SkillWeave

Multi-agent AI skill orchestration — 15 skills, 7 lifecycle phases, any AI coding agent.

SkillWeave orchestrates AI-assisted product development from idea to post-release. It works with Claude Code, Codex, Gemini CLI, Cursor, Windsurf, OpenCode, and any agent that reads Markdown or speaks MCP.

What's new in 1.5.10

• Native execution surfaces: an identity-verified lifecycle store and
  durable execution stream, native Assay, and Scope, Skills, Verify, Cancel and
  Budget enforcement through the authoritative run path.
• Route-lock and MCP control/authority: the route-decision-v1 payload with
  egress confinement, a server-owned MCP authority that gates every tool, and a
  durable harness-neutral agent mailbox served over both transports.
• Trustworthy publishing: one honest publish identity carried to both
  forges, with the Core/SDK contracts qualified against the installed wheel.

See the 1.5.10 changelog
for the full release contract. 1.5.10 is the release in preparation — not
yet released until its v1.5.10 tag exists. Earlier release notes remain in
the changelog as history.

Getting Started

For Power Users

Use meta-commands directly:

/skillweave                    # Where am I? What's next?
/skillweave plan               # View Kanban board
/skillweave build              # Start/continue building
/skillweave test               # Run test pyramid
/skillweave council topic="…"  # Get multi-model opinion
/skillweave report             # Generate release report

Or go straight to individual skills:

/skillweave-blueprint idea="Your SaaS idea"
/skillweave-promptchain-generate inputs='{"prd": ".skillweave/prds/v1.0/prd.json"}'
/skillweave-releasechain inputs='{"prd": ".skillweave/prds/v1.0/prd.json"}' mode="attended"

For Indie Hackers

Start with /skillweave to see where your project stands. The navigator detects your phase automatically and recommends the next step.

New project:

1. /skillweave-blueprint idea="Your idea" — Creates PRD
2. /skillweave-promptchain-generate inputs='{"prd": ".skillweave/prds/v1.0/prd.json"}' — Generates execution plan
3. /skillweave build — Executes the plan with testing gates

Existing project:

1. /skillweave — See current phase
2. Follow the recommendation

For Non-Technical Users

Just type:

/skillweave start

The wizard asks 5 simple questions and starts the right tool for you.

Install

SkillWeave is distributed from source and via Capacium. There is no PyPI
distribution for this repository (see docs/release-flow.md):
installing SkillWeave from PyPI is not a supported install path, and no workflow
publishes it there.

Install paths, all exercised by the packaging suite
(tests/packaging/test_install_provisioning.py):

Via Capacium (recommended)

Works with Claude Desktop, Claude Code, Codex, Gemini CLI, Cursor, Windsurf, OpenCode, and more.

brew install capacium/tap/capacium
cap install skillweave

From a built wheel

python -m pip install ./skillweave-1.5.10-py3-none-any.whl

skillweave declares requires-python >=3.9; Python 3.9 is the supported
floor. See docs/provisioning.md for the tested-Python
matrix and the full offline/venv/uv/pipx flows.

From a source checkout (editable, for development)

git clone https://github.com/SkillWeave-AI/skillweave.git
cd skillweave
pip install -e .

Optional: MCP server

The Model Context Protocol server is an optional extra, never a base
dependency. The base install runs without it and reports an actionable hint if
the server is requested while absent:

python -m skillweave.scripts.check_optional_capability mcp
# -> missing; remedy: pip install 'skillweave[mcp]'

Add it explicitly when you want the real server:

pip install 'skillweave[mcp]'

Providers, credentials and the zero-key default

Search is served through six providers, one of which needs no credential:

• DuckDuckGo Lite is the zero-key default — no API key, no account.
• Serper / SerpApi / Tavily / Brave / Perplexity are keyed providers,
  selected by name. Each reads its key from the environment:
  SERPER_API_KEY, SERPAPI_API_KEY, TAVILY_API_KEY, BRAVE_API_KEY,
  PERPLEXITY_API_KEY. No key is ever required to install SkillWeave.

Honest zero-key failure behavior: if a keyed provider is selected and its
credential is absent, the call fails closed with an explicit
<ENV_VAR> not set / missing-credential refusal — it is never silently served
by DuckDuckGo. An unrecognised provider name likewise fails closed.

envctl is optional convenience (if present); SkillWeave performs no component
and does not require it.

Architecture

Layer 0 — Wizard          /skillweave start         (5 questions, routes to skill)
Layer 1 — Meta-Commands   /skillweave plan|build|…  (7 commands, maps to skills)
Layer 2 — Direct Skills   /skillweave-blueprint …   (15 skills, full control)

7-Phase Lifecycle

Discovery → Blueprint → Design → Build → Release → Launch → Post-Release

Global skills (available in any phase): lifecycle, repo-health, observe, council

15 Skills

The set below is the exact skills/*/capability.yaml inventory; the
machine check in scripts/check-version-truth.py fails if this list and the
on-disk inventory drift apart.

LLM Council

Convene multiple AI models for research, code review, or architecture decisions:

• 3 stages: Independent opinions → anonymous peer review → chairman synthesis
• Profiles: quick (1 model), default (2), deep (2), expert (2) — provider-native Faidata roster ids
• Web search: DuckDuckGo (free), Serper, Tavily, Brave, SerpApi, Perplexity
• Output: Markdown or structured JSON with consensus score
• Truthful attribution: provider-native model ids; each seat records
  requested/resolved/answering model, status, provider and profile revision. The
  answering model is read from the response envelope, never inferred from the
  request, and fewer than the minimum distinct answering models is a degraded
  run — never reported as consensus.

Adapters & extension registry

Add or swap a search provider, register a local tool, or read a package-format
manifest through one neutral extension point, with no vendor edit:

• Provider swap is a data change; the zero-key DuckDuckGo default needs no
  credential, and a keyed provider is selected by name.
• Local tools register directly (no package format); Capacium is an
  optional, versioned package-format reader whose kind stays opaque.
• Fail-closed admission: an unapproved adapter, a denied egress, a missing
  credential and an unknown provider are distinct, attributable refusals.

See docs/adapters.md for the consumer-facing guide
(provider swap, zero-key defaults, local tool and connector-pack examples,
package-author vs consumer responsibilities, and the conformance matrix). Every
claim there is backed by a machine check under tests/contract/.

Testing

5-level test pyramid with automatic gate decisions:

Level 5: Evidence/Groundedness  (are claims backed by code?)
Level 4: Acceptance             (do acceptance criteria pass?)
Level 3: E2E Smoke              (does the system start and respond?)
Level 2: Unit                   (do individual functions work?)
Level 1: Lint                   (is the code clean?)

Gate decisions: PROMOTE (all pass) | HOLD (minor failures, suggest fix) | ROLLBACK (critical failures, block)

Planning

File-based Kanban with YAML frontmatter tickets:

.skillweave/planning/
├── backlog/     # Tickets waiting
├── doing/       # In progress
├── done/        # Completed
└── BOARD.md     # Auto-generated board view

Create tickets, move between states, seed from PRD — all via /skillweave plan.

Configuration

All SkillWeave data lives in .skillweave/ (git-ignored):

.skillweave/
├── config.yaml          # Risk mode, tier, feature flags
├── planning/            # Kanban tickets (backlog/doing/done)
├── prds/                # Product requirement documents
├── sequences/           # Execution sequences
├── testing/             # Test config + results
├── handover/            # Skill-to-skill transition signals
├── reports/             # Release reports (yyyy-mm-dd-topic.md)
├── discovery/           # Discovery phase artifacts
├── schemas/             # JSON schemas for validation
└── templates/           # Report/doc templates

The 27 canonical .skillweave/ areas, their single owning skill/subsystem,
lifecycle phase and mutability are specified in
docs/substrate-map.md; the execution/support matrix
in docs/native-execution-support-1510.md
summarises the substrate count and the one open deviation it records.

Risk Modes

Override hierarchy: CLI parameter > env (SKILLWEAVE_RISK_MODE) > project config > global config > default.

Distribution

SkillWeave ships in two formats depending on agent capabilities:

Both formats use the same Python core (src/skillweave/). SKILL.md is the universal baseline; MCP provides richer tool-based interaction for agents that support it.

Host-Neutral Scope

Shipped skills are host-neutral by contract: they neither invoke a specific host
binary nor default task assignment to a concrete host or model. Task routing is
capability-based (target_agent: any); any retained host-specific invocation
lives under references/adapters/<host>.md as a non-binding adapter example.

Stable, host-neutral Command/Intent semantics and full cross-transport
(Command ⇄ MCP/Intent) parity were scoped as 1.4.0 scope. (Historical note:
that line was written against the 1.4.0 train; this README does not restate a
current-release status for it.)

Support and limits

This section states what the shipped engine does and the limits it carries;
it makes no future-capability or always-free promise.

The execution/support matrix — the versioned surface support table
(skillweave.support/1), the shared run-state interpretation
(skillweave.run-state/1), the CLI/MCP/Assay rows, and the explicit bare
limits, detected-versus-live-tested adapters, unknown-capacity rule, preview
features and plaintext onboarding storage behaviour — is
docs/native-execution-support-1510.md.
Every number and name there is derived from the release tree and asserted by
tests/unit/test_1510_doc_truth.py.

The 15 skills (all 15 skills) and 7 phases above are the ones this repository
ships — the exact skills/*/capability.yaml inventory. The local engine keeps
its raw events on the machine and works offline; the aggregate telemetry service
is labelled preview/unavailable and this release ships no production
service for it (AGGREGATE_SERVICE_GA is pinned False). Edition contracts
exist for Community / Pro / Enterprise; EditionGuard grants only what an
edition's declared contract allows and refuses an unknown edition. Expected
limits — a keyless zero-key search default, a source/Capacium-only install (no
PyPI), a capability handshake where unknown capacity is never treated as
unlimited, and onboarding state written as plaintext YAML — are all
itemised in the matrix above.

Version truth

Version labels on this page distinguish four facts, which are not one
number. The declared, machine-checked inventory is
docs/version-truth.md, verified by
scripts/check-version-truth.py:

The version badge above advertises the candidate 1.5.10, the release in
preparation — it is deliberately not presented as released until its
v1.5.10 tag exists. The released fact, whenever it is needed, is derived by
the peeled-tag rule from the repository rather than printed here. Historical
versions (for example the 1.4.0 note in Host-Neutral Scope) appear only as
explicitly historical statements. See
docs/version-truth.md for the full rule, the
peeled-tag policy and the provenance of each figure.

License

Apache 2.0 — Copyright 2026 LangeVC.com. See LICENSE.