Session 0kk54et5Shz3

<skill_content name="jj-guide">
# Skill: jj-guide

# jj Guide for AI Agents

Jujutsu (jj) is a Git-compatible VCS with mutable commits, automatic snapshotting, no staging area, and first-class conflicts. This skill teaches you how to use it safely from a non-interactive agent environment.

If `.jj/` exists in the repo root, this is a jj repo. **Use `jj` commands, not `git`.** In a colocated repo (`.jj/` *and* `.git/`), git tools can read the state, but mutations should go through `jj` so the operation log stays consistent.

## Critical Rules

- **NEVER** use interactive flags (`-i`, `--interactive`). TUI prompts hang in agent environments. This applies to `jj split -i`, `jj squash -i`, `jj commit -i`, `jj resolve`, `jj diffedit`, etc.
- **ALWAYS** pass `-m "msg"` when describing/committing. Without `-m`, an editor opens and hangs.
- **VERIFY** mutations with `jj st` and `jj log` after `squash`, `abandon`, `rebase`, `restore`, `commit`. ...

<path>/home/byk/Code/opencode-lore-embedding-runtime</path>
<type>directory</type>
<entries>
.craft.yml
.git
.github/
.gitignore
.jj/
.lore.md
.oxfmtrc.json
.oxlintrc.json
AGENTS.md
CHANGELOG.md
codecov.yml
LICENSE
node_modules/
package.json
packages/
patches/
pnpm-lock.yaml
pnpm-workspace.yaml
quality/
README.md
scripts/
stryker.config.mjs
supabase/
tsconfig.base.json
tsconfig.json
vitest.config.ts
vitest.evals.config.ts
vitest.mutation.config.ts

(28 entries)
</entries>

/home/byk/Code/opencode-lore-embedding-runtime/packages/website/package.json
/home/byk/Code/opencode-lore-embedding-runtime/packages/pi/package.json
/home/byk/Code/opencode-lore-embedding-runtime/packages/opencode/package.json
/home/byk/Code/opencode-lore-embedding-runtime/package.json
/home/byk/Code/opencode-lore-embedding-runtime/AGENTS.md
/home/byk/Code/opencode-lore-embedding-runtime/packages/core/package.json
/home/byk/Code/opencode-lore-embedding-runtime/packages/gateway/package.json

<path>/home/byk/Code/opencode-lore-embedding-runtime/AGENTS.md</path>
<type>file</type>
<content>
1: <!-- This section is maintained by the coding agent via lore (https://github.com/BYK/loreai) -->
2: ## Long-term Knowledge
3: 
4: For long-term knowledge entries managed by [lore](https://github.com/BYK/loreai) (gotchas, patterns, decisions, architecture), see [`.lore.md`](.lore.md) in the project root.
5: <!-- End lore-managed section -->
6: 
7: ## Review
8: 
9: See [`quality/REVIEW.md`](quality/REVIEW.md) for the review heuristics and workflow — regression-test discipline, adversarial-order state setup, fan-out registry coverage, recurring bug-class batteries, and the two-reviewer rule. Every non-trivial PR requires an adversarial correctness review before merge.
10: 
11: # Project Overview
12: 
13: Lore is a **three-tier memory architecture** for AI coding agents. ...

<path>/home/byk/Code/opencode-lore-embedding-runtime/quality/REVIEW.md</path>
<type>file</type>
<content>
1: # Review Heuristics — Lore
2: 
3: This document codifies the review bar for pull requests. It exists because
4: adversarial review has repeatedly caught correctness bugs that a green test
5: suite missed — skip/early-return branches, lifecycle sequences, and fan-out
6: registry gaps. Institutionalizing these patterns so they compound.
7: 
8: The review workflow (§6) is referenced from AGENTS.md.
9: 
10: ---
11: 
12: ## 1. Regression-test discipline
13: 
14: Every adversarial-review finding that surfaces a defect **must** land a
15: deterministic regression test in the same PR. The test:
16: - fails on the base branch (proving it reproduces the bug),
17: - passes on the fix (proving the fix works),
18: - drives the *real* precondition, not an artificially-constructed state (e.g. ...

<path>/home/byk/Code/opencode-lore-embedding-runtime/package.json</path>
<type>file</type>
<content>
1: {
2:   "name": "lore-monorepo",
3:   "private": true,
4:   "type": "module",
5:   "license": "FSL-1.1-Apache-2.0",
6:   "description": "Monorepo root for Lore — three-tier memory architecture",
7:   "main": "./packages/opencode/src/index.ts",
8:   "exports": {
9:     ".": "./packages/opencode/src/index.ts"
10:   },
11:   "packageManager": "pnpm@10.28.0",
12:   "volta": {
13:     "node": "24.16.0",
14:     "pnpm": "10.28.0"
15:   },
16:   "pnpm": {
17:     "onlyBuiltDependencies": [
18:       "esbuild",
19:       "onnxruntime-node",
20:       "sharp"
21:     ],
22:     "overrides": {
23:       "basic-ftp": "5.3.1",
24:       "ip-address": "10.2.0",
25:       "fast-xml-parser": "5.7.3",
26:       "@huggingface/hub": "2.11.0",
27:       "onnxruntime-node": "1.27.0",
28:       "yaml": ">=2.8.3",
29:       "esbuild": ">=0.28.1",
30:       "@opentelemetry/core": ">=2.8.0",
31:       ...

<path>/home/byk/Code/opencode-lore-embedding-runtime/packages/core/package.json</path>
<type>file</type>
<content>
1: {
2:   "name": "@loreai/core",
3:   "version": "0.40.0",
4:   "type": "module",
5:   "license": "FSL-1.1-Apache-2.0",
6:   "description": "Shared memory engine for Lore — three-tier storage, distillation, gradient context management",
7:   "main": "./dist/node/index.js",
8:   "types": "./dist/node/index.d.ts",
9:   "exports": {
10:     ".": {
11:       "types": "./dist/types/index.d.ts",
12:       "development": "./src/index.ts",
13:       "bun": "./dist/bun/index.js",
14:       "default": "./dist/node/index.js"
15:     }
16:   },
17:   "imports": {
18:     "#db/driver": {
19:       "bun": "./src/db/driver.bun.ts",
20:       "default": "./src/db/driver.node.ts"
21:     }
22:   },
23:   "scripts": {
24:     "typecheck": "tsc --noEmit && tsc --noEmit -p tsconfig.eval.json",
25:     "build": "tsx script/build.ts"
26:   },
27:   "dependencies": {
28:     "@hpke/core": ...

Found 13 matches
/home/byk/Code/opencode-lore-embedding-runtime/.lore.md:
  Line 29: * **Embedding token batching and temporal chunk cap**: Use embedInTokenBatches for all embedding paths, including temporal messages, to prevent ONNX padding all inputs to the longest sequence and causing OOM. Batch by token area (MAX\_BATCH\_TOKEN\_AREA 4096), but always admit the first item so the loop progresses. Preserve input order and call storeTemporalChunks once. Cap temporal chunks at 64: retain the first 63 and join the remainder into one tail; an oversized tail may occupy its own batch and is truncated downstream.

  Line 41: * **embedding-units 64-chunk cap**: Chose structural \`"\n\x1f"\` boundaries and a 64-chunk temporal cap over splitting on tool-looking text, dropping overflow, or allowing unlimited vec0 rows because tool output may mimic headers and unbounded fan-out expands storage and KNN widening. ...