acdc

ACDC

Part of the AWITHI family — AI with AI. ACDC = Agent Context Delivery Channel — AC ↔ DC, context flowing both ways.

Read this in: English · 한국어

The context of a coding-agent session — what you decided, what's already tried, where you left off — is the most valuable thing in the room. It's also fragile: rate limits stop you mid-flow, compaction leaves the agent foggy on earlier turns, and the trail goes cold across days.

ACDC reads your local Claude Code and Codex sessions to put the context back where you need it.

What ACDC does

Skill	What it does	Example invocation
resume-codex-session (Claude side) / resume-claude-session (Codex side)	Cross-agent handoff: read the prior session and current repo state, then hand a structured summary to the new agent so it continues rather than echoes the last few messages.	"resume the codex session called api-refactor"
recall-session	Self-recall: look up your own past session by name or UUID. Returns a timeline summary — what you decided, what changed, where you left off.	"show me yesterday's api-refactor session"
recall-context	Topic search: find where you discussed something across your sessions, even pre-compaction in the current one.	"where did we decide on the ACDC name"

Three skills per agent, each plugin self-contained.

Install

Claude Code — slash commands inside the CLI:

/plugin marketplace add https://github.com/awithi-co/acdc
/plugin install acdc

Codex — shell command + enable:

codex plugin marketplace add awithi-co/acdc

marketplace add registers ACDC as available. To activate it (so skills are discoverable by natural-language dispatch), either pick ACDC from codex's interactive plugin picker, or append to ~/.codex/config.toml:

[plugins."acdc@acdc"]
enabled = true

After install, trigger skills with natural language — the example invocations above work as-is. The plugin's manifest routes the request to the correct skill in each agent.

How it works

Each skill ships two Python helpers:

• a finder that scores local session candidates by name, cwd hint, and recency
• a bounded summarizer that reconstructs the timeline from the transcript or rollout JSONL

For handoff, the skill cross-checks the reconstructed state against the current git status, branch, and worktree before producing the summary. The new agent acts on the handoff, not on raw transcript content.

For recall by identifier, the same finder/summarizer pair is used without the repo cross-check (the agent already has its own context).

For topic search, a grep_recall.py helper scans JSONL files (current session + recent partition) and returns matched segments with surrounding event context.

Pure Python 3 standard library only. No network access. Read-only against session storage.

Session storage layout

ACDC only reads these paths; it never writes to them.

Claude Code — ~/.claude/

projects/<escaped-cwd>/<session-uuid>.jsonl   transcript (full event stream)
sessions/<pid>.json                           live session metadata (name, cwd, status)
tasks/<session-uuid>/N.json                   structured todos
history.jsonl                                 per-prompt log

• Sessions are identified by UUID. The transcript file is named <uuid>.jsonl, and its parent directory encodes the original working directory (slashes replaced with dashes).
• Session rename events live inside the transcript as custom-title records. Live sessions also expose the current name via ~/.claude/sessions/<pid>.json.
• Tasks are stored as separate JSON files ({id, subject, description, activeForm, status}), one per todo item.

Codex — ~/.codex/

sessions/YYYY/MM/DD/rollout-<timestamp>-<uuid>.jsonl   rollout (event stream, source of truth)
session_index.jsonl                                    {id, thread_name, updated_at} pointer file
shell_snapshots/<...>                                  per-session shell state
history.jsonl                                          per-prompt log

• Sessions are identified by UUID. Rollout files are partitioned by date, not cwd.
• session_index.jsonl is an accelerator for name lookup. It is not guaranteed to contain every rollout on disk, so the rollout tree is the source of truth.
• Todos and plans are not separate files. They live inline in the rollout as function_call events (update_plan, TodoWrite).

Requirements

Python 3 on PATH. Pre-installed on macOS 12.3+ and most Linux distributions; on Windows, install from python.org if unavailable. The skill tries uv run python, then python3, then python, and stops with a clear message if none are found.

Privacy & limitations