source-command-latch-seed

A Cursor skill that previews and applies decision records from the current Cursor conversation. Cursor is an editor with built-in coding-agent features.

In plain words
What is it for?
Use it with `latch-seed` or `/latch-seed` to find decisions in the current chat and prepare them for saving in Latch.
Why use it?
It keeps the skill focused on the conversation the user is currently having and avoids accidentally reading private conversations from elsewhere.

Skill for Claude CodeCodex

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 skills/open-latch/latch/source-command-latch-seed
Any agent
npx skills add open-latch/latch --skill source-command-latch-seed
Clone the repo
git clone --depth 1 https://github.com/open-latch/latch

Made for: Claude Code, Codex.

Per session 37 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 899 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 $0.00037 $0.00899
Opus 5 $0.00018 $0.00449
Sonnet 5 $0.00007 $0.00180
Haiku 4.5 $0.00004 $0.00090

Measured 3d ago against content hash a013daee60d4, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

source-command-latch-seed 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 3d ago.

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.

cursor_skills/source-command-latch-seed/SKILL.md · 68 lines

What it actually says

Latch seed for Cursor

Latch operation id: latch-seed preview

Latch Cursor skill boundary: use only the current sessionStart marker or a transcript path the user explicitly supplied. Never scan private Cursor history directories.

Read the exact Latch Cursor current session id from the current prompt context; the beforeSubmitPrompt hook re-injected it from this chat's payload. Pass --cursor-session-id ID to the wrapper. Never omit it or reuse an id from another chat.

Before any Shell call, read the workspace .cursor/mcp.json and use mcpServers.latch.env.LATCH_PYTHON when present, otherwise mcpServers.latch.command, as <CURSOR_MCP_PYTHON> and LATCH_PYTHON. Never fall back to a PATH python3; the MCP interpreter owns latch's native dependencies. Use latch_home only to construct the absolute wrapper path. Do not export LATCH_HOME or CLAUDE_KB_HOME in the Shell call; managed Cursor operation receipts do not allow those environment assignments.

Resolve latch_home as ${CURSOR_PLUGIN_ROOT} when set, otherwise use the absolute checkout in the project-sync footer. Select native cursor for plugin installs or the backend in that footer. Run the host-appropriate bin/latch_seed.sh or .ps1 wrapper with --source cursor --format json and the seed/model backend environment set. First run preview-only and show the receipt/candidates. The first Shell call must request Cursor required_permissions: ["all"], because the wrapper writes latch-owned budget/session state outside the open workspace. Do not make a sandboxed attempt first and retry with permission: the managed preview receipt is one-shot and consumed by the first exact attempt. Cursor's normal user-approval flow still applies, and this does not authorize apply. The attempt alone does not arm apply: only a matching successful JSON result verified by postToolUse does. Failed, malformed, missing, cross-session, or unexecuted previews require a new preview. Only after the user reviews that preview, ask for an exact selection-bearing confirmation:

  • Whole report: /latch-seed apply all
  • Reject the whole nonempty report: /latch-seed apply none
  • Scoped: /latch-seed apply <ID> [<ID> ...], with every approved cand-... and/or cluster-... review ID as a separate space-delimited token, for example /latch-seed apply cand-0123456789ab cluster-abcdef012345.

Bare /latch-seed apply, prose standing in for IDs, duplicate IDs, and all/none mixed with IDs are invalid. The managed receipt validates every scoped ID against the bound preview and records only that exact selection. none is valid only after a successful preview containing at least one candidate.

For whole-report confirmation, rerun with --preview-digest DIGEST --apply --yes. For reject-all confirmation, rerun with --preview-digest DIGEST --apply --dismiss-all; never add --yes or an approval selector. For scoped confirmation, omit --yes and translate exactly the confirmed tokens to --approve-candidate ID and/or --approve-cluster ID; do not add, remove, or substitute IDs. Unselected items are dismissed for those exact source revisions. Use the exact returned JSON preview_digest. Apply must load the cached reviewed candidates without a second model call; a missing, changed, or stale digest, malformed confirmation, or unknown ID is a hard stop. The apply Shell call must again request Cursor required_permissions: ["all"] on its first and only attempt because it writes staging KB state outside the workspace and consumes a one-shot receipt. Report inserted staging node ids and the catch-demo command. A missing/mismatched marker is a hard stop.

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. 3d ago First seen · 68 lines · 37 tokens per session scan A a013daee60d4

Subscribe to this mod's changes

source-command-latch-seed is a skill published in the GitHub repository open-latch/latch (9 stars, last pushed 3d ago), licensed Apache-2.0. It adds 37 tokens to every session and 899 once invoked, about $0.0002 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.