[malihashar]

In short

txcript check is a new command that tells you what converting a session to another agent would lose, before you do it. For example, txcript check <session> --with codex lists what Codex would drop or change. This is step 3 of 3: the command itself, plus a one-line warning on continue when a conversion is lossy.

Part 3 of 3 for #92. It builds on #94, which builds on #93.

How to review this PR

This branch also contains the commits from #93 and #94 until they are merged. Please review only the last commit, feat(cli): add txcript check, a report of what a conversion loses. After the earlier PRs merge, this PR will show only its own changes.

What this PR adds

The txcript check command, which uses the library from #94.

txcript check <id|file|->[#range] [--with <harness>]... [--from <harness>]
              [--json] [--detail] [--fail-on none|content|accounting|metadata]

• With --with, it prints a full report for each target. Without it, it prints a table with one row per writable target.
• --json prints a stable, versioned format (schema 1). --detail shows longer examples.
• --fail-on makes the exit code 1 when the loss reaches that level, so CI can refuse a lossy conversion.
• check never writes anything and never starts an agent.

New flags on the existing commands:

• continue and crop print one line on stderr when the chosen target loses or changes something. --quiet hides it.
• continue --strict and crop --strict stop before writing anything if content would be lost.
• continue --dry-run does the same as check.

Example

$ txcript check session.json --with codex
check 8d116ece (simple, 23 messages) -> codex
  content    CHANGED  1 tool name renamed  /release -> _release
             LOST     1 tool argument (Bash.description)  Run tests
  accounting LOST     4 usage records
             LOST     8 stop reasons  tool_use
  metadata   LOST     title  Session everything
             LOST     model  claude-opus-5
  structure  23 messages become 30 (the target splits turns)
  codec-level: measured on the target's session text; a store's own additions are not counted

What changes for existing users

• Stdout of every existing command is exactly the same. A test checks this.
• continue and crop print one extra line on stderr when a conversion is lossy. That is the only visible change. Resuming in place (same harness) prints nothing.
• SessionCommand::Continue and SessionCommand::Crop gained fields, so code that builds or destructures them by hand needs .. or the new fields.

Tests

• 36 end-to-end tests run the real binary: the report, the table, --json, --fail-on exit codes, --strict writing nothing, check and --dry-run writing nothing, ranges, files and stdin, the stderr note appearing once, a closed pipe.
• Nine of them check the report against reality. For codex, pi, campfire, cursor, cursor_desktop, grok, fx, antigravity and claude_code, the test converts for real, reads the copy back, and checks that what check said was lost is what really disappeared.
• 15 unit tests for the command, and the whole suite passes. cargo fmt and cargo clippy (0 warnings) are clean.

Hands-on testing

I also used it for about 30 minutes on synthetic sessions, from empty to 18 MB, with every agent as source and target and a 9-step chain of conversions. That found 10 false alarms, all fixed with a test each, and no missed losses. Speed: a 5,000-message session takes about 0.6 seconds for all targets together, and an 18 MB session about 1.6 seconds. Only a 363 MB stress session took over a second per target.

What I could not test

Real agent resumes, OpenCode and Cowork writes, live sources (ChatGPT, Claude Chat), and the interactive crop editor with --strict (it needs a terminal; the check it uses is unit-tested and shared with continue).

Open questions

1. Should the stderr note always print, or only when stderr is a terminal? Today it always prints, because continue already prints similar lines.
2. Should --strict refuse a conversion that only renames a tool argument, or only when text, tool calls or results are lost?