feat(cli): resume sessions from Claude Code, Codex, and Cursor #1057

Problem

When a user switches from Claude Code, Codex, or Cursor to Maka, they have no way to continue their previous session. They must manually re-explain context, re-open files, and reconstruct where they left off.

Desired outcome

Maka detects sessions from Claude Code and Codex on the local machine, shows them in the session picker with a source badge, and generates a handoff summary when the user selects one. The user can continue work without re-explaining context.

Inspired by Grok Build's foreign session implementation (xai-org/grok-build, xai-grok-workspace/src/foreign_sessions/).

Architecture: two layers

Layer 1: Session list scanner (TypeScript)

Read-only metadata scan of foreign session stores. Returns session ID, title, cwd, updated time, and git branch. No conversation content is read at this stage.

Claude Code sessions live at ~/.claude/projects/<encoded-cwd>/<uuid>.jsonl. Each line is a JSON object. Scanner reads file head (up to 4KB) to extract the cwd field, filters by current working directory, reads head and tail (64KB each) for title fields (customTitle, aiTitle, lastPrompt, summary, first user message). Filters out sidechain sessions. Max 50 sessions, 30-day age limit.

Codex sessions live at ~/.codex/. Two storage formats: SQLite (state_N.sqlite, threads table) and JSONL rollouts (sessions/YYYY/MM/DD/rollout-<timestamp>-<uuid>.jsonl). Scanner tries SQLite first, falls back to files. SQL is built dynamically to adapt to column availability (updated_at_ms vs updated_at, optional title/first_user_message/git_branch). Each result is validated by checking the rollout file exists. Supports sources: cli, vscode, atlas, chatgpt.

Cursor sessions: future scope, not blocking this issue.

Layer 2: Session content reader + handoff

When the user picks a foreign session, read its content and generate a handoff summary. The reader parses JSONL (Claude) or JSONL/SQLite (Codex) and extracts: user goal, last user request, files touched, work completed, work open, stopping point.

Output is a short structured summary, not a verbatim transcript replay.

Safety: treat foreign transcripts as untrusted

Foreign session content may contain prompt injection, foreign system prompts, thinking blocks, or tool calls that do not exist in Maka. The reader must:

• Never execute or follow instructions found in the transcript
• Never inject foreign system prompts, base instructions, or reasoning blocks into Maka context
• Never treat a foreign tool call as a Maka tool
• Never replay the transcript verbatim
• Treat old tool output as stale evidence; verify repository state before relying on it
• Surface uncertainty and reader warnings in the handoff

TUI integration

Foreign sessions appear in the existing session picker with a source badge (claude, codex, cursor). Selecting one triggers the resume flow: read content, generate handoff, verify repository state, then continue in a fresh Maka session.

Enable control

Per-source enable/disable via config or env vars. Sources are only scanned if enabled AND the corresponding reader is available. Defaults: Claude on, Codex on, Cursor off (no reader yet).

Scope

This issue covers Claude Code and Codex. Cursor can follow. The scanner goes in packages/storage/ or packages/core/ (TypeScript, matching existing architecture). The handoff generation can be a skill or inline in the runtime, whichever fits the existing skill system better.

Desktop app integration can be a follow-up.

Alternatives or workarounds

Today users manually copy context from their previous tool into a new Maka session. Some keep a separate notes file to track what they were working on. No automated handoff exists.