Praxec

Cognitive verbs and subject namespaces

The closed enum of 10 verbs, the 15 blessed subject roots, and how strict_namespacing controls enforcement.

Why a closed list instead of free-form strings? Because a fixed vocabulary is what keeps your agent’s behavior consistent and auditable: the same verb means the same thing in every workflow, and a typo or an invented verb fails at config load instead of silently mis-routing a step.

Every skill in praxec carries a verb (one of ten) and a subject (a dotted namespace starting with a blessed root). These aren’t free-form strings. They’re a closed vocabulary, enforced at config load. This page lists the verbs, the blessed roots, and the failure modes.

The ten verbs

verb is a closed enum — ten cognitive operations defined in SPEC §5.4.1. Unknown verbs fail config load with INVALID_VERB. There is no escape hatch, no Other(String), no opt-in extension. Adding a verb requires a spec amendment (SPEC §23.7 sets the criterion).

Verb Cognitive posture
triage Classify, prioritize, route. The “what kind of problem is this and who handles it” call.
diagnose Find root cause. Not fix — find. Symptom-to-cause work.
plan Design the approach before acting. ADRs, RFCs, sprint breakdowns.
implement Produce or generate the artifact. Code, config, content, the actual thing.
review Evaluate against criteria. Adversarial reads, checklists, rubrics.
refactor Restructure while preserving behavior. Same outputs, better shape.
explain Build understanding. Self-explain to think, or teach others.
compose Assemble parts into a whole. Integration, glue, end-to-end wiring.
research (v0.3) Gather context from sources — web, local, docs, prior decisions. Open-ended, before you know what you’re looking for.
summarize (v0.3) Condense. Compress what you already understand into a tighter form for the next reader.

These are cognitive postures, not methodologies. TDD, spec-driven, design-by-contract — those are workflow shapes that sequence the ten verbs. Modifiers like “speedrun” or “code-golf” belong in the body of a skill, not in the verb metadata.

Picking the right verb

The closest-neighbor confusions worth disambiguating:

When you’re tempted to use… Use this instead if…
diagnose …the work is open-ended context-gathering (“what do we know about X?”), not a “why is X broken?” investigation → research
triage …you’re researching to inform the classification (not already classifying) → research
explain …you’re compressing rather than expanding → summarize
compose …you’re assembling existing artifacts into a deliverable (compose), not just summarizing what was done (summarize) — both can be right depending on what the output is

The model reads a skill ref as "{verb} {subject}" — review review.style.house-voice — and fetches the body with praxec.query({subject}) only if relevant.

Why a closed enum?

Closed enums let praxec.query({kind:"skill", query:"review"}) return an exact, deterministic set. Open vocabulary would force fuzzy matching (“is evaluate the same as review?”) and synonym tables, and the synonym table is itself a config the operator has to maintain. Closed forces authors to pick: this skill is a review, not an evaluate. The discipline is the point.

It also gives the model a stable cognitive vocabulary across every architecture it encounters. The same ten verbs, everywhere. No re-learning per project.

Subject namespaces

subject is a dotted namespace. Format: lowercase kebab-case, at least two segments, no whitespace. The schema pattern is ^[a-z][a-z0-9-]+(\.[a-z][a-z0-9-]+)+$. An empty subject is rejected with EMPTY_SUBJECT.

The first segment must be a blessed root. Segments below the root are free-form — name them however makes sense for your team.

Examples of well-formed subjects:

The 15 blessed roots, from BLESSED_SUBJECT_ROOTS in crates/praxec-core/src/discovery.rs:

Domain-themed roots (group guidance by topic, regardless of which verb fits):

Verb-mirror roots (group guidance by cognitive operation — one per verb):

plan and review appear in both categories; they’re listed once each. Total: 15 roots.

The dual structure lets you organize either way. review.code.adversarial sits in the verb-mirror tree (everything that primes a review posture). debug.reproduction sits in the domain tree (everything about debugging, whatever verb it uses). Pick whichever clusters better for your library.

strict_namespacing

The strict_namespacing flag, declared under the top-level praxec: key (SPEC §5.4.2), controls what happens when a subject’s first segment isn’t a blessed root.

praxec:
  strict_namespacing: true   # default

The default is strict because the blessed list exists for a reason: it keeps subjects clusterable in search, predictable in audit, and grep-able across architectures. Loose mode is a migration aid — flip it false while you’re ingesting external skill libraries, then flip it back once you’ve renamed.

Error codes

The codes you’ll see in check output or at config-load time, from SPEC §13:

Code When it fires
INVALID_VERB verb field is not one of the ten closed verbs. Payload includes allowed.
INVALID_SUBJECT_ROOT First segment of subject is not blessed. Raised when strict_namespacing: true. Payload includes blessed_roots.
EMPTY_SUBJECT subject string is empty after trim.
UNBLESSED_SUBJECT_ROOT Same condition as INVALID_SUBJECT_ROOT — an unblessed first segment — but raised by the check-lint / structural-analysis path, always at warning severity. (Note: when config-load runs in lenient mode (strict_namespacing: false), it reuses the code INVALID_SUBJECT_ROOT at warn severity rather than this one.)
HASH_MISMATCH The fragment’s stored hash doesn’t match the recomputed hash of the normalized body. Payload includes stored and computed.
VERB_MAPPED Info-level. An ingested external skill used a verb synonym (e.g. fix → implement); the mapping fired cleanly.
INGEST_INVALID_VERB An ingested external skill used a verb that’s neither in the closed ten nor in the synonym table. The skill is rejected.