Session History Mining for Claude/Codex/Cursor

Use this when you want an AI agent to study recent Claude/Codex/Cursor work before proposing a plan.

This is optimized for:

• cross-project history review
• low-noise context transfer
• token efficiency (only new or condensed context)

What to Use

Use Flow's cross-project session browser:

f sessions

f sessions scans Claude + Codex + Cursor sessions across projects, lets you pick one, and copies context to clipboard.

Core Commands

# List sessions across all projects without interactive selection
f sessions --list

# Only Claude sessions
f sessions --provider claude --list

# Only Codex sessions
f sessions --provider codex --list

# Only Cursor sessions
f sessions --provider cursor --list

# Copy selected session context (interactive picker via fzf)
f sessions --provider all

# Copy only the last N exchanges
f sessions --provider all --count 8

# Ignore checkpoints and copy full session
f sessions --provider all --full

# Produce a condensed handoff summary (requires Gemini key)
f sessions --provider all --handoff

Checkpoint Behavior (Important)

Default f sessions copies context since last consumption checkpoint for the current repo.

That means repeated runs do not keep re-copying old context.

Checkpoint file:

.ai/internal/consumed-checkpoints.json

Use --full when you explicitly want full history instead of incremental context.

Current-Repo Deep Pull (When Needed)

If you need more detail from a known session in the current repo:

f ai claude list
f ai codex list
f ai cursor list

# Copy the last 6 exchanges from a selected Claude session for this repo
f ai claude context - /absolute/path/to/repo 6
f ai cursor context - /absolute/path/to/repo 6

Use this after f sessions when you want to zoom in on one thread.

Efficient Workflow (Recommended)

1. In the target repo where you want the plan, run: f sessions --provider all --list
2. Pull 2 to 4 high-signal contexts: f sessions --provider claude --count 6 f sessions --provider codex --count 6 f sessions --provider cursor --count 6
3. For stale/long sessions, prefer condensed transfer: f sessions --provider all --handoff
4. Paste each copied output into labeled blocks in your prompt.
5. Ask for a plan with explicit constraints and ranked execution order.

Prompt Scaffold (Attach This)

Use this format when asking an agent to mine history and propose execution:

I have ~$500 of Claude tokens expiring in <N> day(s) and want to use them efficiently.

Goal:
- study goose and propose a concrete execution plan for token usage
- use ideas from recent Claude/Codex/Cursor histories
- rank ideas by expected impact and execution cost

Constraints:
- avoid low-signal exploration
- maximize useful output per token
- include exact next commands I should run

Session context 1:
<paste from f sessions --provider claude --count 6>

Session context 2:
<paste from f sessions --provider codex --count 6>

Session context 3 (optional handoff):
<paste from f sessions --provider all --handoff>

Session context 4 (optional Cursor):
<paste from f sessions --provider cursor --count 6>

Deliver:
1. top opportunities (ranked)
2. 48-hour execution plan
3. fallback plan if one assumption fails
4. specific commands and owners

Token-Efficiency Rules

• Prefer --count over --full unless you are reconstructing full intent.
• Prefer --handoff for large stale sessions before pasting into expensive models.
• Cursor transcripts use file-modified time rather than per-message timestamps, so repeated copies may include the whole latest transcript after a new edit.
• Merge duplicate context manually before sending to avoid repeated tokens.
• Request ranked outputs with hard deliverables (plan, commands, owners, fallback).

Troubleshooting

• fzf not found: install fzf, or use --list and then run interactive once fzf is available.
• No new context copied: expected if checkpoint is current; rerun with --full.
• --handoff not working: set GEMINI_API_KEY or GOOGLE_API_KEY.