Five stages, a gate on every one, and every claim cited or refused. Runs from your terminal or from Claude Code.
You ask an AI to build something. It goes away, does a lot, and comes back with a summary you have to take on faith. Later you cannot tell what it read, what it assumed, what it cost, or who agreed to any of it.
tldrx puts a shape around that. Work is broken into five stages — What → How → Plan → Build → Watch — and each one ends at a gate. A gate is a stop: nothing after it runs until it is signed. Some gates you sign; some the tool may sign for itself, but only when it can show its work.
Everything a stage produces is a file in your repo — the intent, the design, the plan, the questions it could not answer, the answers you gave, the money spent. Those files are not a report about the state. They are the state: the tool re-reads them to decide what to do next. The canonical state is on disk — inspectable, diffable, committable, recoverable.
1 — Open a piece of work
tldrx run new payments --scope feature --budget 25 writes a folder. Nothing has run yet and nothing has been spent.
2 — Run one stage
tldrx next runs the stage the cursor is on, writes its files, and stops at the gate with exit code 4 — "the work is done, the decision is yours."
3 — Sign it, or send it back
tldrx approve --note "why" re-runs the stage's checks against what is on disk and moves the cursor on. tldrx reject --note "…" sends it back, and the next attempt reads your note.
Repeat until the run is finished. tldrx run auto does the repeating for you and stops the first time something genuinely needs a person.
The three steps above are you at the keyboard. The other case is the demanding one: you hand a run over for the night, inside boundaries you wrote down — which gates an agent may close, and what the whole thing may cost — and what comes back is not a summary you have to believe. It is the run itself: the files each stage wrote, the evidence note under every signature, the money it actually spent, all still on disk and readable in any order you like.
tldrx init # detect, map, and probe your build/test commands. $0.00
tldrx run new payments --scope feature --budget 25 \
--attended-by host --gates what:agent,plan:agent,build:agent,watch:agent
tldrx drive --unattended # the mandate — paste it into a Claude Code sessionThat last command opens no run and spawns nothing: it prints the mandate the session drives from. The mandate is what keeps the night honest — a developer sub-agent, then a fresh reviewer that is never the author, a gate signed only over a written evidence note, and the four things it has to wake you for instead of deciding them.
Attended or unattended — the three ways to run a stage, which to pick, and what each one costs.
Beta, version 0.16.0. Every command is real and tested — the docs on this site were written by running them — and tldrx --help on your machine is the authority, not this site.
The bar for beta was public and it has been cleared: file formats frozen (version: 1 schemas only grow), two or more real workspaces taken through the Build phase, and a documented upgrade path. Releases through 0.3.1 were alpha; 0.4.0 was the first beta. Stable means 1.0 and semver from then on.
Install name is tldr-experts; it gives you two commands, tldrx (short) and tldr-experts (the same binary).