anyllm-ctx 0.1.0

Git for LLM context. Snapshot a dying session, brief the next model in 30 seconds.

anyllm — Git for LLM Context

Snapshot a dying AI coding session. Brief the next model in 30 seconds. Keep moving.

You're deep in a Claude Code session. Context window fills. Or you want a second opinion in Codex. Or you switch machines. Either way — the next tool needs the whole project explained again.

anyllm snapshots that session into a compact briefing and injects it anywhere: paste it, push it to a browser tab, or let a slash command do it automatically in any of 7 supported AI coding CLIs.

And it doesn't just snapshot the latest session. Every pack merges the new snapshot into a rolling project memory — decisions made three sessions ago survive even if today's session never mentioned them.

---

Install

pip install anyllm

Requires Python 3.10+. Set ANTHROPIC_API_KEY or OPENAI_API_KEY for distillation (required). Add to your shell profile.

---

First-Run Setup

cd your-project
anyllm install

This does three things in one command:

1. Creates .anyllm/ in the current project
2. Detects which AI coding CLIs you have installed
3. Installs /anyllm-pack (and 7 other commands) into all of them

After that, you never need to leave your AI coding session:

/anyllm-pack       ← works in Claude Code, OpenCode, Kiro, Kilocode
$anyllm-pack       ← works in Codex
anyllm-pack        ← type as a message in Antigravity/Agy

---

Commands

Command	What it does
anyllm init	Create .anyllm/ in the current project
anyllm pack	Snapshot the most recent session, merge into current.md
anyllm repack	Ingest turns missed since the last pack (delta update, no duplicate work)
anyllm prime [--target MODEL] [--copy] [--write PATH]	Render a briefing for the next model
anyllm push	Silently inject the briefing into a browser tab and press Send
anyllm status	Show task, decisions, and repository context state
anyllm log	Table of every packed session with per-session decision deltas
anyllm diff SESSION_ID	Show the raw snapshot from a past session
anyllm install	First-run setup: init + integrate all detected CLIs
anyllm integrate [NAME]	Install slash commands into one or all detected CLIs
anyllm integrations	Show which CLIs are detected and installed
anyllm uninstall NAME	Remove slash commands from a CLI

---

CLI Integrations

anyllm install (or anyllm integrate --all) installs commands into every AI coding CLI you have on the machine. Run anyllm integrations to see what's detected and installed.

CLI	Detection	Invocation	Slash commands
Claude Code	~/.claude/	/anyllm-pack	Zero-token (no AI call)
Antigravity / Agy	~/.gemini/ or agy binary	type anyllm-pack as a message	AI-triggered skill
OpenCode	opencode binary	/anyllm-pack	AI-triggered
Codex	~/.codex/	$anyllm-pack	AI-triggered
Kiro	~/.kiro/	/anyllm-pack	Steering doc
Kilocode	~/.kilocode/	/anyllm-pack	Zero-token
Cursor	~/.cursor/	/anyllm-pack	Skill directory

Every integration installs the same 8 commands:

/anyllm-init     /anyllm-pack     /anyllm-repack   /anyllm-prime
/anyllm-push     /anyllm-status   /anyllm-log      /anyllm-diff

Claude Code — Zero Token

Claude Code slash commands with disable-model-invocation: true run directly in the shell with no AI tokens consumed. anyllm pack inside Claude Code costs exactly zero tokens.

Antigravity / Agy

Antigravity (Google's agy CLI) does not support custom slash commands — only built-in / commands like /skills, /settings. Instead, type the command name as a plain message:

anyllm-pack

The installed skill activates automatically and runs anyllm pack.

Manual Integration

Install into a specific CLI without the auto-detect:

anyllm integrate claude      # Claude Code
anyllm integrate gemini      # Antigravity / Agy
anyllm integrate opencode    # OpenCode
anyllm integrate codex       # Codex
anyllm integrate kiro        # Kiro
anyllm integrate kilo        # Kilocode
anyllm integrate cursor      # Cursor

anyllm integrate --project   # install to .agents/skills/ instead of user config

---

How It Works

Ingestor → Distiller → Merger → Composer → Adapter

Ingestor reads ~/.claude/projects/*.jsonl (or similar) into a normalized turn-by-turn transcript.

Distiller compresses it into a structured snapshot — task, decisions, code map, failed approaches, next step — using your configured LLM.

Merger combines the new snapshot with the existing current.md instead of overwriting it. Decisions accumulate across sessions. Nothing is silently dropped. Repository analysis (when available) verifies decisions against your actual source code.

Composer wraps the snapshot in role framing and anti-repetition guards so the receiving model doesn't re-explore closed questions.

Adapter renders the briefing in the format expected by the target (chatgpt, claude, cursor, ...).

repack — Zero-Duplication Delta Update

anyllm repack reads only the turns that happened after the last pack, distills just those, and merges the delta in. Use it mid-session without re-processing the entire history.

---

Confidence-Aware Snapshot Merging

Every pack classifies every known decision:

State	Meaning
CONFIRMED	Re-stated this session, or verified against the codebase
ADDED	New this session
UPDATED	Significantly reworded — old wording archived in Superseded Decisions
STALE	Absent from this session, confidence uncertain — surfaced in Stale / Needs Verification
ORPHANED	Code anchor gone and absent for stale_threshold consecutive sessions

Decisions are matched across sessions by normalized hash + character-bigram similarity, so the same decision worded two different ways is recognized as one, not two.

Sections That Never Get Dropped

• Failed Approaches — union of all sessions. If something failed once, every future model knows.
• Open Questions — carried forward until a session explicitly resolves one.

Session Provenance

Every decision tracks which session introduced it and which sessions re-confirmed it. current.md includes a merged_from list, a confidence_report summary, and a Session Provenance table. State persists in frontmatter between packs (decision_provenance).

Graceful Degradation

Merging works without any extra tools. If a merge fails, pack falls back to a plain write rather than losing the snapshot. Disable merging with merge.enabled: false.

---

Config

.anyllm/config.yaml is created by anyllm init with defaults:

distiller:
  model: gpt-4o-mini
  budget_tokens: 2000

targets:
  default: chatgpt

framing:
  extra_rules: []
  tone: direct

merge:
  enabled: true
  stale_threshold: 3      # consecutive sessions absent before ORPHANED

repository_analysis:
  enabled: true
  timeout: 30             # seconds; subprocess timeout
  auto_refresh: true      # analyze repo automatically on pack

push:
  browser: auto           # auto | chrome | firefox | edge
  codex_url: https://codex.openai.com
  send_delay_ms: 500
  open_if_missing: true

---

Storage Layout

.anyllm/
├── config.yaml              # model, target, merge, push settings
├── index.json               # session log with per-session decision deltas
├── current.md               # rolling merged project memory (what prime reads)
└── sessions/
    ├── <date>-<id>.transcript.json   # normalized raw session
    └── <date>-<id>.snapshot.md       # distilled per-session snapshot

All files are plain text — hand-editable, diff-able, committable. The .anyllm/ directory belongs in your repo.

---

Roadmap

• Additional ingestors: Codex, Gemini, OpenCode, raw markdown
• Additional adapters: Cursor, Gemini, local Llama
• Embedding-based decision matching for better paraphrase handling
• Team workspace sync

---

Philosophy

Every AI coding session is stateless. You build up context across dozens of turns, and the moment you switch tools — or the context window fills — that shared understanding evaporates.

anyllm treats context as a first-class artifact: something you build once, version, and carry forward — not something you reconstruct from memory every time.

The .anyllm/ directory belongs in your repo. Commit it.