Phases Agents

phases-agents

English · Français

A local MCP server that discovers, validates and selects skills deterministically. Python standard library only, no runtime dependencies.

One server. Five tools. Nothing executed behind your back.

Why

AI agents improvise. Ask the same question twice and you get two different plans. That is fine for brainstorming, and unacceptable for audit and compliance work.

phases-agents removes the improvisation. It profiles a local project, validates a catalogue of skills against a strict contract, and returns a plan that can be replayed. Same target, same catalogue, same parameters, same decision.

Principle

configured root identifiers
→ bounded discovery
→ official validation
→ immutable registry
→ verified cache
→ detector profile
→ deterministic selection
→ MCP plan

The server selects and exposes. The calling model reads the selected skills and decides what to do with them, using its own tools. The server never executes a skill.

Architecture

The modules live in src/phases_agents/.

The normative contract lives in src/phases_agents/core/SKILLS_CONTRACT.md (French).

Quick start

An example package ships in examples/skills/. Three steps produce a real plan.

git clone https://github.com/Cherridsaid/phases-agents && cd phases-agents

Create skills-roots.json pointing at the example root:

{
  "config_version": "1.0",
  "roots": [
    { "id": "demo", "path": "/absolute/path/to/phases-agents/examples/skills" }
  ]
}

pip install -e .
phases-agents --skills-config /absolute/path/to/skills-roots.json

pip install -e . installs the phases-agents command. Without installing, the same server starts with PYTHONPATH=src python -m phases_agents.server.

The server reads JSON-RPC line by line on standard input. A phases_agents_plan call against a Python project then selects hello-python:

{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"phases_agents_plan",
 "arguments":{"root_ids":["demo"],"target":"/absolute/path/to/a/project",
 "today":"2026-08-27","plan_version":"B3",
 "client_capabilities":["filesystem_read","filesystem_search"]}}}

Two plan formats coexist. "B3" names the versioned format, not a server version; its official schema is src/phases_agents/core/PLAN_B3_SCHEMA.json. Use it for new work. Without plan_version, the legacy format returns a flat list of steps; it is kept so existing callers do not break, and will be deprecated before removal. client_capabilities is only accepted in B3, since declaring what the client can do only makes sense in that format.

Connecting an MCP client

Claude Code (.mcp.json at the root of your project):

{
  "mcpServers": {
    "phases-agents": {
      "command": "phases-agents",
      "args": [
        "--skills-config",
        "/absolute/path/to/skills-roots.json"
      ]
    }
  }
}

Codex uses the same command/arguments pair in its own configuration file. No token and no environment variable is required.

MCP tools

detect(target)
list_skills(root_ids, today)
get_skill(root_ids, today, skill_id)
plan(root_ids, target, today, constraints?)
plan(root_ids, target, today, plan_version, client_capabilities?)
refresh_skills(root_ids, today)

today is injected rather than read from a clock, so every call is replayable. get_skill takes an identifier, never a path, and its content comes from the validated snapshot. Absolute paths and detected secrets are masked in public output. Any encoded JSON-RPC response stays under 1 MiB.

The first call builds the validated registry. Warm calls verify metadata without re-reading contents. refresh_skills forces a rebuild.

Writing a skill package

Each package is a direct child of a root and contains at least:

<root>/<skill-id>/SKILL.md
<root>/<skill-id>/phases.json

The fastest way to start is to copy examples/skills/hello-python/ and rename the identifier.

SKILL.md frontmatter

Five keys are allowed. All optional, all checked when present.

The fourteen required sections

Each is a Markdown heading (##), in any order. Section titles are French, because they belong to the contract; the body is yours to write in any language.

Loi centrale · Ce que ce skill fait · Ce que ce skill ne fait pas · Conditions d'activation · Conditions d'exclusion · Capacites necessaires · Interdictions · Methode d'audit · Contrat de preuve · Format de sortie · Conditions de blocage · Limites connues · Exemples d'entree · Exemple de sortie attendue

phases.json fields

All required: schema_version, id, version, title, description, domain, project_types, platforms, activation, exclusions, requires_capabilities, optional_capabilities, forbidden_capabilities, execution_mode, human_approval, output_schema, rules_path, references_path, scripts_path, tests_path, files.

output_schema uses the symbolic form core:SCHEMA_NAME.json.

Closed vocabularies

project_types must intersect what the detector can emit: apk, python, skill_package, solana, web.

activation.any uses profile facts: collects_personal_data, has_api, has_apk, has_authentication, has_database, has_ecommerce, has_eu_context, has_file_upload, has_javascript, has_python, has_rust, has_skill_packages, has_solana, has_source_code, has_typescript, has_web, uses_ai, uses_payments.

requires_capabilities, optional_capabilities and forbidden_capabilities use: browser, dependency_installation, filesystem_read, filesystem_search, filesystem_write, human_question, shell, target_code_execution, web.

Provided capabilities are an open vocabulary: each catalogue names what it brings, and only the shape is enforced (^[a-z][a-z0-9_]{0,63}$). Only client capabilities are closed, because they describe the protocol rather than your domain.

A domain of legal, juridique, regulatory or compliance triggers an extra regime: every rule cited must carry an official source, a jurisdiction and a verification date.

What the schema does and does not say

SKILL_MANIFEST_SCHEMA.json describes the shape of phases.json: required fields, types, closed vocabularies.

The schema engine is deliberately minimal. It applies enum, minLength and minItems, and nothing else: no pattern, no if/then, no oneOf. A schema using those keywords would itself be rejected.

The consequence matters: conditional rules live in validator.py, which remains the source of truth. The version rule is the example: provides_capabilities is forbidden in a 1.0 manifest and required in a 1.1 one. That rule is enforced and tested, but it is not expressible in the schema. Do not read required as the whole contract.

A package with only SKILL.md fails. An invalid package blocks the registry rather than degrading silently.

Identity

phases.json.id is the identity, and SKILL.md.name must match it. The directory must carry the same key. Keys are normalised with NFKC then casefold, so homoglyphs cannot smuggle in a second identity. Any collision blocks the whole build; no package is elected arbitrarily.

Selection

Every valid skill in the registry lands in exactly one category, with its reason. Nothing is discarded silently.

The only proven automatic signal is:

project_types ∩ profile.types

Platform, domain and capabilities filter only when the caller supplies those constraints. A forbidden capability rejects the skill. No semantic score is invented, and the plan is sorted by identifier.

An empty plan is explicitly valid: it carries NO_COMPATIBLE_SKILL.

The B3 plan classifies every installed skill across skills_selected, skills_not_applicable and skills_blocked; each skill appears exactly once. skills_missing lists capabilities with no executable provider, derived from confirmed facts only. A gap never proves non-compliance: it says an audit deemed necessary is not covered.

Limits

• 16 roots maximum
• direct depth only
• 1,000 packages maximum
• 10,000 entries per root
• SKILL.md capped at 256 KiB
• a single reference capped at 256 KiB, 1 MiB in total
• snapshots capped at 16 MiB
• public result capped at 1 MiB
• 100 issues per package
• fingerprint capped at 100,000 nodes

Callers may only lower these limits, never raise them.

Runtime constraints

• Python >=3.11
• no third-party runtime dependency
• no implicit network
• no runtime shell
• no target code executed
• no implicit clock
• no telemetry
• no skill downloaded

pytest is a development dependency only.

Tests

python -m pytest -q

Expected result, 778 collected, 0 failed on every platform:

Linux           : 771 passed, 7 skipped
Windows (CI)    : 778 passed, 0 skipped
Windows (local) : 776 passed, 2 skipped

The skips are platform capabilities, not failures. Linux skips the Windows-junction tests, which have no POSIX equivalent. Windows skips the two symlink tests only when the local privilege to create symlinks is missing, so a runner that holds it reports no skip at all. Every count above comes from a real run: the two Windows lines are the GitHub Actions runner and a local machine.

Normative texts are checked out with LF endings, enforced by .gitattributes. A couple of Windows symlink tests are skipped: they need a local Windows privilege. Windows junctions are genuinely tested.

Level of proof

The validator confirms one thing only:

STRUCTURALLY_VALIDATED

It does not verify the real target. TARGET_VERIFIED stays forbidden in V1.

Security

The loader refuses reparse points. Reads are bounded and confined. Output is sorted and deterministic.

One design decision deserves your attention: detect and plan take a target path that is not confined to the configured roots, because the point is to profile an arbitrary project. Run this server under an account whose reach you accept, and connect it only to a trusted client. The full threat model is in SECURITY.md.

Non-guarantees

• no universal semantic relevance
• no external skill approved automatically
• no audit of script contents
• no genuinely mounted target proof
• no total Windows atomicity
• no universal HTML recognition
• no universal secret detection
• no guaranteed legal compliance
• no marketplace, no remote source

Licence

Apache-2.0. See LICENSE and NOTICE.