orgx trail

See what your coding agents actually did, stop them relearning the same walls, and prove the fix worked. Reads your Claude Code, Codex, OpenCode and Cursor history into threads of work, finds the walls they keep rediscovering, and writes the fix where they'll read it.

npx @useorgx/trail

• Local. Reads ~/.claude/projects, ~/.codex/sessions, OpenCode's database and Cursor's chat stores, read-only, on your machine. Nothing is uploaded unless you run trail sync.
• Fast. ~3,300 sessions (38 GB) in about 40 seconds; after that, only new work is read.
• Zero dependencies. Small enough to read before you run it.

What you can do

trail	read new work, then open the explorer (overview · walls · threads · sessions · quality)
trail walls	the walls your agents keep hitting, each with a fix (--json for agents)
trail copy <id>	copy a fix: a prompt your agent can act on (default), the rule, or the command
trail adopt <id>	write the fix into AGENTS.md / CLAUDE.md as a marked block (trail unadopt <id> removes it)
trail guard install	prevention: in don't-ask runs, stop a known wall before the agent walks into it
trail card	your trail as a shareable image and post text
trail share <id>	a public page for a fix, measured across everyone who adopted it
trail experiments	did your AGENTS.md / CLAUDE.md edits change agent behavior? (95% intervals)
trail bench	would a model walk into your known walls, with and without your rules?
trail mcp	trail as tools for your agents: claude mcp add trail -- npx -y @useorgx/trail mcp
trail open · trail watch	ledger view in your browser (localhost) · follow the live session
trail connect	sign in to OrgX through @useorgx/wizard; trail never handles your password or key
trail credits	the people whose work trail is built on
trail sync	send thread outlines (never transcripts) to OrgX; --dry-run shows exactly what would be sent

In the explorer, on a wall: c copies a prompt for your agent, r the rule, x the command, a adopts it.

How far to trust it

• Facts — denials, failures and ships — are read straight from the transcript.
• Judgments — intent, discoveries, whether a thread was dropped — are inferred, and the Quality tab says how well. "Dropped" comes from a small tree model (src/model/abandon.json); your own labels are the real test.
• Effects are compared like with like: same repo, same client, same permission mode. When the permission mode changed around a fix, trail says "confounded" instead of claiming a win.
• trail bench measures a model's planned first calls, not executed runs, and says so.

Built on

Trail is built on other people's ideas. None of them endorse it; this is what we took from each. trail credits shows the longer version.

Work	What trail took from it
Clio: privacy-preserving insights into real-world AI use · Alex Tamkin and the Clio team at Anthropic, 2024	A fix’s public numbers appear only once at least 5 separate people have adopted it, and uploads carry counts, not words (thread titles only if you opt in with --with-titles).
Error analysis for AI systems (open coding → axial coding → count → judge) · Hamel Husain and Shreya Shankar, 2025	The walls are failure types counted across real sessions, and trail’s classifier is checked against human labels in a labeling lab. (We wrote our codebook before our notes, which they warn against; the next labeling pass starts from notes.)
Sniffly: a dashboard over your local Claude Code logs · Chip Huyen, 2025	Lead with one surprising number about your own agents, found locally.
Docent: searching agent transcripts against a rubric, with cited evidence · Transluce, 2025	Every adopted fix is reported as a measured before and after, not a claim.
Measuring AGENTS.md: what five runs show that one doesn’t (AAIF) · Andrea Griffiths, 2026	trail experiments reports intervals and says “no detectable change” when the interval spans zero. We learned the same lesson the hard way: our first before/after was confounded by a permission-mode switch.
Measuring the impact of early-2025 AI on experienced open-source developer productivity · Joel Becker, Nate Rush, Beth Barnes and David Rein (METR), 2025	trail compares like with like (same client, same permission mode) and marks a result confounded instead of reporting it.
AGENTS.md: a simple, open format for guiding coding agents · The AGENTS.md contributors (now stewarded by the Agentic AI Foundation), 2025	Fixes are written where agents already look, as a removable block in AGENTS.md or CLAUDE.md.
Entire: agent checkpoints stored in git, next to the code · Thomas Dohmke and the Entire team, 2026	Lessons live in the repo, in files a team reviews like any other change.
Agent Trace: an open, vendor-neutral spec for AI code attribution · Cursor, 2026	trail’s upload format (orgx-trail-threads/v1) is a short, versioned contract you can inspect with --dry-run; publishing it as an open spec is next.
ccusage: token and cost analysis from local agent logs · ryoppippi and the ccusage contributors, 2025	npx, no account, nothing uploaded, a first answer in about 40 seconds.
Portable Game Notation (PGN) · Steven J. Edwards, 1993	Each thread is a move string (probe, run, change, check, ship, failed, denied) you can read at a glance and compare.
Stigmergy: coordination through traces left in the environment · Pierre-Paul Grassé, 1959	A wall one session hit becomes a trace the next session reads before it starts, instead of every session starting from zero.

Data

Everything lives in ~/.orgx/trail (override with TRAIL_HOME). Delete that folder to forget it all.

Built by OrgX.