Sources
057 reads state that already exists on your machine. Every source is optional, every source is read read-only, and none of them is modified, started, configured or made aware that 057 exists.
All available sources contribute to one world at once. Which of them you show is a filter on a surface, never a different world.
The one source that needs nothing installed. A snapshot of the process table every five seconds: what is running, when the kernel says it started, CPU time and resident memory as they change.
It is a sampler, not a kernel hook. An exit is never witnessed — it is inferred from a process being there in one sample and gone from the next, and everything derived from that says so. A process that starts and finishes inside one interval is never seen at all. Command names are recorded; full command lines are not, because they can carry a prompt, a key or a private path.
Sessions, turns, token usage, tool calls and failures, model, and — where the runtime recorded it — which process a session ran in. Detected when ~/.claude exists; contributes once there are sessions or transcripts under it.
Sessions, turns and their outcomes, model and effort per turn, token usage, and typed items such as command executions, MCP tool calls and file changes, with exit codes and durations. Detected when ~/.codex exists; contributes once there are session records under it. Sessions driven through the CLI, the desktop app or the editor extension all land in the same place.
IDE-hosted conversations, their turns and tool calls, the workspace they belong to, edit counts, and subagent parentage — which conversation started which. Detected when Cursor’s state database is present in its application-support directory; contributes once that database holds conversations. It is opened read-only, and never blocks the IDE.
Sessions, turns, token usage, model and tool calls. Detected when ~/.gemini exists; contributes once there are session files under it. No process identity — Gemini records none, and none is invented.
Sessions and turns, run outcomes, token usage, tool calls and failures, sub-agent parallelism, and time spent blocked waiting for a person — which is read from the harness’s own approval records rather than guessed. Detected when ~/.dsh exists; contributes once there are session logs under it.
Tasks and their runs, read from the .work directory inside a workspace. 0x2F is an externally observed system: it need not be running, and 057 integrates from its own side, from files 0x2F already writes.
There is no machine-wide registry of 0x2F workspaces — a workspace is any directory with a .work in it. With nothing set, 057 scans your home directory one level deep, which is where checkouts usually are. Point it at yours with WORK_ROOTS, colon-separated. A root that does not exist is skipped silently; a root that exists and cannot be read is a failure and is reported as one.
A source is in exactly one of four states, and the difference between them is the point.
Not installed is not broken, and 057 will not report it as broken. A machine with none of these runtimes is a valid mode: it has a process table, 057 has something true to say about it, and the header says OS ONLY rather than announcing six failures.
The other half of that bargain is that a real failure is never quiet. Collection outcomes are tracked per source over successive rounds. One failed round is normal — a file being rewritten while it is read — and says nothing. Several in a row makes the source failed, and the header shows SOURCE FAILING, names it, and shows the error. An absent runtime can never trigger it.
Between those there is reduced depth: a source that can be partly read shows as PARTIAL, with what was reduced. The computation is still observed; some of its semantics are not.
A failing source costs you that source and nothing else. Not the other sources, not your OS history, not the store.
Runtimes differ in what they record, and 057 never levels that out. A measure a source does not report is shown as missing, not filled in. Anything computed rather than read is labelled with the assumption it rests on — hover it in the interface to see the assumption. There is no cost figure anywhere: sources report tokens and CPU time, none of them reports money.
057 reads files these runtimes never promised anyone. When one changes its format, 057 reports the read as failed and names the file rather than producing plausible nonsense from it. A source can therefore stop working after an unrelated upgrade to that runtime — expected during the experiment, and visible when it happens rather than silent.
Switch the header control from CAPTURE to SYNTHETIC for a deterministic replay of a made-up day. It is labelled REPLAY while it runs and is never mixed into real telemetry — generated activity sitting alongside observed activity is the one thing this instrument must not do.