[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M72TY9BCDXF5XVAQVBT4BSM6":3,"skill-same:01M72TY9BCDXF5XVAQVBT4BSM6":49},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":27,"commitSha":28,"skillDirectory":29,"forksCount":30,"repositoryHtmlUrl":31,"repositoryHomepageUrl":32,"ownerHtmlUrl":33,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":34,"install":35,"resources":48},"01M72TY9BCDXF5XVAQVBT4BSM6","dispatch",null,"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 (~\u002F.claude\u002Fstate\u002Fdispatch-catalog.json). If neither --demo nor a resolvable catalog is present the skill emits a [CATALOG ERROR] and exits non-zero.","glitchwerks\u002Fclaude-wayfinder","glitchwerks","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F280158622?v=4","01M72TXS0Q6NRNKZYFQ4KFG1VJ","01M72TXS18Q7XJ6RNHV7K0D9XF",4,"MIT","https:\u002F\u002Fgithub.com\u002Fglitchwerks\u002Fclaude-wayfinder\u002Fblob\u002Fmain\u002Fskills\u002Fdispatch\u002FSKILL.md",[],[],"2026-11-10 22:39:42",false,"---\nname: dispatch\ndescription: >\n  Mode-aware dispatch skill for the claude-wayfinder deterministic 7-decision\n  matcher. Real-catalog mode is the default — the skill reads dispatch context\n  JSON from stdin, runs the matcher against your live catalog, and returns\n  the decision JSON verbatim. Pass --demo to opt into bundled fixtures\n  instead. Catalog path resolution: --catalog-path flag > $DISPATCH_CATALOG_PATH\n  env var > canonical default (~\u002F.claude\u002Fstate\u002Fdispatch-catalog.json). If\n  neither --demo nor a resolvable catalog is present the skill emits a\n  [CATALOG ERROR] and exits non-zero.\ntriggers:\n  command_prefixes:\n    - \u002Fdispatch\n---\n\n# Dispatch Skill\n\nThe `\u002Fdispatch` skill is a mode-aware wrapper around the\n**claude-wayfinder** deterministic 7-decision matcher.\n\n## Modes\n\n### Real-catalog mode (default)\n\nThe skill reads a dispatch context JSON from stdin, runs the matcher\nagainst your live catalog, and returns the decision JSON verbatim. This\nis the default — no flag required. Catalog path resolution:\n\n1. `--catalog-path \u003Cpath>` CLI flag.\n2. `$DISPATCH_CATALOG_PATH` env var.\n3. **Canonical default** — `$CLAUDE_HOME\u002Fstate\u002Fdispatch-catalog.json` or\n   `~\u002F.claude\u002Fstate\u002Fdispatch-catalog.json` (see \"Canonical catalog path\"\n   below).\n\n### Demo mode (`--demo` flag)\n\nPass `--demo` to run the matcher against bundled demo fixtures instead\nof the live catalog. Returns decision output for all 7 routing branches\nso you can evaluate the matcher before integrating it into your router.\n`--demo` wins over `--catalog-path` and `$DISPATCH_CATALOG_PATH` — if\nboth are present, demo mode runs and the catalog inputs are ignored.\n\n**Hard-error guarantee:** Without `--demo`, if no catalog can be\nresolved (no flag, no env var, no file at the canonical path), or if\nthe resolved file is missing\u002Funreadable\u002Fmalformed, the skill emits a\n`[CATALOG ERROR]` banner on stderr and exits non-zero. It does **not**\nsilently fall back to demo mode — a broken or missing catalog is\nsurfaced immediately so the consumer knows routing is degraded. The\nbanner names the canonical default path and the repair hint inline.\n\n## Dispatch context JSON (real-catalog mode)\n\nThe consumer's router agent must compose a JSON object and pass it on\nstdin. `task_description` is the only required field:\n\n```json\n{\n  \"task_description\": \"...\",\n  \"file_paths\": [\"...\"],\n  \"agent_mentions\": [\"...\"],\n  \"tool_mentions\": [\"...\"],\n  \"command_prefix\": \"...\",\n  \"session_id\": \"...\",\n  \"domain\": \"code\",\n  \"posture\": \"build\",\n  \"confidence\": \"high\",\n  \"area_span\": 1\n}\n```\n\nAll fields except `task_description` are optional; omit or pass `null`\nfor fields that are not applicable. The four two-axis labels (`domain`,\n`posture`, `confidence`, `area_span`) are documented under **Two-axis\nlabels** below.\n\n`session_id` (optional string, added in fix #294, auto-populated in\n#296) — the Claude Code session identifier for the calling session.\nWhen present, this value is written verbatim into the `matcher_decision`\nlog entry, enabling per-session attribution in the dispatch log.\n\n**Auto-population (issue #296):** when `session_id` is absent from the\ninput JSON and the `CLAUDE_SESSION_ID` env var is not set, the matcher\nautomatically walks its ancestor process chain looking for a PID-keyed\nstate file written by the `session-start-record-session` hook. The hook\nfires at SessionStart, captures the CC process's PID and start time\n(`psutil.Process(ppid).create_time()`), and writes\n`~\u002F.claude\u002Fstate\u002Fwayfinder-sessions\u002F\u003Cppid>-\u003Ccreate_time_int>.txt`. The\nmatcher finds its own CC ancestor's file and reads the session_id from\nit. The result is cached for the matcher process's lifetime.\n\n**Concurrent-session safety:** each CC session gets a unique file keyed\nby both PID and integer create_time. Two concurrent CC sessions write\ntwo separate files; the matcher's process-tree walk reaches only its own\nCC ancestor's PID, so there is no cross-contamination. Do **not**\nsimplify this to a single shared file — a shared file is broken under\nconcurrent sessions (each SessionStart overwrites the prior session's\nID). The per-file-per-session design is load-bearing.\n\n## Two-axis labels (Matcher v3)\n\nFour optional **caller-supplied labels** the matcher consumes for two-axis\n(domain × posture) routing. The matcher runs no encoder — it routes on exactly\nwhat the caller labels — so these are the only signals it cannot derive\nlexically. They describe the **task**, not a target agent: the matcher resolves\ndomain\u002Fposture to an agent internally (via its own routing policy), so the caller\nnever names an agent here.\n\n**Caller: classify each task into these four labels and include them in the\ncontext JSON you compose** — apply the rubric below and obey the `confidence`\nfail-safe (never label `high` on a guess). Label only what the task _is_.\n\nAll four are **optional and additive** — the matcher ignores absent or unknown\nfields, so emitting them never changes current behavior. While the rollout flag\nis off they feed shadow-mode telemetry only (the v3 route is computed and logged\nbeside the live lexical decision); `high`-confidence labels begin steering live\nroutes only after the flag flip.\n\n| Field | Type | Default when absent |\n| --- | --- | --- |\n| `domain` | enum \u002F `null` | `is_any` (no domain gate) |\n| `posture` | enum \u002F `null` | no posture route |\n| `confidence` | `high` \\| `medium` \\| `low` \u002F `null` | **`low`** (fail-safe) |\n| `area_span` | integer ≥ 1 | `1` |\n\n### `domain` — what kind of artifact\u002Farea the task touches\n\n| Value | Task is about |\n| --- | --- |\n| `code` | source code: features, fixes, refactors, tests (`.py` \u002F `.ts` \u002F …) |\n| `docs_prose` | documentation, READMEs, ADRs, plans, specs, changelogs |\n| `project_meta` | harness self-edits — `agents\u002F**`, `skills\u002F**\u002FSKILL.md`, `CLAUDE.md`, `AGENTS.md`, `GEMINI.md`, project governance |\n| `infra_deploy` | infrastructure, deployment, IaC (bicep\u002Fterraform), CI\u002FCD, topology |\n| `is_any` \u002F `null` | cross-cutting, or no single domain dominates (no domain gate) |\n\n### `posture` — what action the task performs\n\n| Value | Intent |\n| --- | --- |\n| `build` | create \u002F implement \u002F add new behavior |\n| `diagnose` | find the root cause of a failure \u002F bug |\n| `assess` | review \u002F evaluate existing code quality |\n| `critique` | adversarial critique (architecture or idea soundness) |\n| `verify` | conformance \u002F consistency check vs. a stated source of truth |\n| `plan` | scope \u002F design \u002F requirements |\n| `research` | prior-art discovery before planning |\n| `operate` | read-only operations (e.g. status \u002F read queries) |\n\n### `confidence` — fail-safe; never label `high` on a guess\n\nHow sure the caller is of the `domain` + `posture` pair. The fail-safe governs\n**delegate** posture-routes: the matcher hard-routes to a preferred agent **only\non `high`**; `medium` \u002F `low` \u002F absent fall through to the lexical scorer. A wrong\n`high` is the one label that can mis-steer a live delegate, so when unsure, omit\n`confidence` (⇒ treated as `low`) rather than inventing a level.\n\n**Exception:** a harness self-edit (`domain: project_meta`, `posture: build`)\nabstains to the router (`self_handle`) **regardless of confidence** — that\nabstention is not a delegate, so the `high`-gate does not apply.\n\n### `area_span` — default 1\n\nNumber of distinct layers\u002Fareas the task genuinely spans (e.g. code + infra +\ndata = 3). Defaults to `1`. Emit `≥ 2` **only** for genuinely multi-layer work — a\nsingle-file bug is `1`; an outage spanning service + database + config is `≥ 2`.\nValues are coerced to `int`; anything missing, non-numeric, or `\u003C 1` becomes `1`.\n\n## Output schema (both modes)\n\nReal-catalog mode returns the matcher's decision JSON verbatim on stdout:\n\n```json\n{\n  \"decision\":     \"delegate\",\n  \"agent\":        \"Explore\",\n  \"skills\":       [\"python\"],\n  \"confidence\":   0.92,\n  \"rationale\":    \"matched keywords: implement.\",\n  \"alternatives\": [{\"agent\": \"Plan\", \"score\": 0.4}]\n}\n```\n\nDemo mode produces human-readable text instead of JSON (one block per\ndecision branch).  The output schema is identical across the 7 branches\nshown in demo mode.  Consumer routers should use real-catalog mode for\nmachine-readable output.\n\n## Consumer router requirements\n\nThe consumer's router agent must include `Skill` in its `tools:` frontmatter\nfor `\u002Fdispatch` to be invocable.  Example of correct frontmatter:\n\n```\ntools: Glob, Grep, Read, Edit, Write, Bash, Skill, ToolSearch\n```\n\nSee `docs\u002Fintegration.md` for the end-to-end wiring guide.\n\n## Prerequisites\n\n`claude-wayfinder` must be installed in a Python ≥ 3.11 environment. After\nrunning `\u002Fsetup-wayfinder`, the plugin venv lives at\n`${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002F` — that is the canonical interpreter to use.\n\n**Use the plugin venv's interpreter explicitly.** Do not rely on bare\n`python` resolving to the right environment via `$PATH` — on Windows in\nparticular, a global Python (e.g. `C:\\Python313\\python.exe`) often takes\nprecedence over the venv and does NOT have `claude-wayfinder` installed,\nproducing `No module named claude_wayfinder` at runtime.\n\n| Platform | Path |\n|----------|------|\n| POSIX    | `${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002Fbin\u002Fpython` |\n| Windows  | `${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002FScripts\u002Fpython.exe` |\n\nThe skill's invocations below show the explicit path. Bare `python` is\nfine as a shorthand **only** when the plugin venv is activated in the\ncalling shell, or its `bin\u002FScripts` dir is first on `$PATH`.\n\n```bash\n# Confirm the package is available\n\"${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002FScripts\u002Fpython.exe\" -m claude_wayfinder dispatch --help   # Windows\n\"${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002Fbin\u002Fpython\" -m claude_wayfinder dispatch --help           # POSIX\n```\n\nIf `claude-wayfinder` is not installed yet, run `\u002Fsetup-wayfinder` — that\nskill materializes the venv at the canonical location and pins the\nmatching plugin version into it.\n\n## Canonical catalog path\n\nThe live catalog is at **`~\u002F.claude\u002Fstate\u002Fdispatch-catalog.json`** (or\n`$CLAUDE_HOME\u002Fstate\u002Fdispatch-catalog.json` when `$CLAUDE_HOME` is set).\nThis is the default real-catalog mode resolves to when neither\n`--catalog-path` nor `$DISPATCH_CATALOG_PATH` is supplied. Override only\nfor test fixtures or unusual deployments. The bundled hooks\n(`refresh-catalog-on-stale.js`, `check-catalog-health.js`) use the same\ndefault.\n\n## Running\n\n```bash\nPY=\"${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002FScripts\u002Fpython.exe\"   # Windows\n# PY=\"${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002Fbin\u002Fpython\"          # POSIX\n\n# DISPATCH_SHADOW gates the v3 shadow-route compute (see \"Two-axis labels\"\n# above). The computed route remains telemetry-only unless its domain is\n# also enabled by DISPATCH_HARD_ROUTING_DOMAINS. Sourced from the userConfig\n# `shadow_enabled` toggle in plugin.json so users can opt out without\n# touching source.\n#\n# Only an exact case-insensitive match of \"0\", \"false\", or \"no\" disables\n# shadow compute. Every other value fails open to enabled — this\n# includes \"off\", \"disabled\", an empty string, and any malformed or\n# whitespace-padded value (e.g. \" false \"). Absent (unset) also fails\n# open to enabled.\nexport DISPATCH_SHADOW=\"${user_config.shadow_enabled}\"\n\n# DISPATCH_HARD_ROUTING_DOMAINS is the surgical serving gate: a comma-\n# separated set of domain tokens such as \"is_any\" or \"is_any,code\".\n# Unset, empty, or whitespace-only values fail closed to OFF. Tokens are\n# stripped, lowercased, and deduplicated; unrecognized tokens are dropped\n# with a stderr warning rather than failing dispatch.\nexport DISPATCH_HARD_ROUTING_DOMAINS=\"is_any\"\n\n# The two gates have opposite safe defaults: absent DISPATCH_SHADOW means\n# Compose compute is ON, while absent DISPATCH_HARD_ROUTING_DOMAINS means\n# hard routing is OFF. DISPATCH_HARD_ROUTING_DOMAINS scopes cutover per\n# domain; DISPATCH_SHADOW=0 is the coarse, all-domain kill switch because it\n# skips Compose entirely and therefore always serves the lexical decision.\n\n# Real-catalog mode — default; resolves to the canonical catalog\necho '{\"task_description\": \"implement auth module\", \"file_paths\": [\"src\u002Fauth.py\"], \"agent_mentions\": [], \"tool_mentions\": [], \"command_prefix\": null}' \\\n  | \"$PY\" -m claude_wayfinder dispatch\n\n# Demo mode — opt in with --demo\n\"$PY\" -m claude_wayfinder dispatch --demo\n\n# Explicit catalog override (e.g. test fixture)\nexport DISPATCH_CATALOG_PATH=\u002Fpath\u002Fto\u002Ftest-catalog.json   # POSIX\n# $env:DISPATCH_CATALOG_PATH = \"C:\\path\\to\\test-catalog.json\"  # PowerShell\necho '{...}' | \"$PY\" -m claude_wayfinder dispatch\n```\n\n## Stale-catalog warning\n\nWhen `$DISPATCH_SKILLS_DIR` and\u002For `$DISPATCH_AGENTS_DIR` are set and any\nsource file within them has a modification time newer than the catalog\nfile, the skill emits a warning to stderr:\n\n```\n[DISPATCH WARNING] Catalog mtime is older than source files: ...\nConsider running `claude-wayfinder catalog build` to refresh.\nProceeding with stale catalog.\n```\n\nExecution **proceeds** with the stale catalog — staleness is a\ndegraded-quality signal, not an error.  Run\n`\"$PY\" -m claude_wayfinder catalog build` (using the same explicit\ninterpreter path resolved in the **Running** section above) to refresh.\n\n## The 7 decision branches\n\n| Branch              | When it fires                                                   |\n|---------------------|-----------------------------------------------------------------|\n| `needs_more_detail` | Feature density \u003C 2; provide more context to route accurately. |\n| `delegate`          | One agent scores ≥ 0.85 with a gap ≥ 0.2 above the next.      |\n| `self_handle`       | At least one skill scores ≥ 0.5; no dominant agent.            |\n| `mixed_content`     | Gap \u003C 0.2; ≥ 2 agents clamped at 1.0 on path-disjoint lanes. Output includes `lanes[]` (agent, score, matched_paths, skills per lane) and `unassigned_paths[]`. |\n| `advisory`          | Best agent ≥ 0.5. Covers both tie (gap \u003C 0.2, rationale includes `gap=`) and marginal (gap ≥ 0.2 but score \u003C 0.85) cases. Top agent named; alternatives populated. |\n| `ask_user`          | Reserved — not produced by the v0.1 matcher.                   |\n| `self_handle_unaided` | Nothing scores above threshold; proceed without delegation. |\n","# Dispatch Skill\n\nThe `\u002Fdispatch` skill is a mode-aware wrapper around the\n**claude-wayfinder** deterministic 7-decision matcher.\n\n## Modes\n\n### Real-catalog mode (default)\n\nThe skill reads a dispatch context JSON from stdin, runs the matcher\nagainst your live catalog, and returns the decision JSON verbatim. This\nis the default — no flag required. Catalog path resolution:\n\n1. `--catalog-path \u003Cpath>` CLI flag.\n2. `$DISPATCH_CATALOG_PATH` env var.\n3. **Canonical default** — `$CLAUDE_HOME\u002Fstate\u002Fdispatch-catalog.json` or\n   `~\u002F.claude\u002Fstate\u002Fdispatch-catalog.json` (see \"Canonical catalog path\"\n   below).\n\n### Demo mode (`--demo` flag)\n\nPass `--demo` to run the matcher against bundled demo fixtures instead\nof the live catalog. Returns decision output for all 7 routing branches\nso you can evaluate the matcher before integrating it into your router.\n`--demo` wins over `--catalog-path` and `$DISPATCH_CATALOG_PATH` — if\nboth are present, demo mode runs and the catalog inputs are ignored.\n\n**Hard-error guarantee:** Without `--demo`, if no catalog can be\nresolved (no flag, no env var, no file at the canonical path), or if\nthe resolved file is missing\u002Funreadable\u002Fmalformed, the skill emits a\n`[CATALOG ERROR]` banner on stderr and exits non-zero. It does **not**\nsilently fall back to demo mode — a broken or missing catalog is\nsurfaced immediately so the consumer knows routing is degraded. The\nbanner names the canonical default path and the repair hint inline.\n\n## Dispatch context JSON (real-catalog mode)\n\nThe consumer's router agent must compose a JSON object and pass it on\nstdin. `task_description` is the only required field:\n\n```json\n{\n  \"task_description\": \"...\",\n  \"file_paths\": [\"...\"],\n  \"agent_mentions\": [\"...\"],\n  \"tool_mentions\": [\"...\"],\n  \"command_prefix\": \"...\",\n  \"session_id\": \"...\",\n  \"domain\": \"code\",\n  \"posture\": \"build\",\n  \"confidence\": \"high\",\n  \"area_span\": 1\n}\n```\n\nAll fields except `task_description` are optional; omit or pass `null`\nfor fields that are not applicable. The four two-axis labels (`domain`,\n`posture`, `confidence`, `area_span`) are documented under **Two-axis\nlabels** below.\n\n`session_id` (optional string, added in fix #294, auto-populated in\n#296) — the Claude Code session identifier for the calling session.\nWhen present, this value is written verbatim into the `matcher_decision`\nlog entry, enabling per-session attribution in the dispatch log.\n\n**Auto-population (issue #296):** when `session_id` is absent from the\ninput JSON and the `CLAUDE_SESSION_ID` env var is not set, the matcher\nautomatically walks its ancestor process chain looking for a PID-keyed\nstate file written by the `session-start-record-session` hook. The hook\nfires at SessionStart, captures the CC process's PID and start time\n(`psutil.Process(ppid).create_time()`), and writes\n`~\u002F.claude\u002Fstate\u002Fwayfinder-sessions\u002F\u003Cppid>-\u003Ccreate_time_int>.txt`. The\nmatcher finds its own CC ancestor's file and reads the session_id from\nit. The result is cached for the matcher process's lifetime.\n\n**Concurrent-session safety:** each CC session gets a unique file keyed\nby both PID and integer create_time. Two concurrent CC sessions write\ntwo separate files; the matcher's process-tree walk reaches only its own\nCC ancestor's PID, so there is no cross-contamination. Do **not**\nsimplify this to a single shared file — a shared file is broken under\nconcurrent sessions (each SessionStart overwrites the prior session's\nID). The per-file-per-session design is load-bearing.\n\n## Two-axis labels (Matcher v3)\n\nFour optional **caller-supplied labels** the matcher consumes for two-axis\n(domain × posture) routing. The matcher runs no encoder — it routes on exactly\nwhat the caller labels — so these are the only signals it cannot derive\nlexically. They describe the **task**, not a target agent: the matcher resolves\ndomain\u002Fposture to an agent internally (via its own routing policy), so the caller\nnever names an agent here.\n\n**Caller: classify each task into these four labels and include them in the\ncontext JSON you compose** — apply the rubric below and obey the `confidence`\nfail-safe (never label `high` on a guess). Label only what the task _is_.\n\nAll four are **optional and additive** — the matcher ignores absent or unknown\nfields, so emitting them never changes current behavior. While the rollout flag\nis off they feed shadow-mode telemetry only (the v3 route is computed and logged\nbeside the live lexical decision); `high`-confidence labels begin steering live\nroutes only after the flag flip.\n\n| Field | Type | Default when absent |\n| --- | --- | --- |\n| `domain` | enum \u002F `null` | `is_any` (no domain gate) |\n| `posture` | enum \u002F `null` | no posture route |\n| `confidence` | `high` \\| `medium` \\| `low` \u002F `null` | **`low`** (fail-safe) |\n| `area_span` | integer ≥ 1 | `1` |\n\n### `domain` — what kind of artifact\u002Farea the task touches\n\n| Value | Task is about |\n| --- | --- |\n| `code` | source code: features, fixes, refactors, tests (`.py` \u002F `.ts` \u002F …) |\n| `docs_prose` | documentation, READMEs, ADRs, plans, specs, changelogs |\n| `project_meta` | harness self-edits — `agents\u002F**`, `skills\u002F**\u002FSKILL.md`, `CLAUDE.md`, `AGENTS.md`, `GEMINI.md`, project governance |\n| `infra_deploy` | infrastructure, deployment, IaC (bicep\u002Fterraform), CI\u002FCD, topology |\n| `is_any` \u002F `null` | cross-cutting, or no single domain dominates (no domain gate) |\n\n### `posture` — what action the task performs\n\n| Value | Intent |\n| --- | --- |\n| `build` | create \u002F implement \u002F add new behavior |\n| `diagnose` | find the root cause of a failure \u002F bug |\n| `assess` | review \u002F evaluate existing code quality |\n| `critique` | adversarial critique (architecture or idea soundness) |\n| `verify` | conformance \u002F consistency check vs. a stated source of truth |\n| `plan` | scope \u002F design \u002F requirements |\n| `research` | prior-art discovery before planning |\n| `operate` | read-only operations (e.g. status \u002F read queries) |\n\n### `confidence` — fail-safe; never label `high` on a guess\n\nHow sure the caller is of the `domain` + `posture` pair. The fail-safe governs\n**delegate** posture-routes: the matcher hard-routes to a preferred agent **only\non `high`**; `medium` \u002F `low` \u002F absent fall through to the lexical scorer. A wrong\n`high` is the one label that can mis-steer a live delegate, so when unsure, omit\n`confidence` (⇒ treated as `low`) rather than inventing a level.\n\n**Exception:** a harness self-edit (`domain: project_meta`, `posture: build`)\nabstains to the router (`self_handle`) **regardless of confidence** — that\nabstention is not a delegate, so the `high`-gate does not apply.\n\n### `area_span` — default 1\n\nNumber of distinct layers\u002Fareas the task genuinely spans (e.g. code + infra +\ndata = 3). Defaults to `1`. Emit `≥ 2` **only** for genuinely multi-layer work — a\nsingle-file bug is `1`; an outage spanning service + database + config is `≥ 2`.\nValues are coerced to `int`; anything missing, non-numeric, or `\u003C 1` becomes `1`.\n\n## Output schema (both modes)\n\nReal-catalog mode returns the matcher's decision JSON verbatim on stdout:\n\n```json\n{\n  \"decision\":     \"delegate\",\n  \"agent\":        \"Explore\",\n  \"skills\":       [\"python\"],\n  \"confidence\":   0.92,\n  \"rationale\":    \"matched keywords: implement.\",\n  \"alternatives\": [{\"agent\": \"Plan\", \"score\": 0.4}]\n}\n```\n\nDemo mode produces human-readable text instead of JSON (one block per\ndecision branch).  The output schema is identical across the 7 branches\nshown in demo mode.  Consumer routers should use real-catalog mode for\nmachine-readable output.\n\n## Consumer router requirements\n\nThe consumer's router agent must include `Skill` in its `tools:` frontmatter\nfor `\u002Fdispatch` to be invocable.  Example of correct frontmatter:\n\n```\ntools: Glob, Grep, Read, Edit, Write, Bash, Skill, ToolSearch\n```\n\nSee `docs\u002Fintegration.md` for the end-to-end wiring guide.\n\n## Prerequisites\n\n`claude-wayfinder` must be installed in a Python ≥ 3.11 environment. After\nrunning `\u002Fsetup-wayfinder`, the plugin venv lives at\n`${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002F` — that is the canonical interpreter to use.\n\n**Use the plugin venv's interpreter explicitly.** Do not rely on bare\n`python` resolving to the right environment via `$PATH` — on Windows in\nparticular, a global Python (e.g. `C:\\Python313\\python.exe`) often takes\nprecedence over the venv and does NOT have `claude-wayfinder` installed,\nproducing `No module named claude_wayfinder` at runtime.\n\n| Platform | Path |\n|----------|------|\n| POSIX    | `${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002Fbin\u002Fpython` |\n| Windows  | `${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002FScripts\u002Fpython.exe` |\n\nThe skill's invocations below show the explicit path. Bare `python` is\nfine as a shorthand **only** when the plugin venv is activated in the\ncalling shell, or its `bin\u002FScripts` dir is first on `$PATH`.\n\n```bash\n# Confirm the package is available\n\"${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002FScripts\u002Fpython.exe\" -m claude_wayfinder dispatch --help   # Windows\n\"${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002Fbin\u002Fpython\" -m claude_wayfinder dispatch --help           # POSIX\n```\n\nIf `claude-wayfinder` is not installed yet, run `\u002Fsetup-wayfinder` — that\nskill materializes the venv at the canonical location and pins the\nmatching plugin version into it.\n\n## Canonical catalog path\n\nThe live catalog is at **`~\u002F.claude\u002Fstate\u002Fdispatch-catalog.json`** (or\n`$CLAUDE_HOME\u002Fstate\u002Fdispatch-catalog.json` when `$CLAUDE_HOME` is set).\nThis is the default real-catalog mode resolves to when neither\n`--catalog-path` nor `$DISPATCH_CATALOG_PATH` is supplied. Override only\nfor test fixtures or unusual deployments. The bundled hooks\n(`refresh-catalog-on-stale.js`, `check-catalog-health.js`) use the same\ndefault.\n\n## Running\n\n```bash\nPY=\"${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002FScripts\u002Fpython.exe\"   # Windows\n# PY=\"${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002Fbin\u002Fpython\"          # POSIX\n\n# DISPATCH_SHADOW gates the v3 shadow-route compute (see \"Two-axis labels\"\n# above). The computed route remains telemetry-only unless its domain is\n# also enabled by DISPATCH_HARD_ROUTING_DOMAINS. Sourced from the userConfig\n# `shadow_enabled` toggle in plugin.json so users can opt out without\n# touching source.\n#\n# Only an exact case-insensitive match of \"0\", \"false\", or \"no\" disables\n# shadow compute. Every other value fails open to enabled — this\n# includes \"off\", \"disabled\", an empty string, and any malformed or\n# whitespace-padded value (e.g. \" false \"). Absent (unset) also fails\n# open to enabled.\nexport DISPATCH_SHADOW=\"${user_config.shadow_enabled}\"\n\n# DISPATCH_HARD_ROUTING_DOMAINS is the surgical serving gate: a comma-\n# separated set of domain tokens such as \"is_any\" or \"is_any,code\".\n# Unset, empty, or whitespace-only values fail closed to OFF. Tokens are\n# stripped, lowercased, and deduplicated; unrecognized tokens are dropped\n# with a stderr warning rather than failing dispatch.\nexport DISPATCH_HARD_ROUTING_DOMAINS=\"is_any\"\n\n# The two gates have opposite safe defaults: absent DISPATCH_SHADOW means\n# Compose compute is ON, while absent DISPATCH_HARD_ROUTING_DOMAINS means\n# hard routing is OFF. DISPATCH_HARD_ROUTING_DOMAINS scopes cutover per\n# domain; DISPATCH_SHADOW=0 is the coarse, all-domain kill switch because it\n# skips Compose entirely and therefore always serves the lexical decision.\n\n# Real-catalog mode — default; resolves to the canonical catalog\necho '{\"task_description\": \"implement auth module\", \"file_paths\": [\"src\u002Fauth.py\"], \"agent_mentions\": [], \"tool_mentions\": [], \"command_prefix\": null}' \\\n  | \"$PY\" -m claude_wayfinder dispatch\n\n# Demo mode — opt in with --demo\n\"$PY\" -m claude_wayfinder dispatch --demo\n\n# Explicit catalog override (e.g. test fixture)\nexport DISPATCH_CATALOG_PATH=\u002Fpath\u002Fto\u002Ftest-catalog.json   # POSIX\n# $env:DISPATCH_CATALOG_PATH = \"C:\\path\\to\\test-catalog.json\"  # PowerShell\necho '{...}' | \"$PY\" -m claude_wayfinder dispatch\n```\n\n## Stale-catalog warning\n\nWhen `$DISPATCH_SKILLS_DIR` and\u002For `$DISPATCH_AGENTS_DIR` are set and any\nsource file within them has a modification time newer than the catalog\nfile, the skill emits a warning to stderr:\n\n```\n[DISPATCH WARNING] Catalog mtime is older than source files: ...\nConsider running `claude-wayfinder catalog build` to refresh.\nProceeding with stale catalog.\n```\n\nExecution **proceeds** with the stale catalog — staleness is a\ndegraded-quality signal, not an error.  Run\n`\"$PY\" -m claude_wayfinder catalog build` (using the same explicit\ninterpreter path resolved in the **Running** section above) to refresh.\n\n## The 7 decision branches\n\n| Branch              | When it fires                                                   |\n|---------------------|-----------------------------------------------------------------|\n| `needs_more_detail` | Feature density \u003C 2; provide more context to route accurately. |\n| `delegate`          | One agent scores ≥ 0.85 with a gap ≥ 0.2 above the next.      |\n| `self_handle`       | At least one skill scores ≥ 0.5; no dominant agent.            |\n| `mixed_content`     | Gap \u003C 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[]`. |\n| `advisory`          | Best agent ≥ 0.5. Covers both tie (gap \u003C 0.2, rationale includes `gap=`) and marginal (gap ≥ 0.2 but score \u003C 0.85) cases. Top agent named; alternatives populated. |\n| `ask_user`          | Reserved — not produced by the v0.1 matcher.                   |\n| `self_handle_unaided` | Nothing scores above threshold; proceed without delegation. |\n",{"name":5,"triggers":23,"description":26},{"command_prefixes":24},[25],"\u002Fdispatch","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 (~\u002F.claude\u002Fstate\u002Fdispatch-catalog.json). If neither --demo nor a resolvable catalog is present the skill emits a [CATALOG ERROR] and exits non-zero.\n","main","e906955a5bf962247a87d755ae63495538ce36b5","skills\u002Fdispatch",1,"https:\u002F\u002Fgithub.com\u002Fglitchwerks\u002Fclaude-wayfinder","https:\u002F\u002Fpypi.org\u002Fproject\u002Fclaude-wayfinder\u002F","https:\u002F\u002Fgithub.com\u002Fglitchwerks","2026-11-14 06:56:04",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":47},[37,40,43],{"tool":38,"command":39},"npx","npx skills add glitchwerks\u002Fclaude-wayfinder --skill dispatch",{"tool":41,"command":42},"bunx","bunx skills add glitchwerks\u002Fclaude-wayfinder --skill dispatch",{"tool":44,"command":45},"pnpm","pnpm dlx skills add glitchwerks\u002Fclaude-wayfinder --skill dispatch","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fglitchwerks\u002Fclaude-wayfinder\u002Fblob\u002Fmain\u002Fskills\u002Fdispatch\u002FSKILL.md",true,[],[50,57,64,71,77,84],{"skillNo":51,"name":52,"nameZh":6,"description":53,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":54,"categories":55,"keywords":56,"publishedAt":18},"01M72V35HRZHT8N6NZWAGY4CDS","refactoring-discipline","Behavior-preserving refactoring discipline. Activates when the task involves restructuring, extracting, or renaming existing code without adding new behavior.","https:\u002F\u002Fgithub.com\u002Fglitchwerks\u002Fclaude-wayfinder\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskills\u002Frefactoring-discipline\u002FSKILL.md",[],[],{"skillNo":58,"name":59,"nameZh":6,"description":60,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":18},"01M72V33XS39XGHFYH5RM0HSW3","legacy","Pre-migration skill.","https:\u002F\u002Fgithub.com\u002Fglitchwerks\u002Fclaude-wayfinder\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskills\u002Flegacy\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":18},"01M72V289JXSDJBQK1GW5RCJAP","csv-utils","Helpers for reading and writing CSV files in Python.","https:\u002F\u002Fgithub.com\u002Fglitchwerks\u002Fclaude-wayfinder\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskills\u002Fcsv-utils\u002FSKILL.md",[],[],{"skillNo":72,"name":73,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":18},"01M72V1GJFXNZ8CVZZG8T66E3G","broken","https:\u002F\u002Fgithub.com\u002Fglitchwerks\u002Fclaude-wayfinder\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskills\u002Fbroken\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":18},"01M72V0SWXW4XJE2EWCZWM5EDP","agent-authoring","Expert guidance for writing, editing, and debugging agent and skill frontmatter, CLAUDE.md harness rules, and SKILL.md files.","https:\u002F\u002Fgithub.com\u002Fglitchwerks\u002Fclaude-wayfinder\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskills\u002Fagent-authoring\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":18},"01M72V00NQ401XMPQZV0PH01J3","setup-wayfinder","Materialize the claude-wayfinder Python venv at ${CLAUDE_PLUGIN_DATA}\u002Fvenv\u002F\nand write the setup-state flag. Use when:\n- User types \u002Fsetup-wayfinder\n- User says \"set up claude-wayfinder\", \"install wayfinder dependencies\"\n- User says \"wayfinder isn't working\", \"fix wayfinder\", \"repair wayfinder\"\n- SessionStart banner indicates setup is required and the user wants to proceed\n- Plugin version bumped and re-setup needed\nDo NOT trigger on casual mentions of \"wayfinder\" without setup\u002Finstall\u002Ffix intent.","https:\u002F\u002Fgithub.com\u002Fglitchwerks\u002Fclaude-wayfinder\u002Fblob\u002Fmain\u002Fskills\u002Fsetup-wayfinder\u002FSKILL.md",[],[]]