okto-pulse-core 0.1.14

Okto Pulse Core — spec-driven project management engine with MCP support for AI agents

okto-pulse-core

Core engine for Okto Pulse — shared models, services, API routes, and MCP server.

You probably want to install okto-pulse instead. This package is the internal engine. The okto-pulse package provides the CLI, frontend, and everything you need to get started.

What's inside

• 26 SQLAlchemy models — Boards, Cards, Specs, Ideations, Refinements, Sprints, Agents, Knowledge, Mockups, Validations, etc. (Skills entity dropped in 0.1.14.)
• 17 service classes — Full business logic with governance rules (Skills service dropped in 0.1.14.)
• 11 API route modules — FastAPI REST endpoints
• 150+ MCP tools — Complete Model Context Protocol server for AI agent integration, including:
  • Pipeline CRUD (Ideation, Refinement, Spec, Sprint, Card)
  • Q&A and choice questions across every entity
  • Mockups (HTML+Tailwind, sanitised) and Knowledge Bases at spec/refinement/card scope
  • Decisions with supersedence and coverage gates
  • Per-card Knowledge attachment lifecycle (add_card_knowledge and friends)
  • 22 Knowledge Graph tools (consolidation, query primario/power, health, dead-letter, schema-migrate, decay tick controllability)
• App factory — create_app() with dependency injection for auth and storage providers
• Embedded Knowledge Graph — per-board Kùzu instance + global discovery meta-graph, deterministic + cognitive workers, 11 node types and 10 relationship types

Architecture

┌─────────────────────────────────────────────────────────┐
│  Single Python process  (single Kùzu lock holder)        │
│                                                          │
│   ┌─────────────────┐         ┌──────────────────┐       │
│   │  uvicorn :api   │         │  uvicorn :mcp    │       │
│   │   FastAPI app   │         │  MCP ASGI app    │       │
│   │   /api/v1/*     │         │  /mcp            │       │
│   │   /static, SPA  │         │  streamable-http │       │
│   └────────┬────────┘         └────────┬─────────┘       │
│            │                           │                  │
│            └─────────┬─────────────────┘                  │
│                      │                                    │
│           shared module-level state:                      │
│           - Mongo session factory (init via lifespan)     │
│           - _global_db (Kùzu cache)                       │
│           - _active_api_key (ContextVar)                  │
└─────────────────────────────────────────────────────────┘

build_mcp_asgi_app() and mount_mcp(app) are the two helpers exposed from okto_pulse.core.mcp — pick build_mcp_asgi_app() to drive a separate uvicorn Server (the community edition does this for the --mcp-port listener) or mount_mcp(app) to mount the MCP sub-app under an arbitrary path on an existing FastAPI app.

Docker

This repo has no Dockerfile. The deployable artifact is a single image built from the sibling okto-pulse repo:

• okto-pulse/Dockerfile target local-runtime builds wheels from this repo and okto-pulse/ as siblings (used by okto-pulse/docker-compose.yml and the release pipeline's smoke build).
• okto-pulse/Dockerfile target pypi-runtime installs okto-pulse==<version> from PyPI, which transitively pulls this package off PyPI per the floor in okto-pulse/pyproject.toml (used by okto-pulse/docker-compose.prod.yml).

To run the published image:

docker run -d --name okto-pulse \
  -e HOST=0.0.0.0 -e MCP_HOST=0.0.0.0 \
  -p 8100:8100 -p 8101:8101 \
  -v okto-pulse-data:/data \
  ghcr.io/oktolabsai/okto-pulse:latest

See okto-pulse/README.md for the full Docker quickstart, and okto-pulse/CLAUDE.md for the multi-stage build architecture.

Release Notes

0.1.14 — current

Fix C: single-process, dual-port serve (Kùzu lock contention)

okto-pulse serve now runs API/UI and MCP from a single Python process but on two different ports (--api-port defaults to 8100, --mcp-port defaults to 8101). Two uvicorn.Server instances run concurrently inside one asyncio.gather — the embedded Kùzu DB is owned by exactly one OS process (no inter-process lock contention), and the two listeners share the module-level state (the registered session factory, the _global_db Kùzu cache, the _active_api_key ContextVar).

What you get:

• No Kùzu file-lock thrash — the embedded DB does not support multiple writers, so a single Python process is the only safe topology. The kg.db_open.lock_retry path=... attempt=N/5 warnings disappear.
• Independent ports — keep :8100 for the SPA fetches and :8101 for the MCP HTTP transport, unchanged from earlier releases.
• One lifespan — init_db, KG worker startup, scheduler boot, and register_session_factory all run once on the API listener; the MCP sub-app picks up the registered factory automatically.

Public surface:

• okto_pulse.core.mcp.build_mcp_asgi_app() — returns the MCP ASGI app wrapped in the ApiKeySessionMiddleware (handles ?api_key= / X-API-Key / Authorization: Bearer and binds the key to the request ContextVar).
• okto_pulse.core.mcp.mount_mcp(app, mount_path="/mcp") — mounts the same ASGI app onto an existing FastAPI app at the given path (community does this when an embedded mount is needed).
• okto_pulse.core.mcp.register_session_factory(factory) — call from the API lifespan so the MCP sub-app finds the DB. Idempotent.

Spec Skills entity removed in its entirety

The experimental "skills" feature on the spec entity is gone. Adoption was zero in real boards and knowledge entries already cover the reusable-context use case more naturally — the dedicated tab, MCP tools, REST endpoints and ORM table were paying recurring maintenance cost without return.

What goes away:

• 5 MCP tools removed — okto_pulse_create_spec_skill, okto_pulse_delete_spec_skill, okto_pulse_spec_skill_retrieve, okto_pulse_spec_skill_inspect, okto_pulse_spec_skill_load.
• 4 REST endpoints removed — GET / POST / PATCH / DELETE /api/v1/specs/{spec_id}/skills (and the {skill_id} variants).
• 5 permission flags removed — spec.skills.{read,load,create,delete,recall} from the registry and from every preset.
• Database table dropped — spec_skills. Migration is idempotent (DROP TABLE IF EXISTS); no downgrade — the data is gone.
• Pydantic schemas removed — SkillSectionSchema, SpecSkillCreate, SpecSkillUpdate, SpecSkillResponse, SpecSkillSummary. The skills field is gone from SpecResponse.
• agent_instructions.md scrubbed — Quick Navigation, the dedicated Spec Skills section, the spec-authoring workflow step and the destructive-operations row no longer reference skills.

Reader-side defensive handling: BaseSchema now sets extra="ignore" so historical payloads still carrying a skills field validate silently — no warning, no log, no error. There is nothing to migrate; the field is dropped on read.

Use knowledge entries (spec_knowledge, card_knowledge) and decisions for the same use case.

Agent instructions overhaul

agent_instructions.md was reviewed end-to-end. Three behavioural sections were added in response to repeated drift patterns observed across production sessions:

• § 2.1a Ambiguity-killer protocol — at ideation, the agent must scan the user's request against a table of ambiguity symptoms (vague verbs, undefined nouns, multiple plausible interpretations, implicit success criteria, implicit scope) and post Q&A items for every gap before advancing the ideation. "Just make a reasonable choice" is permission, not silence — it must be recorded explicitly.
• § 2.2a Investigação profunda obrigatória (refinement) — refinement is research, not paraphrasing. The agent must exhaust all applicable sources (project files, source code, KE, Knowledge Graph, mockups, web docs, online discussions, runtime evidence, stakeholder context) and the refinement body must cite each finding with path:line, KE titles, KG node ids or URLs.
• § 2.8 Card-level artifact attachment (MANDATORY) — every card must be self-contained. KE/mockup dependencies must be attached directly to the card via copy_knowledge_to_card / copy_mockups_to_card / add_card_knowledge / add_screen_mockup(entity_type="card"). Vague references to "see the spec" are a protocol violation.

Cleanup:

• Quick Navigation header Multi-value Parameters — Two Accepted Formats corrected to Three Input Shapes (the section was extended to native list[str] in 0.1.4).
• Obsolete delete_task_validation reference removed from the Available Tools → Evaluations & Validations table (the tool never shipped).
• okto_pulse_create_sprint parameter list aligned with the schema (objective? and expected_outcome? were missing).
• Duplicate "Startup Protocol" subsection deleted — Pre-Flight Checklist is the single source of truth.

Other improvements

• MCP ApiKeySessionMiddleware rewritten on top of ContextVar — required because the FastAPI process serves multiple concurrent requests and the previous module-level global would leak identities across requests. Token-based set/reset pattern protects against exception leaks.
• run_mcp_server retained for legacy callers, now defers to build_mcp_asgi_app for ASGI-mode embeds.

To upgrade an existing install: pip install -U okto-pulse okto-pulse-core and then okto-pulse init --agents to regenerate .mcp.json (the URL still points at port 8101 by default; override with --mcp-port if you remapped). No downstream contract changes for MCP clients — the wire protocol and tool catalog (sans 5 skills tools) are unchanged.

0.1.3 — previous stable (PyPI)

First hardening pass on the card lifecycle, the analytics contract, and the MCP instruction set.

• CardService.delete_card cascades through every spec-side JSON list (test_scenarios[].linked_task_ids, business_rules[], api_contracts[], technical_requirements[], decisions[]) and through bug cards' linked_test_task_ids. The transactional cascade unblocks the delete→recreate flow that previously tripped _validate_spec_linked_refs.
• Analytics card-type classifier uses enum identity instead of str(card.card_type).endswith(...). total_cards_impl/test/bug, task_validation_gate.total_submitted, velocity[].test/bug, and bug_rate_per_spec now report real counts.
• parse_multi_value helper consolidated the scattered .split("|") pattern; pipe-separated and JSON-array inputs are autodetected.
• MCP agent instructions rewritten (1830 → 2050 lines) with new sections for Multi-value Parameters, Destructive Operations, Versioning & Concurrent Edits, Security, Analytics-Driven Closure.

0.1.1 — initial PyPI release

26+1 SQLAlchemy models, 17+1 service classes, 11 API route modules, 119 MCP tools, embedded Kùzu Knowledge Graph with deterministic workers. (Spec Skills shipped here and was removed in 0.1.14.)

(Version 0.1.2 was published to TestPyPI only as a release candidate for 0.1.3.)

License

Elastic License 2.0 — free for personal and commercial use. Cannot be offered as a hosted/managed service.