Roadmap
SkillWeave generalises SigMap's proven primitives into an open standard, shipped version by version. Stats below reflect the current build.
Stats: 137 tests passing · 3 frozen base skills · 7 domain skills · 3 SigMap adapters · 4 provider adapters · 8 registry schemas · 4 capabilities · 3 trust tiers · 7 trigger types · 4 event types · 6 composition patterns · 19 CLI commands · 4 runtime deps
Shipped
v0.1.0 — Prototype chain ✓
The 4-skill chain proving the mechanics end to end: parse-input → validate-coverage → boundary-judge → memory-update. Core primitives (SKILL, PIPELINE, STATE, ASSERTION), the three frozen base skills, a multi-LLM boundary judge with an offline heuristic fallback, NDJSON tracing, and STATE checkpoints.
Tags: prototype · ask→validate→judge→learn · multi-LLM judge Impact: runs end to end offline; failure caught and surfaced with full diagnostics.
v0.2.0 — Reliability layer ✓
Systematic non-determinism handling at the probabilistic boundary: confidence routing (≥0.85 / 0.65–0.85 / <0.65), an auto-inserted boundary judge, and retry-with-negative-context (budget 2). Golden anchors in the skill contract; the extract-highlights probabilistic skill; deterministic skills stay overhead-free.
Tags: confidence routing · auto-judge · retry-with-negative-context · golden anchors Impact: failing probabilistic skills recover instead of halting.
v0.3.0 — Production runtime ✓
A real runtime driven by a skillweave CLI: a pipeline YAML loader that resolves each step against a skill registry (with per-step confidence_threshold / retries overrides), and the commands run · validate · test · list · trace · new, plus an installable bin.
Tags: skillweave CLI · pipeline YAML loader · skill registry · bin Impact: pipelines are declared in YAML and run/validated from the CLI. (Full JSON-schema registry with version resolution deferred to v0.6.0.)
v0.4.0 — SigMap adapters ✓
The SigMap adapter layer: ContextProvider / CostManager / ObservabilityProvider interfaces with implementations that wrap SigMap's local artifacts (the .context/ directory and the SigMap usage.ndjson-compatible metric stream) — no shell spawn, no rebuild. OBSERVE computes a composite 0–100 health score on SigMap's grade scale, surfaced via skillweave health and a footer on every run.
Tags: CONTEXT · COST · OBSERVE · skillweave health · sigmap command Impact: health grading from the shared NDJSON contract. (TRIGGER / EVENT / MEMORY adapters and install @sigmap/adapters deferred to later milestones.)
v0.5.0 — Multi-LLM provider layer ✓
A formal LLMProviderAdapter interface with anthropic / google / openai / ollama adapters, capability profiles in provider-profiles/*.profile.yaml, an executor that auto-selects a model by tier + required capabilities and runs primary → fallback, and a Neutral Skill Language validator. The boundary judge now routes its LLM call through the layer.
Tags: LLMProviderAdapter · provider profiles · primary→fallback executor · neutral-language validator Impact: the same pipeline runs on any configured provider, with capability-driven model selection and automatic fallback.
v0.6.0 — Schema governance ✓
A versioned schema registry (schemas/registry/<name>@<version>.json), input_schema / output_schema pins on skills, a schema differ (additive vs breaking), an additive-only rule across same-major versions, and skillweave check-schemas to catch breaking changes before they ship.
Tags: versioned registry · schema pins · diffSchemas · additive-only · check-schemas Impact: 43-test node:test suite; schema changes are governed — a removed/retyped/newly-required field within a major fails the gate.
v0.7.0 — Security model ✓
A per-skill security model: a capability vocabulary (fs:read · fs:write · net · env:read), a default-deny SecurityPolicy, and a filesystem sandbox (guardWrite) that contains writes within the policy's roots and blocks ../ traversal. Skills declare a capabilities field; the orchestrator runs a pre-flight permission check and halts an over-privileged skill — with secret-redacted diagnostics — before it executes. skillweave check-permissions audits the whole registry against the policy.
Tags: capability permissions · default-deny · guardWrite sandbox · redactSecrets · check-permissions Impact: 56-test node:test suite; a skill can do only what its declared capabilities allow, enforced before execution. Pure skills carry zero overhead.
v0.8.0 — SigMap pipeline integration ✓
SkillWeave becomes SigMap's internal execution architecture. A load-context skill sources raw_input from SigMap's CONTEXT artifact, the sigmap-verify pipeline runs the ask→validate→judge→learn verify flow, and runSigMapVerify() (exported from src/index.ts) returns a structured VerifyResult so SigMap can embed the runtime in-process — no shell spawn, no CLI. Also skillweave verify.
Tags: load-context · sigmap-verify pipeline · runSigMapVerify · VerifyResult · public API barrel · check verify Impact: 63-test node:test suite; SigMap consumes SkillWeave as a library, verifying its own context through the same reliability layer.
v1.0.0 — Registry + public launch ✓
The first stable release. A tiered, quality-gated skill registry: the 9-point gradeSkill quality gate assigns a trust tier (verified / community / experimental), each entry carries a quality-derived reputation, and a local-first store (.registry/skills.json) backs publish / install / registry. The src/index.ts API is now the supported integration surface.
Tags: skill catalog · 9-point quality gate · trust tiers · reputation · publish/install/registry · stable public API Impact: 71-test node:test suite; skills can be graded, tiered, published, and installed locally. (Runtime reputation is seeded from the quality score; trace-history reputation is a follow-up.)
v1.1.0 — Developer experience ✓
First-run onboarding so a Level 1 user never has to feel the runtime's depth. skillweave doctor gives a one-command readiness report (Node · active judge provider or the offline heuristic · registered skills · registry/artifacts), and the CLI now suggests the closest match when a command or skill name is mistyped (verfy → verify).
Tags: skillweave doctor · runDoctor · did-you-mean · closest/levenshtein Impact: 79-test node:test suite; a newcomer confirms "ready to run, offline" in one command, and typos guide instead of dead-end.
v1.2.0 — TRIGGER + EVENT primitives ✓
The two activation/observability primitives, local-first. TRIGGER declares how a pipeline activates (manual / cron / webhook / pipeline_completion + condition + human_checkpoint), with a pure 5-field cronMatches and a shouldActivate resolver. EVENT is a typed signal model — an EventBus with declarative { on, emit, notify, continue } subscriptions that fan out across trace-log / webhook / human routes; the orchestrator emits low_confidence_detected / skill_failed / pipeline_succeeded. The loader parses + validates both pipeline blocks.
Tags: TriggerSpec · cronMatches · shouldActivate · EventBus · trace-log/webhook/human routes · continue=halt Impact: 97-test node:test suite; pipelines declare activation and route typed signals — without any network or shell (delivery is the host's job).
v1.3.0 — MEMORY primitive ✓
Pipelines learn from past executions. A local-first MemoryStore on .context/ records outcomes and failures and reads them back across sessions, with a decay model (stale records stop informing adaptation), last-write-wins keyed records plus a conflict log, and failure-pattern learning (failurePatterns / recommend). Skills declare memory_reads / memory_writes (scoped writes, default-deny), and skillweave memory surfaces the trend + recommendations.
Tags: MemoryStore · decay/isStale · conflict log · failurePatterns · recommend · per-skill scope · skillweave memory Impact: 108-test node:test suite; the run log becomes adaptive knowledge — failure patterns and recommendations from real history, all local and offline.
v2.0.0 — COMPOSE + OBSERVE ✓
The capstone. COMPOSE completes all composition patterns as pure async combinators — sequential / parallel / mapPattern / reducePattern / conditional / loop — plus dagLayers for DAG resolution into parallelizable layers. OBSERVE adds the local-first observability layer: checkAlerts threshold rules (routable via the EventBus), visualise (ASCII/Mermaid pipeline diagrams + skillweave visualise), and abTest skill-version comparison. This release also carries the v1.3.0 MEMORY primitive to a published tag.
Tags: sequential/parallel/map/reduce/conditional/loop · dagLayers · checkAlerts · visualise · abTest · skillweave visualise Impact: 120-test node:test suite; the v0.1 → v2.0 roadmap is complete — every primitive (CONTEXT · COST · OBSERVE · TRIGGER · COMPOSE · MEMORY · EVENT · SECURITY) shipped, all local-first.
v2.0.1 — npm-publishable ✓
Packaging release. The runtime now compiles to dist/ (JS + .d.ts), drops private, and exposes a stable main / exports so it can be installed from npm and imported as a library. Bundled data (schemas, provider profiles) resolves relative to the install, so a global CLI works from any directory.
Tags: dist build · main/exports/types/files · package-relative data · dist-first bin Impact: 123-test node:test suite; import { runPipeline, EventBus, MemoryStore, … } from "skillweave" once built.
v2.0.1.x → v2.0.2 — published to npm ✓
The package is live on npm (npm i skillweave), published by a tag-gated workflow with provenance. This patch also hardened the CLI: run --doc / test --input / verify --input fail gracefully on a missing or malformed file (clean error + exit 2) instead of an uncaught stack trace — found by running every feature end-to-end before publishing.
Tags: npm publish workflow · provenance · graceful CLI file errors · readArgFile Impact: 126-test node:test suite; npm i skillweave works.
v2.1.0 — todo-flagger skill + examples ✓
A sixth domain skill, todo-flagger, that flags content blocks containing a TODO / FIXME / XXX marker — a complete worked example of authoring a skill: it extends State with a new flags field, adds a todo-flag@1.0 registry schema, and grades 9/9 → verified. Plus eight runnable, offline examples (one per feature area) surfaced from the docs and README.
Tags: todo-flagger · new STATE field · todo-flag@1.0 schema · examples/ · Examples guide Impact: 130-test node:test suite; 6 domain skills · 7 registry schemas; a copy-paste template for new skills.
v2.2.0 — summarize skill ✓
A seventh domain skill, summarize — a probabilistic extractive summarizer that ranks content blocks by salience and takes the top ones verbatim, so the auto-judge sees grounded output. On a retry it takes fewer, stronger sentences, so a low-confidence first attempt recovers — exercising the full reliability layer (confidence routing → auto-judge → retry-with-negative-context) from a second skill (only extract-highlights did).
Tags: summarize · probabilistic · DocSummary · doc-summary@1.0 · confidence routing · retry-recovers Impact: 134-test node:test suite; 7 domain skills (2 probabilistic) · 8 registry schemas.
v2.2.1 — Gemini array fix + use-case examples ✓
Fixes structured array output on the Google provider: toGemini() now maps array JSON schemas to Type.ARRAY (they previously fell through to the string default, so Gemini returned a comma-joined string — breaking extraction and the boundary judge on Gemini). Adds examples/use-cases/ — six runnable real-world scenarios (support triage, code-review gate, secret-safe logs, batch digest, nightly report, and a live-LLM A/B) that show what the runtime adds over a plain prompt.
Tags: Gemini Type.ARRAY · structured array output · examples/use-cases · live-LLM A/B Impact: 137-test node:test suite; structured array output works on all three providers.
Planned
The published roadmap (v0.1 prototype → v2.0 COMPOSE + OBSERVE) is complete. Deferred, host-facing extensions remain: a hosted observability dashboard and live alert/webhook delivery (network), orchestrator-level DAG auto-execution, and published performance benchmarks.
Current milestone: roadmap complete — v2.0.0 shipped. Next work is host-facing (dashboard, live delivery) and an optional public npm publish.