agentfootprint/dist/esm/injection-engine/index.js

The explainable agent framework — backtrack a wrong answer to the exact context that caused it (evidence, not guesses). Built on footprintjs.

/**
 * agentfootprint Injection Engine — public barrel.
 *
 * The unifying primitive of agentfootprint context engineering.
 * One `Injection` type. One `InjectionEngine` subflow. N typed sugar
 * factories. See `README.md` in this folder for the full concept.
 */
// Bind the pure core's dev-warn seam to footprintjs's `isDevMode` (9.34.0).
// Side-effect import, on purpose: the skill-graph core no longer imports the
// engine, so SOMEONE has to hand it the dev flag, and this barrel is on every
// path inside this package that can reach `skillGraph()`. Without it,
// `enableDevMode()` would stop switching skill-graph warnings on — the one
// behavior the fence refactor must not change. `agentfootprint/skill-graph`
// deliberately skips it; see `devWarn.ts`.
import './devWarnHost.js';
// POJO projection — used by slot subflows + advanced consumers
export { projectActiveInjection } from './types.js';
// The iteration's tool-result batch with the singular fallback applied
// (9.16.0) — the one reader `rule` predicates should use to see every call of
// a parallel batch, not only the last (`ctx.toolResults ?? [ctx.lastToolResult]`).
export { toolResultsOf } from './types.js';
// The messages a PERSON wrote (9.84.0) — the one reader for a predicate that
// judges what was said. `ctx.history` also holds the five user-role messages
// this library writes itself (compaction frame, drop notice, schema-check
// correction, evidence-check correction, delivered message), and a rule
// scanning it raw matches our own bookkeeping.
export { saidByPerson } from './types.js';
// Engine
export { evaluateInjections } from './evaluator.js';
export { buildInjectionEngineSubflow, } from './buildInjectionEngineSubflow.js';
// Sugar factories — Ships four; more flavors planned (RAG / Memory / Guardrail)
export { defineInstruction } from './factories/defineInstruction.js';
export { defineRelevanceHint, READS_ENTRY_SCORES_METADATA_KEY, } from './factories/defineRelevanceHint.js';
export { defineMenuHint, MENU_HINT_METADATA_KEY, } from './factories/defineMenuHint.js';
export { defineSkill, resolveSurfaceMode, } from './factories/defineSkill.js';
// Steps-as-data (9.18.0) — the procedure grammar. `skillSteps.ts` is the ONE
// owner of the types, the pointer's shape and every sentence the model reads;
// the public slice here is what a consumer needs to DECLARE steps, READ the
// pointer off a snapshot (`pointerOf(sharedState.stepPointer)`), and compose
// the integrity tool into custom wiring (the buildReadSkillTool precedent).
export { pointerOf, SKIP_STEP_TOOL_NAME, skipStepDescriptor, } from './skillSteps.js';
export { defineStepsHint, STEPS_HINT_METADATA_KEY, } from './factories/defineStepsHint.js';
// Artifact vocabularies (9.25.0) — `produces`/`consumes` on a skill or a step.
// The public slice is what a consumer needs to READ a skill's declared data
// legs off its metadata (`vocabularyOf`, for a lens or an inventory) and to run
// the satisfiability check itself over a list of skills the library did not
// assemble. `assertArtifactVocabulary` stays private: it is the authoring-time
// refusal `defineSkill` already applies.
export { checkArtifactVocabularies, vocabularyOf, } from './skillVocabulary.js';
export { SkillRegistry } from './SkillRegistry.js';
// File-authored skills — a loader over `defineSkill`, not a second mechanism.
// Node-only (reads the filesystem); node:fs is imported lazily inside the call
// so this barrel stays safe to import from a browser bundle.
// `runbookFromDir` is the same loader read as a whole runbook — skills PLUS the
// edges each file's `routes:` declares, ready for `skillGraph({ ...runbook })`.
export { skillsFromDir, runbookFromDir, } from './skillsFromDir.js';
// Skill-tool builders — used by SkillRegistry.toTools() and the Agent's
// auto-attach path. Exported so consumers building custom tool wiring
// (e.g., gatedTools chains) can compose the same `list_skills` /
// `read_skill` tools directly.
export { buildListSkillsTool, buildReadSkillTool, buildSkipStepTool, } from './skillTools.js';
// The graph's own DESCRIPTIONS of the tools it needs (9.34.0) — the POJOs the
// builders above wrap. Exported because a host composing its own tool objects
// (another framework, a custom ToolProvider) should read `read_skill`'s enum,
// reachability offer and result sentences from the library rather than
// re-writing them. Also the whole surface of `agentfootprint/skill-graph`.
export { listSkillsDescriptor, readSkillDescriptor } from './skillToolDescriptors.js';
export { TOOL_RESULT_STATUSES } from '../semantics/toolOutcome.js';
export { useSkillGraphDevMode } from './devWarn.js';
export { defineSteering } from './factories/defineSteering.js';
export { defineFact } from './factories/defineFact.js';
// Unified factory — a `type` discriminant routes to the four named factories
// above. Use when the flavor is chosen programmatically; prefer the named
// factories when you know the flavor at author time.
export { defineInjection, } from './factories/defineInjection.js';
// Declarative skill graph (proposal 002) — declare skills + routing edges →
// graph-derived triggers + a drawable topology. Sugar over the trigger model.
export { skillGraph, 
// The official vocabulary (9.51.0): you declare the SkillMap; the agent is
// the SkillWalker; the recording carries both. `defineSkillMap` is a
// permanent reference-equal alias of `skillGraph`, `SkillMap` of
// `SkillGraph` — both names ship forever, neither is a rename.
defineSkillMap, decideSkill, 
// 8.7.0 — render a `graph.checkup()` for a log line or a CI failure. The formatter
// the library itself uses; without it every consumer wrote their own. `checkupGraph`
// stays private: its input is the graph's internal wiring shape, and `graph.checkup()`
// is already the door to it.
formatCheckup, SKILL_GRAPH_METADATA_KEY, 
// Guards-as-data on route edges (9.51.0) — the SkillWalker's guard mover as
// data: `guard:` beside the code `when`, compiled once into the predicate,
// the serializable `SkillGuardData` (edges/provenance/`skill.graph_declared`)
// and the per-condition evidence (`cursorMove.guard`/`guardsClosed`).
GUARD_HOP_KEYS, plainGuardCaption, 
// The graph's note that its body-contract checks wait for Agent build (present
// only when built without `knownTools`) — see SkillGraph.deferredBodyContract.
// The note rides the graph AND each compiled skill's metadata (under
// SKILL_GRAPH_DEFERRED_CONTRACT_KEY), so Agent build honors it whichever way
// the skills arrive — `.skillGraph(graph)` or `.skills({ list: () => graph.skills })`.
SKILL_GRAPH_DEFERRED_CONTRACT_KEY, 
// The ONE owner of "is this target the cursor?" (9.86.0) — every gate,
// description and refusal switches on it instead of re-deriving it.
classifySkillTarget, } from './skillGraph.js';
export { keywordScorer, embeddingScorer, rankEntries, } from './entryScorer.js';
export { classifierScorer, CLASSIFIER_SCORER_QUESTION, CLASSIFIER_SCORER_UNAVAILABLE, } from './classifierScorer.js';
// The turn-start routing cascade (SG-C, 9.17.0) — the intent-scorer port +
// the built-in LLM classifier, the tie policy that judges tier 2, and the
// graph surfaces the agent consumes. keywordScorer/embeddingScorer above ARE
// IntentScorers too (same factories, a second arity).
export { validateIntentScores, } from './intentScorer.js';
export { llmClassifier } from './llmClassifier.js';
// The constrained-enum machinery behind the classifier AND the tier-3
// decider (9.19.0) — one model call that can only answer from a fixed list.
export { constrainedEnumPick, } from './constrainedEnumPick.js';
export { NEAR_TIE_MARGIN, MENU_SIZE, DEFAULT_ROUTING_POLICY, decideTier2, menuOutstanding, } from './routingPolicy.js';
export { checkSkillContract, checkSkillContracts, skillToolNames, } from './skillContract.js';
//# sourceMappingURL=index.js.map