rlat

name: rlat description: >- Orchestrate rlat CLI workflows on the user's behalf for any task involving an rlat knowledge model (.rlat file). Triggers cover: setting up rlat for a fresh project (init), building or refreshing a knowledge model on drift, running grounded retrieval to ground an answer in citations, generating skill-context for an Anthropic skill !command block, recalling past decisions from memory, comparing two corpora, switching storage modes (bundled / local / remote), and the opt-in MRL optimise pipeline. Trigger when the user mentions rlat by name, references a .rlat file, asks to build/query/refresh/sync/watch a knowledge model, asks "what does the project know about X" on a project with a .rlat file, or describes any task that maps onto an rlat workflow. This skill executes the workflow on the user's behalf — it does not just document the CLI. For multi-hop research with cross-file synthesis (plan → retrieve → refine → synthesize loop), defer to the deep-research skill instead. Not for: Pinecone, Chroma, FAISS, or other vector databases; raw embedding ops without rlat. allowed-tools: Bash(rlat:*), Read, Write, Edit, Glob, Grep

rlat — workflow orchestration

You are running an rlat workflow on the user's behalf. The CLI is deterministic; your job is to (1) identify which workflow fits the user's request, (2) execute it with the right flags for their context, and (3) surface what changed and what's next.

Step 1: Identify the workflow

Step 2: Pre-flight (before any workflow that builds or runs)

• Encoder installed? First-time builds need a one-shot encoder install (~2 min on CPU). If rlat build or rlat search errors with "encoder not found", run rlat install-encoder once and tell the user it's a one-time setup (cached at the standard XDG cache location).
• Find the right .rlat file. Use Glob "*.rlat" to discover. If multiple, prefer the one whose name matches the project (resonance-lattice.rlat, project.rlat) or ask the user.
• Check freshness when the workflow assumes a current build: rlat profile <km>.rlat reports drift count.

If a precondition fails, fix it first or surface it clearly to the user before running the workflow.

Init

When: project has no .rlat file or the user explicitly asks to set up rlat for this project.

Run:

rlat init-project

Auto-detects docs/, src/, top-level *.md/*.rst/*.txt. Builds the knowledge model and generates .claude/resonance-context.md (the project primer) in one command.

For non-default source paths:

rlat init-project --source ./<dir1> --source ./<dir2>

After:

• Verify the .rlat and .claude/resonance-context.md both exist.
• Tell the user the next-step paths: single-shot retrieval is now rlat search ./<km>.rlat "<query>"; for multi-hop research, the deep-research skill takes over.

Refresh

When: source files changed since the last build, drift detected on a recent search, or rlat profile reports any non-zero drift count.

Run (local-mode KMs):

rlat refresh <km>.rlat

Atomic. Re-encodes only the changed passages; preserves any existing optimised band by reprojecting it from the new base for free. If the optimised band would be invalidated by the changes, the command aborts unless --discard-optimised is passed — surface this to the user before deciding.

Run (remote-mode KMs):

rlat sync <km>.rlat

Same shape; fetches the manifest delta over HTTP first.

After:

• Run rlat profile <km>.rlat — drift count should be 0.
• If a primer was generated from this KM, regenerate: rlat summary <km>.rlat -o .claude/resonance-context.md.

Watch (live refresh)

When: the user is actively editing source files and wants the knowledge model to stay current without remembering to refresh after each save. Also: CI / pre-commit hooks (--once) where a synchronous one-shot reconciliation is the right shape.

Run (live, foreground):

rlat watch                          # zero-arg: discover *.rlat in cwd
rlat watch <km>.rlat                # explicit single archive
rlat watch --verbose                # one line per refresh (default: silent)

Default UX is silent. Errors are loud, success is invisible. On Ctrl-C, prints a one-line summary (elapsed time, refreshes, errors). Local-mode only — bundled / remote archives fail preflight with a rlat convert / rlat sync hint.

Run (CI / pre-commit shape):

rlat watch --once                   # synchronous one-shot reconcile + exit

--once walks every preflighted archive, runs bucketise + apply_delta against current disk state, and exits. No observer, no event wait — files are typically already changed before the command runs in CI. Returns 0 even when archives are already up to date.

Pre-flight requirements:

• Local-mode KM (run rlat convert <km> --to local first if bundled).
• [watch] extra installed (pip install rlat[watch] — pulls watchdog + the build extras for re-encoding).
• The recorded source_paths exist on disk.

Concurrency: per-archive threading.Lock serialises refreshes so concurrent FS events can't race the atomic-write path. Same incremental delta-apply pipeline as refresh; optimised bands re-project for free on every fire.

Override: debounce window via RLAT_WATCH_DEBOUNCE_MS env var (default 1000).

Search

When: user has a one-shot question whose answer plausibly lives in the indexed corpus.

Run:

rlat search <km>.rlat "<question>" --top-k 5 --format context --mode <mode>

Pick the mode based on the user's risk tolerance:

• --mode augment (default) — passages as primary context, blended with training. The right pick for general-purpose use.
• --mode constrain — passages are the only source of truth, refuse on thin evidence. Pick for compliance / audit / regulatory work where wrong-but-confident is unacceptable. Bench: 91.7% distractor refusal at 2.0% answerable hallucination — the safety floor.
• --mode knowledge — passages supplement training. Pick when the LLM has solid base domain knowledge.

Add --strict-names when the question references specific proper nouns / acronyms / SKUs and a name-mismatch should refuse rather than silently substitute.

After:

• Use the returned passages as the basis for your answer; cite source paths inline (docs/foo.md, src/bar.py).
• If retrieval coverage looks thin (top-1 score < 0.30, all hits drifted), surface the gap honestly; don't extrapolate.

Skill-context

When: user is authoring an Anthropic skill that needs corpus-grounded context injected via !command blocks.

Run (preset + dynamic-query pattern):

rlat skill-context <km>.rlat \
  --query "<skill-author preset>" \
  --query "$user_query" \
  --top-k 5 --token-budget 4000

Multi-query order matters: skill-author preset first, $user_query last. When the token budget runs out, later queries drop first; the first preset always survives.

Compliance variant (drift + name-mismatch are hard refusals):

rlat skill-context <km>.rlat --strict --strict-names --mode constrain \
  --query "$user_query"

--strict aborts on drift (rc=2). --strict-names aborts on name-mismatch (rc=3). The skill loader gates on the exit code.

Memory

When: user asks about past sessions, prior decisions, recurring themes — "have we discussed X before?", "what did we decide about Y?", "what was the rationale agreed for Z?".

The memory directory is selected with --memory-root on the parent memory command. Default: ./memory/ (project-local). For Claude Code session memory, use .claude/memory.

Run (recall):

rlat memory --memory-root ".claude/memory" recall "<query>"

Returns ranked memory entries fused across the working / episodic / semantic tiers (recency-weighted retention; cross-session knowledge survives in the semantic tier).

Add a memory (when the user wants to record a decision for future sessions):

rlat memory --memory-root ".claude/memory" add "<text>" --tier semantic

Generate a session-start primer from the memory state:

rlat memory --memory-root ".claude/memory" primer -o .claude/memory-primer.md

Consolidate episodic near-duplicates into the semantic tier:

rlat memory --memory-root ".claude/memory" consolidate

Garbage-collect entries past their retention half-life:

rlat memory --memory-root ".claude/memory" gc

Compare

When: user wants to compare two corpora — typically two snapshots of the same project, or two related products.

Run:

rlat compare a.rlat b.rlat

Always uses the base band (cross-knowledge-model rule). Outputs centroid cosine + asymmetric mutual coverage (both directions).

For deeper analysis of a single KM (cross-passage relationships, evidence per claim, drift detection, corpus diff across snapshots), use the RQL ops via the Python API:

from pathlib import Path

from resonance_lattice.field.encoder import Encoder
from resonance_lattice.rql import corpus_diff, drift, evidence
from resonance_lattice.store import archive, open_store

contents_a = archive.read("a.rlat")
store_a = open_store(Path("a.rlat"), contents_a, None)

# Evidence + ConfidenceMetrics for a claim (embed the claim first):
claim_emb = Encoder().encode(["<claim>"])[0]
evidence(contents_a, store_a, claim_emb, top_k=10)
# Within-KM stale passages:
drift(contents_a, store_a)
# Added / removed / unchanged across two snapshots:
corpus_diff(contents_a, archive.read("b.rlat"))

These ops are Python-only in v2.0; CLI verbs for them are not in scope. For contradiction detection use the rlat-contradictions skill (the self-audit pairs stored in every .rlat, judged in-session) — the old experimental RQL contradictions op was removed 2026-06.

Convert

When: user wants to switch a .rlat between storage modes (bundled / local / remote).

Run:

rlat convert <km>.rlat --to <mode>
# remote requires:
rlat convert <km>.rlat --to remote --remote-url-base <url>

In place; embeddings preserved (semantically identical pre/post). Conversion validates every kept passage's content_hash against live bytes; if any have drifted, conversion aborts. Run rlat refresh (local) or rlat sync (remote) first if so.

Optimise (opt-in, requires Anthropic API key)

⚠ Cost: ~$2-8 in Anthropic API calls + ~20-50 min Kaggle T4 GPU time per corpus. Two of three publicly-tested corpora regressed (BEIR fiqa: -0.042 nDCG@10, BEIR nfcorpus: -0.043). Lift correlates with synth-query/test-query distribution alignment, not "natural-language vs keyword" surface form. Per-corpus decision; run --estimate first.

When: user explicitly asks for the optimised band, OR retrieval recall is the bottleneck on a natural-language Q&A corpus where this can pay off.

Run:

export CLAUDE_API=sk-ant-...                        # or ANTHROPIC_API_KEY
rlat optimise <km>.rlat --estimate                  # cost preview FIRST
rlat optimise <km>.rlat --corpus-description "<one-line corpus description>"

The corpus description conditions the synth-query generator. Without it, lift drops materially — always pass it for non-trivial corpora.

API key setup if missing: see docs/user/API_KEYS.md.

Programmatic deep-search (CLI verb, requires API key)

The rlat deep-search CLI verb runs the same multi-hop research loop the deep-research skill drives interactively, but as a standalone CLI command suitable for non-Claude-Code agents, CI pipelines, batch scripts, or any consumer that needs a programmatic surface with structured output.

⚠ Requires an Anthropic API key (~$0.009-0.025/q). For interactive Claude Code research, the deep-research skill drives the same loop with no API cost — prefer that skill in nearly every Claude Code scenario.

When to invoke this CLI verb instead of the skill:

• The deliverable is a machine-readable artefact (JSON, markdown for piping into another LLM, file output)
• The user is authoring a CI pipeline, agent, or batch script
• The user explicitly asks for rlat deep-search

Run:

rlat deep-search <km>.rlat "<question>" \
  [--max-hops 4] [--top-k 5] \
  [--format text|json|markdown] \
  [--strict-names]

Exit codes: 0 success, 2 no-answer, 3 name-mismatch refusal.

When NOT to use this skill

• Question is answerable from training alone (no .rlat involved). Just answer.
• User is asking about Pinecone, Chroma, FAISS, or other vector DBs — different products.
• User wants to read or edit specific source files. Use Read / Edit / Grep directly.
• Question requires multi-hop synthesis across files — defer to the deep-research skill.
• User just wants a one-line answer that doesn't need grounding — answer directly without the retrieval round-trip.

Reference

• docs/user/CLI.md — full CLI reference (every command, every flag — the canonical source)
• docs/user/CORE_FEATURES.md — seven things rlat enables
• docs/user/STORAGE_MODES.md — bundled / local / remote decision guide
• docs/user/API_KEYS.md — only for optimise and the deep-search CLI verb
• docs/user/FAQ.md — common questions including licence