehavior / resumeBehavior / listBehaviors, added in mcp_knowledge 0.2.4). flowbrain_core's own code is otherwise unchanged; this guarantees the capability for consumers (e.g. brain_kernel) that reach behavior through system.ops. Consumers should bump to ^0.1.3.

0.1.2

Changed (cascade)

• mcp_bundle caret bumped from ^0.3.2 to ^0.4.0 (mcp_bundle 0.4.0 UiSection.pages spec realignment).
• mcp_knowledge caret bumped from ^0.2.2 to ^0.2.3 (sibling cascade).

flowbrain_core does not touch UiSection.pages directly — caret-only cascade. Consumers should bump to ^0.1.2.

0.1.1

Dependencies

• mcp_bundle: ^0.3.1 → ^0.3.2 — cascade alignment to pick up McpBundle.factGraphSection wire (additive — fact instance round-trip slot alongside the existing factGraphSchema type catalogue).
• mcp_knowledge: ^0.2.1 → ^0.2.2 — same cascade.

No code changes in flowbrain_core itself.

---

0.1.0

Initial release.

• Knowledge Subsystem — five facades (facts, skill, profile, philosophy, ops) wrapped from mcp_knowledge 0.2.x. Four-layer knowledge structure (L0 FactGraph → L1 Skill → L2 Profile → L3 Philosophy) plus Ops, with default L0 auto-wiring and InfraPorts.inMemory() smoke infrastructure.
• Agent Subsystem — flowbrain-native, fully on-package:
  • Self-contained agents — own LLM context, own model (ModelSpec), own forked 4-axis instances.
  • Three roles via AgentRole — worker / manager / reviewer.
  • 4-axis fork assignment from workspace pool or from another already-evolved agent (PoolForkSource / AgentForkSource) across skill · profile · philosophy · facts.
  • agents.ask / agents.stream for conversation, agents.route for manager dispatch, agents.review for reviewer evaluation.
  • Growth Tracker records evolution kinds (variation / adjustment / revision / accretion) per agent.
  • Conversation Store with getHistory(limit).
• Event bus — system.eventBus.stream (broadcast Stream<KnowledgeEvent>); domain-typed agent events (AgentCreatedEvent, AgentDeletedEvent, …).
• Stub LLM — StubLlmPort auto-wired by KnowledgeSystem.withAgents() when no provider is supplied; deterministic short replies for tests / smoke runs.
• Public surface — single barrel import package:flowbrain_core/flowbrain_core.dart exposes:
  • KnowledgeSystem (defaults / stub / withAgents factories)
  • KnowledgeConfig + 9 sub-configs
  • InfraPorts + standard infrastructure port interfaces
  • AgentFacade + Agent / AgentReply / AgentRole / AgentAxis / ModelSpec / RoutingDecision / ReviewResult / ReviewVerdict / GrowthKind / ConversationTurn
  • Philosophy domain helper exceptions and result types
• Examples — example/flowbrain_example.dart, example/multi_role.dart, example/lifecycle_events.dart. All run end-to-end against the stub LLM with no external services.