Dispatch authoring

Frontmatter Authoring Guide

This skill covers everything you need to write, improve, and troubleshoot trigger frontmatter for claude-wayfinder agents and skills. It explains what fields the matcher reads, how scoring works, the seven routing decisions the matcher can produce, and the most common mistakes that cause an entry to match poorly or never match at all.

---

1. What the Matcher Consumes

Source precedence

The catalog generator accepts trigger configuration from two places: inline frontmatter inside the agent or skill file itself, and a sidecar file placed next to the main file. The sidecar always wins when both are present.

The current sidecar format is v6. Under v6:

• A skill stores its trigger configuration in a file named triggers.yml placed in the same directory as SKILL.md:

  skills/python/SKILL.md
  skills/python/triggers.yml   ← authoritative trigger config
  

  Per issue #150, a colocated <name>.triggers.yml next to SKILL.md (owned and project agent sidecars) overrides inline frontmatter. The SKILL.md file itself should not contain triggers:, applicable_agents:, or applicable_skills: keys; the generator emits a warning and ignores them if they appear there.

• A plugin-shipped agent stores its trigger override at triggers/<plugin>/agents/<name>.yml (per issue #142). This file activates a dormant plugin agent that would otherwise score zero.

For the canonical field reference see docs/schema.md. For the design discussion behind the sidecar approach see docs/design/trigger-schema.md.

Schema fields

Field	Where it lives	Purpose
command_prefixes	triggers: block	Slash commands that immediately short-circuit to score 1.0.
agent_mentions	triggers: block	Agent names whose explicit mention in the prompt immediately short-circuits to score 1.0.
path_globs	triggers: block	fnmatch-style globs matched against the file_paths dimension of the input. Each matched glob adds 0.4 to the score.
keywords	triggers: block	List of {term, weight} mappings. Each term found in the input keywords adds 0.5 × weight to the score. Terms are Porter2-stemmed unless no_stem: true is set (see § 1 "Stemming").
no_stem	per-keyword flag	Optional no_stem: true on a {term, weight} mapping exempts that term from stemming — it matches the raw (unstemmed) input token verbatim. Use for acronyms, product names, and CLI flags (aws, gh, ps1).
keyword_groups	triggers: block	AND-group conjunctive triggers (added v0.6.0, #135). List of {slots: [{name, terms: [...]}, ...], weight} groups. A group fires only when all of its slots match — each slot must have ≥ 1 of its terms present in the input keywords. On match, the group adds 1.0 × weight (via _GROUP_MULTIPLIER, distinct from the 0.5 _KEYWORD_MULTIPLIER so a satisfied weight-1.0 group can solo-decide delegate). Use when a routing decision should require co-occurrence of two or more terms (e.g. verb + noun) rather than either alone.
tool_mentions	triggers: block	Tool names matched against the input tool_mentions dimension. Each match adds 0.5.
excludes	triggers: block	Terms that hard-zero the entry's score when found in the input keywords.
applicable_agents	skill sidecar top-level	Hard allowlist of agent names that may receive this skill. ["*"] means any agent. [] means no agent — the skill is dormant.
applicable_skills	agent sidecar top-level	Hard allowlist of skill names to attach when routing to this agent. ["*"] means any. [] means no skills.
routable	agent frontmatter top-level	When false, the agent is excluded from the scored-agent pool. Set to false on the router itself so it is never selected as a delegation target. Absent on skill entries.

How each field type feeds the matcher

command_prefixes and agent_mentions are checked before any additive scoring begins — a match on either field returns 1.0 immediately without examining any other field. excludes is also a pre-scoring check: a match zeroes the entry and stops further evaluation.

path_globs, keywords, and tool_mentions contribute additively to a running score that is capped at 1.0 before being returned. An entry with no triggers in any of these three fields will score 0.0 on any prompt that does not happen to name the entry directly by command prefix or agent mention.

Stemming (symmetric Porter2, issue #304)

The matcher applies symmetric Porter2 (Snowball English) stemming to keyword matching. Catalog terms are stemmed at catalog-build time; input tokens are stemmed at dispatch time; matching is stem-vs-stem. The practical consequence for authors: list the base form only — you do not need to enumerate inflections. implement matches implementing and implemented; refactor matches refactored; lint matches linting.

This applies to flat keywords terms and to keyword_groups slot terms alike. Two distinct catalog terms that collapse to the same stem are deduped at parse time so a single input token cannot double-count.

no_stem opt-out. Add no_stem: true to a {term, weight} mapping to exempt that term from stemming. A no_stem term is matched verbatim against the raw (unstemmed) input token. Use it for tokens that must not collapse — acronyms, product names, and CLI flags:

triggers:
  keywords:
    - { term: "deploy", weight: 1.0 }                 # stemmed: matches "deploying", "deployed"
    - { term: "aws", weight: 1.0, no_stem: true }     # exact: matches "aws" only
    - { term: "gh", weight: 0.5, no_stem: true }      # exact: matches "gh" only

Stemming changes nothing about the scoring formula or thresholds — it only changes which input tokens count as a match. See docs/dispatch-authoring-guide.md for the canonical reference, and § 9 for the --check-stems collision pre-flight.

---

2. The Seven-Decision Ladder

The matcher evaluates the catalog against the input features and emits exactly one of seven decisions. The ladder is evaluated in order; the first branch whose conditions are satisfied wins.

1. needs_more_detail — the input's extracted features populate fewer than two distinct dimensions (paths, keywords, tools, command prefixes, agent mentions); the matcher did not attempt to score the catalog at all.

2. delegate — one routable agent scored ≥ 0.85 and its gap above the second-place agent is ≥ 0.2; high-confidence single winner, delegation is appropriate.

3. self_handle — no dominant agent, but at least one skill scored ≥ 0.5; the router handles the task itself with the matched skills attached.

4. mixed_content — the gap between the top two agents is < 0.2 and ≥ 2 agents are each clamped at 1.0 on path-disjoint lanes (every qualifying agent has at least one path-glob match, and no input path is claimed by more than one agent). The decision carries a lanes[] array (one entry per specialist, each with agent, score, matched_paths, and skills) and an unassigned_paths[] list for input paths not claimed by any lane. Fires after self_handle and before advisory.

5. advisory — the best agent scored ≥ 0.5 but the delegate threshold was not met (gap < 0.2 or score < 0.85) and mixed_content conditions were not satisfied; delegation is suggested but not certain.

6. ask_user — reserved in v0.1 and v0.2; the current matcher never produces this decision. Include a handler for forward compatibility.

7. self_handle_unaided — no agent and no skill scored above threshold; the router proceeds without delegation or skill attachment.

Input-side density floor vs. entry-side weak scoring

These are two separate concerns. Conflating them leads to incorrect diagnosis when an entry fails to match.

Input-side density floor (the needs_more_detail branch above). The matcher emits needs_more_detail when the user prompt's extracted Features populate fewer than two dimensions — paths, keywords, tools, command prefixes, or agent mentions. This is a property of how thin the input is, not of how thin any catalog entry's triggers are. A two-word prompt with no file paths and no recognised keywords triggers this branch regardless of how rich the catalog is. The fix is to provide a richer prompt, not to change the entry's triggers.

Entry-side weak scoring (a calibration footgun — see Section 5). An entry whose triggers populate only one dimension — for example keywords-only with no path_globs, tool_mentions, or command_prefixes — will score weakly on most inputs, because it can only accumulate score when the input happens to mention one of its specific terms. This is not an unreachability theorem; the entry can score if the input fills its one dimension. But it is a calibration smell: the score ceiling on matching inputs is limited (+0.5 × weight per keyword hit, clamped at 1.0), and any prompt that doesn't mention one of the entry's specific terms scores it at zero. Section 5 elaborates the practical guidance.

---

3. Scoring Math

The matcher computes a per-entry score using the following rules, applied in order. Short-circuits fire before any additive contribution is calculated.

Short-circuit rules (evaluated first, in this order):

• command_prefixes match → score = 1.0 immediately.
• agent_mentions match → score = 1.0 immediately.
• excludes match in features.keywords → score = 0.0 immediately.

Additive scoring (when no short-circuit fired):

• Per path_globs match: +0.4
• Per keywords match: +0.5 × weight (verified at src/claude_wayfinder/match.py:84 — _KEYWORD_MULTIPLIER = 0.5; raised from 0.3 to fix single-keyword skills never attaching)
• Per keyword_groups match: +1.0 × weight per group that fires (a group fires only when all of its slots are satisfied; _GROUP_MULTIPLIER = 1.0 is deliberately distinct from _KEYWORD_MULTIPLIER = 0.5 so a satisfied weight-1.0 group can solo-decide delegate; #135 / v0.6.0)
• Per tool_mentions match: +0.5

The final additive score is clamped to 1.0 before being returned.

AND-group worked example

An entry has a keyword_groups block requiring both a "verb" and a "noun" to co-occur:

keyword_groups:
  - slots:
      - {name: verbs, terms: [create, open, file]}
      - {name: nouns, terms: [issue, ticket, bug]}
    weight: 1.0

Prompt	verbs slot	nouns slot	Group fires?	Score from this group
"open the issue tracker"	✓	✓	yes	+1.0 × 1.0 = +1.0
"create the issue body"	✓	✓	yes	+1.0
"open the file"	✓	✗	no	+0.0
"tell me about issues"	✗	✓	no	+0.0

A satisfied weight-1.0 group contributes 1.0 to the entry — enough to clamp at the ceiling on its own and solo-decide delegate if the gap to the runner-up is ≥ 0.2. That is the design intent of _GROUP_MULTIPLIER = 1.0: a group is a stronger signal than any single flat keyword (which maxes at +0.5) because the group already encodes co-occurrence — the matcher has more evidence that the prompt intends this entry, so the score reflects that.

The flat-keywords equivalent of the same term lists would score on either side firing alone — scoring +0.5 × weight on "open the file" and "tell me about issues" too, since each individual term contributes independently. The AND-group structurally requires co-occurrence, which is the right shape when the trigger only makes sense for the combined intent.

Worked example

An entry has one path_globs entry (**/*.py) and one keywords entry (python, weight 1.0). For a prompt that mentions a .py file and the word "python":

path_glob match:   +0.4
keyword match:     +0.5 × 1.0 = +0.5
total:              0.9   (below 1.0, no clamping needed)

A score of 0.9 with a gap ≥ 0.2 above the second-place entry would yield a delegate decision.

Clamping footgun

Because the final score is hard-clamped at 1.0, stacking additional high-weight keywords past the ceiling adds nothing. Consider an entry with one path-glob hit (+0.4) and two weight-1.0 keyword hits (+0.5 + +0.5): the additive total is 1.4, which clamps to 1.0 — exactly the same as one path-glob plus one weight-1.0 keyword. The second high-weight keyword is dead weight on any input that already crosses the ceiling.

The practical guidance: once an entry can plausibly reach ≥ 1.0 on its highest-signal inputs, prefer broadening coverage — add more distinct terms at 0.25 or 0.5 weight — over stacking duplicate 1.0 weights. Broadening raises the score on a wider range of inputs; stacking only inflates the sum on the inputs where the ceiling would already be hit, and the clamp throws that extra score away.

---

4. Trigger Field Rules

These are the validation rules the catalog generator enforces. Violating them produces a warning at build time and may cause the entry to score unexpectedly.

• Weight ladder is exactly {0.25, 0.5, 1.0}. Any other numeric value is clamped to the nearest ladder step with a validator warning (see _clamp_weight in build_catalog.py). There is no weight of 0.75, 0.3, or 2.0.

• keywords is a list of {term, weight} mappings. Bare strings are rejected by the generator. Every keyword entry must be an object with exactly two keys: term (a string) and weight (one of 0.25, 0.5, 1.0).

• Keyword terms are Porter2-stemmed unless no_stem: true. List base forms; stemming matches inflections automatically (see § 1 "Stemming"). Set no_stem: true on terms that must match verbatim — acronyms, product names, CLI flags. A no_stem term is compared against the raw input token, not its stem. Authoring inflected variants of a stemmed term (deploy, deploying, deployed) is redundant — they all collapse to one stem and are deduped.

• path_globs uses Python fnmatch semantics, not gitignore semantics. The matcher calls fnmatch.fnmatch(path, glob). This has important consequences — see the footguns section for the most common mistake.

• tool_mentions is case-sensitive. The matcher compares tool names as literal strings. Bash and bash are different values; the correct casing matches what the Claude Code harness uses. Wrong case silently fails to match.

• excludes matches against features.keywords only. The excludes list is not checked against file_paths, tool_mentions, or agent_mentions. An exclude term that appears only in a file path will not zero the score.

• command_prefixes should start with /. A prefix like dispatch without a leading slash will not match a user-typed /dispatch command, because the dispatcher passes the slash as part of the string.

• keyword_groups requires ≥ 2 slots per group. A single-slot group is semantically equivalent to a flat keywords entry — the catalog generator warns and drops single-slot groups. Use plain keywords instead when you only want one set of alternatives. Each slot needs a non-empty terms list; empty terms makes the slot unsatisfiable and drops the group.

• keyword_groups weights follow the same {0.25, 0.5, 1.0} ladder as flat keywords. Off-ladder weights on a group are clamped with a validator warning, same as on flat keyword entries.

---

5. Footguns

These are the most common authoring mistakes, in roughly descending order of frequency.

fnmatch *.py does not match nested files. Python's fnmatch matches only within a single path component when the glob contains no path separator. *.py matches foo.py but not src/foo.py. Use **/*.py if you mean "any .py file anywhere under the tree." This is the most common path-glob mistake and the one most likely to cause silent non-matching on real inputs.

Tool names are case-sensitive. The harness passes tool names with the casing it uses internally: Bash, Read, Edit, WebFetch, Glob. Lowercase variants like bash or webfetch will silently fail to match. When in doubt, check the tool name in the harness output rather than guessing the casing.

applicable_skills: [] mutes the agent's skill attachment entirely. Setting this to an empty list means the agent will never have any skill attached, regardless of how well those skills score against the input. Only set [] when you genuinely want no skills auto-attached to this agent. The most common unintended form of this is inheriting a default empty list in a new agent sidecar template and forgetting to change it.

One-dimensional triggers are a calibration footgun (entry-side). A routable agent or skill with only keywords and no path_globs, tool_mentions, or command_prefixes will score zero on any prompt that does not mention at least one of its specific keyword terms. Even on matching prompts, the score is bounded by +0.5 × weight per hit, clamped at 1.0. This is not the same as the input-side needs_more_detail floor described in Section 2, which fires when the user's prompt is too thin. This is about the entry being weakly reachable across the distribution of prompts the matcher actually sees. The fix is to pair keywords with at least one path_globs, tool_mentions, or command_prefixes entry, giving the matcher a second scoring dimension to work with.

Flat keywords over-fire when terms only make sense together. If your entry should match "open issue" but not "open file" or "recent issues" alone, flat keywords score each term independently and accumulate on either side firing alone. Use keyword_groups to require co-occurrence: each slot carries one set of alternatives, and the group fires only when all slots are satisfied. See the §3 worked example.

Conflict pairs produce advisory decisions. Two entries whose keywords lists share three or more overlapping case-insensitive terms, with no discriminating path_globs, tool_mentions, or command_prefixes to break the tie, will both score similarly on inputs that mention those shared terms. The gap falls below 0.2, so neither reaches the delegate floor, and the matcher emits advisory — the top agent is suggested but not confirmed. (Prior to v0.9.0 this scenario produced an ambiguous decision; that outcome was removed. The runtime now emits advisory for all gap-tied and below-threshold cases.) Heavy keyword overlap is a design smell. The remedy is to introduce a discriminator: a path_globs entry that is unique to one of the two entries, a tool_mentions entry that only one of them legitimately fires on, or a command_prefixes entry that explicitly routes one of them. If the overlap is fundamental — the two entries genuinely do the same thing in the same context — consider whether they should be merged into one.

Stemming can make two distinct terms collide silently. Because matching is stem-vs-stem, two terms you intended to be different can collapse to the same Porter2 stem and fire as one — universe and university both stem to univers, operating and operate both stem to oper. When that collision spans two competing catalog entries, it manufactures a conflict pair you did not author by hand. The fix is either to accept the collision (often it is harmless) or to add no_stem: true to one of the terms so it matches verbatim. Run python -m claude_wayfinder catalog build --check-stems (see § 9) to surface cross-entry stem collisions as STEM_COLLISION warnings before they reach the live catalog. Acronyms and product names are the most common over-stem casualties — prefer no_stem: true on those by default.

---

6. Authoring Workflow

Use this workflow when writing trigger frontmatter for a new agent or skill from scratch.

Step 1 — Read the body in full. Open the agent or skill file and read it completely before writing a single trigger. Triggers that are drafted from a one-line summary routinely miss the recurring terminology the body actually uses.

Step 2 — Identify prominent terms and assign weights. As you read, note terms by how central they are to the entry's purpose:

• The skill or agent name, or the core verb it acts on (e.g. refactor for the refactoring agent, dispatch for the routing skill) → weight 1.0. At most one or two terms should sit here; if many terms seem equally central, that is a sign the scope is too broad.
• Recurring concept terms — vocabulary the body uses repeatedly and that distinguishes this entry from neighboring ones → weight 0.5.
• Supporting or contextual terms — words that appear in the body and hint at the use-case, but are not discriminating on their own → weight 0.25.

A well-calibrated entry typically has one or two 1.0 terms, three to five 0.5 terms, and a handful of 0.25 terms. If the 1.0 bucket is full of generic words (code, file, run) the entry will conflict with half the catalog.

List base forms only — terms are Porter2-stemmed, so implement already covers implementing and implemented (see § 1 "Stemming"). Mark acronyms, product names, and CLI flags with no_stem: true so they are not over-stemmed into a collision with another entry's term.

Step 3 — Add path_globs for any file patterns the body implies. If the body directs attention to specific file types or directory trees, add a path_globs entry for each. Always use **/*.ext for extension-based patterns (not *.ext — see the fnmatch footgun in Section 5). A path_globs entry adds +0.4 per match and provides a second scoring dimension that significantly improves disambiguation against other entries with similar keyword sets.

Step 4 — Add tool_mentions for any tools the body explicitly names. If the body tells the user to reach for a specific tool — Bash, Edit, WebFetch, etc. — list those tools in tool_mentions. Use the exact casing the harness uses (capitalize the first letter; see the case-sensitive footgun in Section 5).

Step 5 — Decide applicable_skills (for agents). Read the body for skill-task language: verbs like "plan", "test", "review", "debug". List the skill names whose purpose aligns with those verbs in applicable_skills. Use ["*"] only if the agent is genuinely purpose- agnostic. Do not leave the field blank or set it to [] unless you intentionally want no skills attached.

Step 6 — Prefer the v6 sidecar location over inline frontmatter. Place the resulting YAML in the sidecar rather than embedding it in the agent or skill file's frontmatter block. Sidecars isolate trigger config from body copy, which makes diffs cleaner and code review faster. The correct paths are:

• Owned skills: skills/<name>/triggers.yml
• Plugin agents: triggers/<plugin>/agents/<name>.yml

Inline frontmatter is still read by the generator, but a sidecar at the location above always overrides it.

---

7. Tuning Workflow

Use this workflow when improving trigger frontmatter that already exists but is producing poor match results.

Step 1 — Read the body and the current triggers side-by-side. Open both the sidecar (or inline frontmatter) and the agent or skill body at the same time. This side-by-side read is the only reliable way to spot divergence between what the entry does and what the triggers say it does.

Step 2 — Find stale keywords. Look for terms in triggers.keywords that no longer appear in the body. Bodies change over time; triggers often do not. A stale keyword raises the entry's score on inputs that no longer reflect its actual purpose, creating misleading matches. Remove or replace stale terms.

Step 3 — Find missing keywords. Scan the body for recurring terms that are absent from triggers.keywords. If a concept appears in every paragraph but is not listed as a keyword, the entry will miss prompts that use that concept. Add the term at the weight level appropriate to how central it is.

Step 4 — Check weight alignment. For each existing keyword, ask whether its weight still reflects its centrality to the body. The most common drift pattern is a term that was elevated to 1.0 during an early iteration and was never revisited as the scope of the entry narrowed. A 1.0 weight that should be 0.5 inflates the score on prompts that mention that term even tangentially and widens conflict-pair risk.

Step 5 — Check conflict-pair risk. Eyeball the catalog for other entries that share several of the same keywords. If two or more entries have three or more overlapping terms at 0.5 or 1.0 weight and no differentiating path_globs or tool_mentions, they will produce advisory decisions on the prompts where those terms overlap (the gap falls below 0.2, preventing delegate). Introduce a discriminator (see Section 5) or run audit-catalog (see Section 9) to surface all conflict pairs at once.

Step 6 — Check for structural violations. Before committing, verify that:

• Every keywords entry is a {term, weight} mapping, not a bare string.
• Every weight is exactly one of 0.25, 0.5, or 1.0.
• No term contains leading or trailing whitespace (the generator does not strip these; " python" and "python" are different terms).
• command_prefixes entries each start with /.
• Acronyms, product names, and CLI flags carry no_stem: true so they are not over-stemmed; run python -m claude_wayfinder catalog build --check-stems to confirm no unintended cross-entry stem collisions were introduced.

---

8. Troubleshooting Workflow

When an agent is not being dispatched or a skill is not attaching as expected, work through the symptom table below to identify the cause.

Symptom	Likely cause
Routable agent scores 0 on prompts that should match	Unreachable routable: triggers is empty or every keyword weight is 0.
Score never crosses the delegation floor (0.85)	One-dimensional triggers — the entry has only keywords and no path_globs or tool_mentions; max reachable score is limited.
Agent matches everything indiscriminately	Keyword set too generic (code, file, run); conflict-pair risk against many other entries.
Skill never attaches to the expected agent	applicable_agents on the skill sidecar excludes that agent name; or applicable_skills: [] on the agent sidecar mutes all skill attachment.
Weight you set in the sidecar is not what the matcher uses	Non-ladder weight (e.g. 0.75) was silently clamped to the nearest step; check catalog-generation.log.
A specific term never contributes to the score	The term appears in the entry's own excludes list, zeroing the entry whenever it is present in the input.
A tool_mentions entry never matches	Case mismatch — the harness passes Bash, not bash; WebFetch, not webfetch. Use the exact harness casing.
A term matches inputs that share only a word stem (university firing on universe)	Porter2 over-stemming collapsed two distinct terms to one stem. Add no_stem: true to the term, or run catalog build --check-stems to find the collision (§ 1, § 5, § 9).
An acronym / product name matches unexpected inflections	The token was stemmed when it should be literal. Add no_stem: true to match it verbatim.

Diagnostic sequence when none of the above is obvious:

1. Run audit-catalog (see Section 9) and check the output for structural warnings on the entry.
2. Compare the entry's trigger YAML against the schema in docs/schema.md field by field.
3. Check whether a sidecar overrides the inline frontmatter you edited — both exist, the sidecar wins, and the edit may have gone to the wrong file.
4. Confirm the entry is marked routable: true (or that the key is absent, which defaults to true). An explicit routable: false removes the entry from the scored-agent pool entirely.

---

9. When to Run the CLI

The matcher-aware checks the LLM cannot do consistently across all ~70 catalog entries — conflict-pair detection, unreachable-routable scans, structural validation across the whole catalog — live in python -m claude_wayfinder audit-catalog. Run it whenever you add or substantially edit a routable agent, before opening a PR that ships new frontmatter, or as a periodic catalog sanity check. See docs/dispatch-authoring-guide.md for the rule reference and exit-code contract.

The CLI is the authoritative source for catalog-wide problems. It is not a substitute for the field-by-field checks in Sections 6 and 7, but it catches conflict pairs and unreachable entries that would require reading the full catalog by hand to detect otherwise.

Stem-collision pre-flight. Run python -m claude_wayfinder catalog build --check-stems whenever you add or edit keyword terms. It detects pairs of distinct keyword terms — across different catalog entries — that share the same Porter2 stem, printing each to stderr as a STEM_COLLISION line. Review each: accept the collision if it is harmless, or add no_stem: true to one of the terms to disambiguate (see the § 5 stemming footgun). This check is distinct from audit-catalog's conflict-pair detection — a stem collision can manufacture a conflict pair that the hand-authored keyword lists do not reveal on inspection.

---

10. References

• docs/schema.md — canonical trigger field reference; start here when looking up the exact name, type, or default for any trigger field.
• docs/design/trigger-schema.md — design rationale for the schema; explains why certain field shapes were chosen and what alternatives were considered.
• docs/dispatch-authoring-guide.md — extended worked-examples companion to this skill; covers edge cases and advanced calibration patterns not addressed above.
• agent-authoring skill (in ~/.claude/skills/agent-authoring/) — broader harness authoring discipline covering agent structure, routing configuration, and the full lifecycle of a new agent; this dispatch-authoring skill is its matcher-specific counterpart.