Get the override model for a provider, or undefined if none set.

* Sort by severity (error > warning > info), then by pass order (the * canonical REVISION_PASSES ordering), so the most urgent, most structural

* Set (or clear) a provider's model override and persist. * Passing an empty/whitespace model clears the override (revert to default).

* Select the best provider for a given task type using tiered routing. * Priority: per-project override → global preference → tier routing. * Wh

Write a fixture SKILL.md at skills/<category>/<name>/SKILL.md.

* Send completion request to the selected provider. * Tracks system prompt cache hits to estimate token savings.

* Analyze a sample of the author's prose and build a style profile.

Run the deterministic mechanical screen. Cheap; safe to run on every draft.

Write a {projectId}.json fixture into workspace/context and load it.

* Diff a chapter against the entity DB + prior summaries and return * categorized, evidence-chained contradictions. * * Never throws on AI tr

* Persist all project state to disk (debounced — max once per second). * Non-fatal: if save fails, projects continue to work in-memory.

* Search the ARCHIVAL tier (FTS5 BM25 over conversations + manuscripts + * project steps) and return a labeled, budgeted excerpt block. * * F

* A 3-chapter project: * - Aria: protagonist, recurring across all 3 chapters (baseline top-K). * - Kael: recurring, named in the active chapter

ISO timestamp `days` before NOW.

* Escape special XML characters.

A recording aiSelectProvider — captures the task types requested.

Directly seed a preference with a chosen source + updatedAt (bypassing set()'s clock).

* Assemble the CORE block for a project, budgeted per the design's priority * table and hard-clamped to CORE_BUDGETS.total chars. * * Pure re

* A stubbed aiComplete that returns a fixed body, and records the request so * tests can assert on the provider/task routing.

* Match skills against user input, ranked by match quality, capped at the * top MAX_MATCHED_SKILLS, with a total injected-content budget so skill

Structured stats for dashboard Morning Briefing

* Returns the cached chapter summaries for a project (chapter-sorted, as * generateSummary keeps them). Pure in-memory read — never AI-calls, never

Edge TTS is always available (only needs internet); ElevenLabs needs an API key.

* Combined evaluation. Returns a verdict including whether to retry. * * Default scoring weights: * - mechanical 30%, judge 70% (judge catc

* Get autonomous mode status for dashboard/API

* Returns cached entities filtered by type. Pure in-memory read over cached * ProjectContext — never AI-calls, never throws. Returns [] if nothing

* Read the materialized core digest. Never throws — returns null on missing * file or malformed JSON so the hot path degrades to live reads.

* Create a new project from a template or custom definition. * Returns the project with auto-planned steps.

* Compute the promoted set of character sheets for CORE. * * PROMOTE rule: a character is promoted if its name/alias appears * (case-insensit

Run a pass body, converting throws into a failed PassOutcome (never rethrows).

* Core message handler — processes input from any channel. * Optional extraContext is appended to the system prompt (used by goal engine).

* Build a minimal fake ServiceContainer covering exactly the members * handleMessage() touches on the happy path. `overrides` lets a test swap in a

* Cron entry point (CronHandler shape). Consolidates one project (when * payload.projectId is given) or every enumerable project. * * Never r

Run one chat turn through the pipeline and return the assembled system prompt.

* Write the materialized core digest atomically (tmp + rename). Best-effort: * ensures the context dir exists first. Throws only on a genuine write

Get the persona currently tagging new conversation turns.

* Infer a preference key from a value phrase.

* Infer the best project type from a natural language description. * Used when the user just says what they want without specifying a type.

* Sleep-job maintenance pass (Tiered Memory Chunk C, step 6). * * Removes low-value inferred/observed preferences that have gone stale, plus

* Resolve the active model for a provider using the precedence: * model-config.json override → this.config.<provider>.model → hardcoded default.

* Set (or clear) a provider's model override, persist it, and reinitialize * so the change takes effect without a restart. Passing an empty model