Register `constructor` under `name`. Explicit registration wins over any later call with the same name. Safe to call from `main` before processing be

Confirm detected entities with the user, then persist to the global registry and `entities.json`. Calls [`confirm_entities`] (auto or interactive), t

Append `/mempalace.yaml`, `/entities.json`, and `/corpus_origin.json` (anchored to repo root) to the `.gitignore` in `directory` if the directory is t

Validate drawer/diary content. Returns `Ok(trimmed)` if valid, or `Err(error_json)` if not.

Persist the registry data to disk atomically. Writes JSON to a sibling temp file, fsyncs it, renames over the target, then fsyncs the parent director

Process a single mega-file: split it into per-session files and return the number written.

Return aggregated palace statistics.

Sweep a single Claude Code `.jsonl` file into the palace. Parses every user/assistant message, skips those already stored (identified by UUID), and i

Query all knowledge-graph facts for an entity, with optional direction and as-of filters.

List drawers with optional wing/room filtering and cursor-style pagination.

Try to parse a Gemini CLI JSONL session into transcript text. Returns `None` when the content lacks a `session_metadata` entry, fewer than 2 messages

Extract concatenated text from an Anthropic Messages API response. The response `content` array may contain multiple text blocks. This function joins

Build a system-prompt preamble from a corpus-origin detection result. Returns an empty string when the origin adds no useful context (no platform, no

Called by `check_text` to detect mentioned names close in edit distance to other registry names. Only O(m × n) comparisons are made, where m is the n

Compute Jaccard similarity between two word sets. Returns 0.0 when both sets are empty, and 1.0 when sets are identical.

Classify conversation content into the best-matching topic room. Scans the first 3000 characters for keyword matches across the `TOPIC_KEYWORDS` tabl

Tier 2 LLM-assisted detection — confirms platform and extracts persona names. Passes up to [`MAX_SAMPLES_FOR_LLM`] excerpts to `provider.classify` wi

Find lines where true new sessions begin (Claude Code v header not followed by context restore). Returns an empty Vec for empty input — callers treat

Read `topics_by_wing` from the registry. Returns an empty map when the registry is absent, unreadable, or lacks a valid `topics_by_wing` object. Wing

Return `true` if a previous background mine process named by `pid_file_name` is still running. Each mine target writes its own pid file (see [`PID_FI

Open (or create and append to) `hook.log` in `state_dir`.

Strip all characters except alphanumeric, dash, and underscore. Prevents path traversal when the session ID is used as a filename stem.

Layer 1: Essential story — top drawers grouped by room.

Process detected entity candidates and insert high-confidence people. Helper for [`EntityRegistry::learn_from_text`]. Returns names of newly inserted

Append any `include_ignored_paths` entries not already in `scanned`. Called by [`mine`] after the gitignore-aware scan to honour the `--include-ignor

Print the three mode choices and read the user's selection. Returns one of `"work"`, `"personal"`, or `"combo"`. Loops until valid.

Read one person from stdin (name + optional nickname, relationship, context). Returns a `PersonEntry` with an empty `name` when the user presses ente

Read an optional string argument from the MCP tool args, treating empty or whitespace-only values as absent. Mirrors the Python pattern `value if val

Resolve the adapter name per RFC 002 §3.3 priority order. 1. Explicit `--source` flag (`explicit`) 2. Per-source config value (`config_value`) 3. Def

Run `mempalace init` for `directory`. Discovers entities, optionally refines via LLM, confirms with the user, detects rooms, then writes `entities.js

Print the init summary including detected entities and rooms. Called by [`run`] to keep that function within the 70-line limit.

Dispatch a hook from a pre-parsed JSON payload. Extracted from `run()` so tests can inject JSON without requiring stdin. Returns `Err` for unknown ho

Knowledge graph stats.

Collect all `.jsonl` files under `dir_path` up to `WALK_DEPTH_LIMIT` deep. Extracted from `sweep_directory` to keep that function within the 70-line l

Update `report` and the removable lists for one classification result. Extracted from `run` so the main loop stays under the 70-line ceiling.

Union the components containing `a` and `b`.

Remove an adapter registration (primarily for tests).

Called by `check_text_kg_contradictions` to extract structured claims from text. Supports two surface forms — `"X is Y's Z"` and `"X's Z is Y"` — bot

Split content into chunks by paragraph (blank-line boundary) or line groups. Falls back to 25-line groups when the content has only one paragraph but

Classify `name` into `"person"`, `"project"`, or `"uncertain"` based on its signal scores and occurrence frequency.

Find and remove near-duplicate drawers across the palace. Groups drawers by `source_file`, computes pairwise Jaccard similarity using `drawer_words`,

Fetch all source-file groups with 2+ drawers (optionally filtered by wing). Returns a map from `source_file` → Vec<DrawerInfo>, ordered by content le

Hex-encoded SHA-256 of the diary's full text. Used by the change-detection rule to catch same-size edits (e.g. `"teh" → "the"`) that `(size, mtime)` a

Percent-encode a word for use as a Wikipedia REST API URL path segment. Only unreserved characters (A-Z, a-z, 0-9, `-`, `_`, `.`, `~`) are kept as-is

Extract a JSON block from `text` — handles markdown fences and raw brace pairs. Tries ```` ```json ```` fences first, then bare ```` ``` ```` fences,

Extract the most memorable sentence from `text` for the AAAK quote field. Tries the locale's `quote_pattern` first (verbatim quoted strings are ideal

Extract timestamp from session lines.

Return a fresh adapter instance for `name`, or `None` if not registered. Each call creates a new instance via the registered constructor.

Inverse of [`header_field_encode`]. `?` decodes to empty; the five known `%XX` escapes decode to their literal characters; any other `%XX` (or a trai

Extract the last `RECENT_MSG_COUNT` user messages from a JSONL transcript.

Extract the top `THEMES_MAX` topic words from recent messages. Uses a static stopword list; English-only. Non-English messages produce noisier themes

Spawn `mempalace mine <transcript_parent> --mode convos` in the background. The transcript's parent directory is always mined as a conversation sourc

Stop hook: auto-save every `SAVE_INTERVAL` messages. In silent mode, writes a diary entry directly to the palace DB and emits a `systemMessage`. In b

Validate a transcript path: must have `.jsonl`/`.json` extension and no `..`.

Derive a project wing name from a Claude Code transcript path. Claude Code encodes the project directory with dashes, e.g.: `~/.claude/projects/-User

Suppress the benign `SQLite` error that occurs when an `ALTER TABLE ADD COLUMN` targets a column that already exists (idempotent migration). Any othe

Extract a positive integer argument, coercing floats and strings. MCP JSON transport sometimes delivers integers as floats (`5.0`) or strings (`"5"`)

Disambiguate a word that is both a registered person name and a common English word. Scores context patterns for person usage vs concept usage and re

Rebase pre-scanned file paths onto the canonicalised `project_dir`. Callers (`init::run` → `app::run_mine_after_init`) scan with their own idea of th

Canonicalize `directory` and resolve the wing name. Returns `(canonical_path, wing_name)` where `wing_name` is either the explicit override from `opt

Open (or create) a local turso database and return a connection. For file-backed databases, WAL journal mode is enabled for better concurrency and cr

Parse a Claude Code `.jsonl` file and return all user/assistant messages. Each JSONL line is parsed independently; malformed lines are silently skipp

Parse an LLM response into a name → `(label, reason)` map. Extracts a JSON block, validates each label against [`VALID_LABELS`], and filters to entri

Refine a [`DetectedDict`] using an LLM to drop false positives and reclassify entities. Sends candidates in batches of [`BATCH_SIZE`] to `provider.cl

Called by `regenerate_closets_process_drawer` to parse the LLM JSON response. Strips optional markdown code fences (` ```json `) before parsing. Retu

Run the export command, writing drawers to `output_dir`.

Run a sync pass over the palace. `apply = false` returns the classification report without touching the database. `apply = true` deletes the gitignor

Derive the wing name: prefer the user's own project name from scan results, fall back to the directory name sanitized to `snake_case`. Called by [`ru

Print a labelled entity list. Called by [`run_print_summary`].

Print the privacy warning and optionally prompt for consent when an external LLM is configured via an env-fallback API key. Returns `false` if the us

Inner execution after the LLM provider is confirmed available. Logs progress messages, calls `regenerate_closets`, asserts stat invariants, and logs

Validate tunnel label: trim, non-empty, reject null bytes and length violations.

Scan a `.claude/projects/` directory and return one [`ProjectInfo`] per project. `total_commits` is repurposed as session count — a density signal th

Get chronological timeline of facts.

Convert `scan` results into [`DetectedDict`] — the same shape as `entity_detect::detect_entities`. Capped at `PROJECT_CAP` projects and `PEOPLE_CAP`

Human-readable evidence string used as the `signals` entry in [`DetectedEntity`].

Tokenize text into lowercase words, filtering stop words and short words.

Tokenize a query string into searchable words. The minimum length of 3 matches the indexing threshold in `index_words`: shorter tokens (single letter

Delete an explicit tunnel by its 16-character hex ID.

Try to parse a Codex CLI JSONL session into transcript text. Returns `None` if the file lacks a `session_meta` entry, contains fewer than 2 messages,

Try to parse a Slack JSON message export into transcript text. Valid messages are JSON objects with `type == "message"`, non-empty `text`, and non-em

Bind a random port then immediately drop it, leaving the port unreachable. Used to produce a predictable "connection refused" in network error path t

Iteratively walk `directory` (depth-limited DFS) and collect conversation files. Symlinks are skipped; directories named `tool-results` or `memory` a

Filter AI agent persona names out of `detected` to prevent them from being confirmed as human entities during the init pipeline. When `origin.agent_p

Build the palace graph from drawer metadata. Returns (nodes, edges) where nodes are rooms and edges are tunnels.

Verify that `extracted` matches the palace row count before any destructive write. Two signals trip the guard: 1. **Strong** — the DB reports more dr

Flush an assembled exchange into `chunks`, splitting on `CHUNK_SIZE` boundaries. When content fits in one chunk it is pushed as-is (guarded by `CHUNK

Split conversation content into chunks, routing to `chunk_by_exchange` or `chunk_by_paragraph` depending on whether the content uses the quoted-turn f

Called by [`classify_entity`] to keep that function within the 70-line limit. Constructs the [`DetectedEntity`] from pre-computed ratio and corrobora

clamped to `max(1, min_count)` so a value of 0 still requires one match. Returns the number of tunnels created or refreshed.

Return the active locale language code. Resolution order: `MEMPALACE_LANG`, then `MEMPAL_LANG`, then `"en"`. The result is always lowercased and vali

Compute pairwise Jaccard similarity and return IDs of drawers to delete. For each pair (i, j) where i < j: if Jaccard ≥ `threshold`, the shorter draw