latch-heal

latch-heal is a command for coding agents from open-latch/latch. It costs 11 tokens per session (499 once invoked), scanned A, original, Apache-2.0.

A maintenance command that checks Latch's project knowledge base for broken links and similar or conflicting records. A knowledge base is a store of project information used for later decisions.

In plain words
What is it for?
Use it for scheduled maintenance: remove orphaned links, synchronize stored records, and mark one of two similar records as outdated or related when appropriate.
Why use it?
It cleans up inconsistent records and handles possible duplicates, while keeping an audit trail and respecting daily limits on language-model calls.

Command

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add commands/open-latch/latch/latch-heal
Clone the repo
git clone --depth 1 https://github.com/open-latch/latch

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for latch-heal

README.md
[![agentmods](https://agentmods.dev/badge/commands/open-latch/latch/latch-heal.svg)](https://agentmods.dev/commands/open-latch/latch/latch-heal)
Your own site
<a href="https://agentmods.dev/commands/open-latch/latch/latch-heal"><img src="https://agentmods.dev/badge/commands/open-latch/latch/latch-heal.svg" alt="Measured on agentmods" height="20"></a>
Per session 11 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 499 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00011 $0.00499
Opus 5 $0.00005 $0.00249
Sonnet 5 $0.00002 $0.00100
Haiku 4.5 $0.00001 $0.00050

Measured today against content hash 85f0c81c8d9c, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

latch-heal scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured today.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

commands/latch-heal.md · 44 lines

What it actually says

Run the nightly heal against the per-project knowledge base under <KB_HOME>. Two phases:

  1. Integrity — clean up orphan edges and sync the vec_nodes virtual table with the main nodes table (delete orphans, backfill missing).
  2. Contradiction sweep — for every non-stale node with an embedding, find near-duplicates at cosine similarity ≥ 0.70 (the range the on-insert heal deferred). For each pair:
    • Skip if an edge already exists between them (idempotent — prior sweeps' decisions stick).
    • Apply three-pass arbitration:
      • Pass A (recency): newer wins if age diff > 30d AND newer is still fresh (updated in last 30d).
      • Pass B (ref_count): the dominant side wins if ratio ≥ 3 AND both have been referenced at least once.
      • Pass C (LLM): only invoked when A and B are inconclusive. Uses the selected maintenance model backend (Claude by default; adapter env can select Codex) and counts against the daily budget.
    • Verdict: supersede marks loser stale + adds a supersedes edge. keep_both adds a related_to edge so this pair is skipped on future sweeps.
    • If the heal budget is exhausted before arbitration, leave the pair untouched and edge-free. Record a structural heal_deferred row with the pair ids, similarity, and tier so the same pair can be audited and retried automatically on a later run.

Run with the Bash tool:

python "<KB_HOME>/src/latch/pipeline/maintenance.py" nightly "$(pwd)"

Report the JSON summary: examined, collisions, superseded, kept_both, deferred, per-path counts (recency / ref_count / llm / deferred), and budget_blocked if any collisions were left pending because the daily cap was hit. Budget-blocked pairs do not receive an edge and remain eligible for the next heal run.

If anything fails, check <KB_HOME>/maintenance.log.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. today Changed 85f0c81c8d9c
  2. 4d ago First seen · 44 lines · 11 tokens per session scan A b06422033954

Subscribe to this mod's changes

latch-heal is a command published in the GitHub repository open-latch/latch (9 stars, last pushed yesterday), licensed Apache-2.0. It adds 11 tokens to every session and 499 once invoked, about $0.0001 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.