What's new in 0.1.0-rc23

Your project keeps working as it is, and your earlier records keep their meaning. In this release, every agent session that commits work keeps a trace record, and one command does the record’s routine before each commit.

What changed	What you notice
A record for every agent commit	The shared instructions now say it outright: a session in which an agent commits work keeps a trace record. A commit you type yourself owes none. See Record a session.
One command before each commit	Your agent runs the record’s routine in one command. It finds the running Claude Code or Codex session’s own transcript, brings the record up to date, screens it, records the disclosure review, moves it to ai/records/, checks it and drafts the commit body with its Record: line and execution marker. A later commit of the same session updates the same record, without renumbering a turn or rewriting a choice or check already in it.
What that command leaves alone	It writes no summary, choice or check, and it stages and commits nothing. It clears the disclosure review only when the agent declares it has read every captured turn and the local screen finds nothing. It leaves a turn out only with a stated reason, or when another promoted record already summarizes that turn, which it names. When something is missing it says what and moves nothing, and nothing it reports stops a commit.
Your name and the model’s	On a session’s first record, you are named from your Git name (git config user.name), unless that reads as a model’s name, and the record says how it knew. The model in the execution marker is read off the session’s own transcript; a model id the agent gives that the session did not report is marked self-reported in the record and the commit body.
What waits for you	The overview of waiting work now lists the commits that name no trace record, each with its subject and who executed it (the first ten, and a count of the rest), where it used to count them. See Accept, return or refuse work.
Review against sources	A reviewer’s view of a trace record names the commits under review and, separately, every commit that names the record, so a reviewer no longer reads the record’s other commits as missing. See Review work.
Stated for every agent	The shared instructions now state what was left to each model’s judgment: no em dash anywhere, chat replies included; who you are (your Git name, unless you named yourself otherwise, and the agent asks when another name for you is in view); and what to do when a sandbox refuses a commit. See Commit messages.
Your earlier records	Records, registries and ledger rows written by an earlier release keep their meaning.
The release file	trace-os-0.1.0-rc23.zip, with everything under one folder, trace-os/.
Updating	Supported from 0.1.0-rc22, and from earlier releases through a chain.

If you edited the shared instructions

The update replaces the shared section of AGENTS.md, core-workflow, now version 6 because it states the trace-record rule. If your project kept its own version of that section, and declared it compatible with version 5, it is reported unresolved until the declaration names version 6, so the new rule does not pass an edited section silently. Ask your agent to show you what changed, then say whether your version still fits.

After the update, the overview of waiting work names the commits that belong to no record, where it used to count them. Nothing in a record changes.

Limits of this release

• The one-command routine was rehearsed with Claude Code only, with two models and three runs each: every run kept a record, against two of six with 0.1.0-rc22. Codex is covered only by tests on synthetic sessions, and Cline was not run.
• A subagent shares its parent’s session, so the routine reads the parent’s transcript. A subagent gives its own model id, which the record marks self-reported.
• A turn the routine leaves out before the summary cites it stays left out after the summary cites it, an open defect. Cite each turn before the first run, or have your agent remove that exclusion from the staged record.
• Whether a resumed session writes a new transcript file was not checked. If it does, the routine asks for a new record id.
• In the rehearsal, one run lost a note it had written and not yet committed while the workflow check ran in the background, and the cause was not found, an open defect. Commit your work before running long checks.
• Only macOS on Apple silicon with Python 3.12 was run; Linux and Windows were not.

In 0.1.0-rc22

The kit did what each of the five principles asks of a research team, as the authors now state them, and named the principles as verbs: track records, review against sources, accept independently, control transitions and evaluate corrections. Your agent records each consequential choice when it makes it, with why it took it over the alternatives. The table of waiting work shows each record’s owners, labels a record that logs no choice needs you, and your decision covers the version the brief showed you. A review can be given the files to check against, and keeps its report and findings in ai/reviews/. Outside material waits in incoming/ inside the shared folder until someone admits, returns or refuses it. A defect lists every output it reaches and verifies only when each has its own passing recheck.

In 0.1.0-rc21

The shared folder arrived: one shared place for your team’s large and bulky files, in a folder your sync app keeps on every coauthor’s computer, or in a bucket. You say where it is, and the agent connects it, keeps it out of Git and checks everyone’s copy against a list Git keeps. See Connect your shared folder.

In 0.1.0-rc20

The kit was renamed TRACE-OS. It was called TRACEware in 0.1.0-rc19, and the TRACE kit (trace-kit) up to 0.1.0-rc18. A project’s state folder is .trace-os/, and the update moves .traceware/ or .trace-kit/ there, whole; every earlier name is still read until then. After that update, a skills operation makes its rollback refuse, so roll back first if you may want to, then adopt the skill.

See Update TRACE-OS.