AGENT SKILL

dispatch

Mode-aware dispatch skill for the claude-wayfinder deterministic 7-decision matcher. Real-catalog mode is the default — the skill reads dispatch context JSON from stdin, runs the matcher against your live catalog, and returns the decision JSON verbatim. Pass --demo to opt into bundled fixtures instead. Catalog path resolution: --catalog-path flag > $DISPATCH_CATALOG_PATH env var > canonical default (~/.claude/state/dispatch-catalog.json). If neither --demo nor a resolvable catalog is present the skill emits a [CATALOG ERROR] and exits non-zero.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 4
Forks
1

01 前言 Front Matter

name
dispatch
triggers
{ "command_prefixes": [ "/dispatch" ] }
description
Mode-aware dispatch skill for the claude-wayfinder deterministic 7-decision matcher. Real-catalog mode is the default — the skill reads dispatch context JSON from stdin, runs the matcher against your live catalog, and returns the decision JSON verbatim. Pass --demo to opt into bundled fixtures instead. Catalog path resolution: --catalog-path flag > $DISPATCH_CATALOG_PATH env var > canonical default (~/.claude/state/dispatch-catalog.json). If neither --demo nor a resolvable catalog is present the skill emits a [CATALOG ERROR] and exits non-zero.

03 SKILL.md 正文

Dispatch Skill

The /dispatch skill is a mode-aware wrapper around the claude-wayfinder deterministic 7-decision matcher.

Modes

Real-catalog mode (default)

The skill reads a dispatch context JSON from stdin, runs the matcher against your live catalog, and returns the decision JSON verbatim. This is the default — no flag required. Catalog path resolution:

  1. --catalog-path <path> CLI flag.
  2. $DISPATCH_CATALOG_PATH env var.
  3. Canonical default — $CLAUDE_HOME/state/dispatch-catalog.json or ~/.claude/state/dispatch-catalog.json (see "Canonical catalog path" below).

Demo mode (--demo flag)

Pass --demo to run the matcher against bundled demo fixtures instead of the live catalog. Returns decision output for all 7 routing branches so you can evaluate the matcher before integrating it into your router. --demo wins over --catalog-path and $DISPATCH_CATALOG_PATH — if both are present, demo mode runs and the catalog inputs are ignored.

Hard-error guarantee: Without --demo, if no catalog can be resolved (no flag, no env var, no file at the canonical path), or if the resolved file is missing/unreadable/malformed, the skill emits a [CATALOG ERROR] banner on stderr and exits non-zero. It does not silently fall back to demo mode — a broken or missing catalog is surfaced immediately so the consumer knows routing is degraded. The banner names the canonical default path and the repair hint inline.

Dispatch context JSON (real-catalog mode)

The consumer's router agent must compose a JSON object and pass it on stdin. task_description is the only required field:

{
  "task_description": "...",
  "file_paths": ["..."],
  "agent_mentions": ["..."],
  "tool_mentions": ["..."],
  "command_prefix": "...",
  "session_id": "...",
  "domain": "code",
  "posture": "build",
  "confidence": "high",
  "area_span": 1
}

All fields except task_description are optional; omit or pass null for fields that are not applicable. The four two-axis labels (domain, posture, confidence, area_span) are documented under Two-axis labels below.

session_id (optional string, added in fix #294, auto-populated in #296) — the Claude Code session identifier for the calling session. When present, this value is written verbatim into the matcher_decision log entry, enabling per-session attribution in the dispatch log.

Auto-population (issue #296): when session_id is absent from the input JSON and the CLAUDE_SESSION_ID env var is not set, the matcher automatically walks its ancestor process chain looking for a PID-keyed state file written by the session-start-record-session hook. The hook fires at SessionStart, captures the CC process's PID and start time (psutil.Process(ppid).create_time()), and writes ~/.claude/state/wayfinder-sessions/<ppid>-<create_time_int>.txt. The matcher finds its own CC ancestor's file and reads the session_id from it. The result is cached for the matcher process's lifetime.

Concurrent-session safety: each CC session gets a unique file keyed by both PID and integer create_time. Two concurrent CC sessions write two separate files; the matcher's process-tree walk reaches only its own CC ancestor's PID, so there is no cross-contamination. Do not simplify this to a single shared file — a shared file is broken under concurrent sessions (each SessionStart overwrites the prior session's ID). The per-file-per-session design is load-bearing.

Two-axis labels (Matcher v3)

Four optional caller-supplied labels the matcher consumes for two-axis (domain × posture) routing. The matcher runs no encoder — it routes on exactly what the caller labels — so these are the only signals it cannot derive lexically. They describe the task, not a target agent: the matcher resolves domain/posture to an agent internally (via its own routing policy), so the caller never names an agent here.

Caller: classify each task into these four labels and include them in the context JSON you compose — apply the rubric below and obey the confidence fail-safe (never label high on a guess). Label only what the task is.

All four are optional and additive — the matcher ignores absent or unknown fields, so emitting them never changes current behavior. While the rollout flag is off they feed shadow-mode telemetry only (the v3 route is computed and logged beside the live lexical decision); high-confidence labels begin steering live routes only after the flag flip.

Field Type Default when absent
domain enum / null is_any (no domain gate)
posture enum / null no posture route
confidence high | medium | low / null low (fail-safe)
area_span integer ≥ 1 1

domain — what kind of artifact/area the task touches

Value Task is about
code source code: features, fixes, refactors, tests (.py / .ts / …)
docs_prose documentation, READMEs, ADRs, plans, specs, changelogs
project_meta harness self-edits — agents/**, skills/**/SKILL.md, CLAUDE.md, AGENTS.md, GEMINI.md, project governance
infra_deploy infrastructure, deployment, IaC (bicep/terraform), CI/CD, topology
is_any / null cross-cutting, or no single domain dominates (no domain gate)

posture — what action the task performs

Value Intent
build create / implement / add new behavior
diagnose find the root cause of a failure / bug
assess review / evaluate existing code quality
critique adversarial critique (architecture or idea soundness)
verify conformance / consistency check vs. a stated source of truth
plan scope / design / requirements
research prior-art discovery before planning
operate read-only operations (e.g. status / read queries)

confidence — fail-safe; never label high on a guess

How sure the caller is of the domain + posture pair. The fail-safe governs delegate posture-routes: the matcher hard-routes to a preferred agent only on high; medium / low / absent fall through to the lexical scorer. A wrong high is the one label that can mis-steer a live delegate, so when unsure, omit confidence (⇒ treated as low) rather than inventing a level.

Exception: a harness self-edit (domain: project_meta, posture: build) abstains to the router (self_handle) regardless of confidence — that abstention is not a delegate, so the high-gate does not apply.

area_span — default 1

Number of distinct layers/areas the task genuinely spans (e.g. code + infra + data = 3). Defaults to 1. Emit ≥ 2 only for genuinely multi-layer work — a single-file bug is 1; an outage spanning service + database + config is ≥ 2. Values are coerced to int; anything missing, non-numeric, or < 1 becomes 1.

Output schema (both modes)

Real-catalog mode returns the matcher's decision JSON verbatim on stdout:

{
  "decision":     "delegate",
  "agent":        "Explore",
  "skills":       ["python"],
  "confidence":   0.92,
  "rationale":    "matched keywords: implement.",
  "alternatives": [{"agent": "Plan", "score": 0.4}]
}

Demo mode produces human-readable text instead of JSON (one block per decision branch). The output schema is identical across the 7 branches shown in demo mode. Consumer routers should use real-catalog mode for machine-readable output.

Consumer router requirements

The consumer's router agent must include Skill in its tools: frontmatter for /dispatch to be invocable. Example of correct frontmatter:

tools: Glob, Grep, Read, Edit, Write, Bash, Skill, ToolSearch

See docs/integration.md for the end-to-end wiring guide.

Prerequisites

claude-wayfinder must be installed in a Python ≥ 3.11 environment. After running /setup-wayfinder, the plugin venv lives at ${CLAUDE_PLUGIN_DATA}/venv/ — that is the canonical interpreter to use.

Use the plugin venv's interpreter explicitly. Do not rely on bare python resolving to the right environment via $PATH — on Windows in particular, a global Python (e.g. C:\Python313\python.exe) often takes precedence over the venv and does NOT have claude-wayfinder installed, producing No module named claude_wayfinder at runtime.

Platform Path
POSIX ${CLAUDE_PLUGIN_DATA}/venv/bin/python
Windows ${CLAUDE_PLUGIN_DATA}/venv/Scripts/python.exe

The skill's invocations below show the explicit path. Bare python is fine as a shorthand only when the plugin venv is activated in the calling shell, or its bin/Scripts dir is first on $PATH.

# Confirm the package is available
"${CLAUDE_PLUGIN_DATA}/venv/Scripts/python.exe" -m claude_wayfinder dispatch --help   # Windows
"${CLAUDE_PLUGIN_DATA}/venv/bin/python" -m claude_wayfinder dispatch --help           # POSIX

If claude-wayfinder is not installed yet, run /setup-wayfinder — that skill materializes the venv at the canonical location and pins the matching plugin version into it.

Canonical catalog path

The live catalog is at ~/.claude/state/dispatch-catalog.json (or $CLAUDE_HOME/state/dispatch-catalog.json when $CLAUDE_HOME is set). This is the default real-catalog mode resolves to when neither --catalog-path nor $DISPATCH_CATALOG_PATH is supplied. Override only for test fixtures or unusual deployments. The bundled hooks (refresh-catalog-on-stale.js, check-catalog-health.js) use the same default.

Running

PY="${CLAUDE_PLUGIN_DATA}/venv/Scripts/python.exe"   # Windows
# PY="${CLAUDE_PLUGIN_DATA}/venv/bin/python"          # POSIX

# DISPATCH_SHADOW gates the v3 shadow-route compute (see "Two-axis labels"
# above). The computed route remains telemetry-only unless its domain is
# also enabled by DISPATCH_HARD_ROUTING_DOMAINS. Sourced from the userConfig
# `shadow_enabled` toggle in plugin.json so users can opt out without
# touching source.
#
# Only an exact case-insensitive match of "0", "false", or "no" disables
# shadow compute. Every other value fails open to enabled — this
# includes "off", "disabled", an empty string, and any malformed or
# whitespace-padded value (e.g. " false "). Absent (unset) also fails
# open to enabled.
export DISPATCH_SHADOW="${user_config.shadow_enabled}"

# DISPATCH_HARD_ROUTING_DOMAINS is the surgical serving gate: a comma-
# separated set of domain tokens such as "is_any" or "is_any,code".
# Unset, empty, or whitespace-only values fail closed to OFF. Tokens are
# stripped, lowercased, and deduplicated; unrecognized tokens are dropped
# with a stderr warning rather than failing dispatch.
export DISPATCH_HARD_ROUTING_DOMAINS="is_any"

# The two gates have opposite safe defaults: absent DISPATCH_SHADOW means
# Compose compute is ON, while absent DISPATCH_HARD_ROUTING_DOMAINS means
# hard routing is OFF. DISPATCH_HARD_ROUTING_DOMAINS scopes cutover per
# domain; DISPATCH_SHADOW=0 is the coarse, all-domain kill switch because it
# skips Compose entirely and therefore always serves the lexical decision.

# Real-catalog mode — default; resolves to the canonical catalog
echo '{"task_description": "implement auth module", "file_paths": ["src/auth.py"], "agent_mentions": [], "tool_mentions": [], "command_prefix": null}' \
  | "$PY" -m claude_wayfinder dispatch

# Demo mode — opt in with --demo
"$PY" -m claude_wayfinder dispatch --demo

# Explicit catalog override (e.g. test fixture)
export DISPATCH_CATALOG_PATH=/path/to/test-catalog.json   # POSIX
# $env:DISPATCH_CATALOG_PATH = "C:\path\to\test-catalog.json"  # PowerShell
echo '{...}' | "$PY" -m claude_wayfinder dispatch

Stale-catalog warning

When $DISPATCH_SKILLS_DIR and/or $DISPATCH_AGENTS_DIR are set and any source file within them has a modification time newer than the catalog file, the skill emits a warning to stderr:

[DISPATCH WARNING] Catalog mtime is older than source files: ...
Consider running `claude-wayfinder catalog build` to refresh.
Proceeding with stale catalog.

Execution proceeds with the stale catalog — staleness is a degraded-quality signal, not an error. Run "$PY" -m claude_wayfinder catalog build (using the same explicit interpreter path resolved in the Running section above) to refresh.

The 7 decision branches

Branch When it fires
needs_more_detail Feature density < 2; provide more context to route accurately.
delegate One agent scores ≥ 0.85 with a gap ≥ 0.2 above the next.
self_handle At least one skill scores ≥ 0.5; no dominant agent.
mixed_content Gap < 0.2; ≥ 2 agents clamped at 1.0 on path-disjoint lanes. Output includes lanes[] (agent, score, matched_paths, skills per lane) and unassigned_paths[].
advisory Best agent ≥ 0.5. Covers both tie (gap < 0.2, rationale includes gap=) and marginal (gap ≥ 0.2 but score < 0.85) cases. Top agent named; alternatives populated.
ask_user Reserved — not produced by the v0.1 matcher.
self_handle_unaided Nothing scores above threshold; proceed without delegation.

04 同仓库更多 Skills