Lamina Init Agent Skill

/lamina-init

Before any evidence or graph mutation, read and apply

../lamina/orchestrator/prerequisites/cli-required.md.

Read ../lamina/orchestrator/audit-profiles.yaml. Load the init.always

references, then only those init.conditional references whose when signal

is present. For each selected entry, open the named capability SKILL.md and

then its exact reference; do not load sibling topics.

For a directory that is not yet a Git project, initialize Git metadata as

specified by that prerequisite before writing evidence. Do not ask for or

create an initial commit.

Init writes evidence files only under .lamina/ and canonical knowledge through graphd. It never edits application source.

Evidence artifacts

Write .lamina/business-context.md with frontmatter containing lamina.maturity, platform, and last_updated, followed by exactly these non-placeholder sections: Problem statement, Business goals, Success metrics, Scope, Users & market, Product posture, Constraints, Stakeholders, Risks & unknowns, Research posture, Triad check.

Write .lamina/personas.json as evidence-source JSON with evidence-grounded personas. Goals, constraints, and evidence are arrays. Do not invent demographics. These files are indexable evidence, not canonical graph state.

Run the shipped init/persona validators when available, then `lamina graph

observe` so CocoIndex produces explicit source Observation envelopes. The CLI

automatically replaces an incompatible graphd and retries observation once; do

not repeatedly rebuild observations to repair daemon compatibility.

If observation exits nonzero after that recovery, treat observation as

degraded and optional. Core graph initialization may continue, but do not

create Evidence Resources or attach evidence claims for the unavailable

snapshot. Do not describe aliases, observation view names, or generation

labels as evidence.

When observation succeeds, use observed.resource_ids from its output (or the

exact Resource.id values returned by querying that active observation view)

as graph evidence. Never substitute a path alias, source key, view name, or

generation for an Observation Resource id.

Canonical graph

Start one explicit session. Propose:

• one inferred Product Resource proposal grounded in the user's explicit product intent;
• every evidence-grounded Persona as a Persona Resource;
• corresponding Actor Resources when authority/ownership is known;
• lamina:canAssume Statements between Personas and Actors;
• Evidence Resources referencing the relevant source observations.

Publish atomically. Agents must not submit epistemic class or approval. Never cap Personas.

Record the actual returned Resource ids for Product, Personas, Actors, and

Evidence; generated canonical Resources normally use res_* ids. Do not

report input aliases as ids.

All agent-accessible proposal methods use inferred ingress, including claim.add. Never select an epistemic class by choosing a method name. Intended knowledge requires a trusted engine-owned intent ingress; until that ingress supplies it, preserve the user's words as provenance and keep the proposal inferred.

Update mode

Merge changed business evidence, append a dated changelog, rerun observations, and propose new Statements or aliases without replacing stable Resource identity. Preserve conflicting valid facts as Contradictions.

Completion

Report canonical graph initialization and observation as separate outcomes.

Always report the GraphVersion, source revision, actual Product/Persona/Actor

Resource ids, contradictions, and evidence gaps. Install or refresh passive

provider rules with lamina setup --agent <provider>, then state that future

ordinary product requests automatically prepare graph-backed implementation

context, complete design gaps, implement, and verify. Do not recommend another

slash command.

Only report observation coverage and Observation Resource ids when `lamina

graph observe` exited zero and its completion checks passed. After a nonzero

observation command, report observation as degraded/unavailable and explicitly

state that the published GraphVersion has no current observation-backed

evidence. Never claim complete observation coverage on that path.