ancha pull

ancha pull writes a test your workspace saved from a production trace into scenes/from-traces/<name>/, where ancha run finds it. Name a trace that has no saved test yet and ancha pull saves one first. With --answer, it saves a new version with your answers to the test’s questions, then writes that.

ancha pull [options] <id>

Argument	Description
`id`	A test or a trace: its id, its first 8 characters, or the trace id your tracing tool shows

Option	Description
`--dest <dir>`	Where the tests go (default: `scenes/from-traces`)
`--answer <question>=<value>`	Answer one of the test’s questions, saving a new version first. Repeat it for more.
`--expect <what went wrong>`	Say in a sentence what the agent did wrong in production, saving a new version with a check for it.
`--force`	Overwrite `scene.yaml` even if it has changes besides checks you added

Files

File	Holds
`scene.yaml`	The test: what the agent is asked, its tools, starting data, what happens during the run, and the checks. Run it with `ancha run`.
`observations.jsonl`	Every call production made to the agent’s own tools, with its answer.
`production.jsonl`	Production’s run: its input, its tool calls and its answer.
`production.json`	What production did on each check: `passed`, `failed` or `unknown`. `ancha run` computes the unknown ones from `production.jsonl`, also for checks you add, and compares with it.
`questions.json`	What one trace cannot settle, each with the default the test uses.
`provenance.json`	Every part of the test, and whether it was seen in the trace (`observed`, with its span ids), `inferred`, or `missing`.

When your ancha.yaml lists scenes: globs that do not include the new test, ancha pull prints a note with the glob to add.

Answers

ancha traces show lists a test’s questions, each with how to answer it:

Question	`--answer` value
`task`	What the agent is told, in your words. You can always answer it, even when the trace said it and nothing asks.
`tools.mcp`	The tools your agent reaches over MCP, comma-separated, or `all` or `none`
`tools.simulate_writes`	The writes a test may simulate when your agent makes one production never made, comma-separated, or `all` or `none`
`checks.success`	What the agent should have done, in your words, when nothing in the trace says what went wrong
`deploy-…`	Seconds until production has a merge, such as `60`, or `none` when nothing deploys
`recovery-…`	`floor=<earliest second the problem can end>,recovery=<seconds it then takes>`; either part alone works
`fix-…`	`undo` when undoing the change the question names ends the problem, or `self=<second>` when it ends on its own

A value can also be the JSON the console sends, such as '{"floor":140,"recovery":10}'. Answers add to the test’s earlier answers; <question>=default drops one. Questions with no --answer, such as issues your agent never read, are answered in scene.yaml.

Pulling again

Pull a test again to get a newer version. Checks you added to scene.yaml by hand carry over to the new version. Any other change to scene.yaml would be lost, so ancha pull stops before it saves or writes anything and says so; move the change into a check of your own, or pass --force to overwrite the file. ancha pull tells the two apart with the # ancha pull: line at the top of scene.yaml, so keep that line.

Example

ancha pull fc7f4f10 --answer tools.mcp=get_deploy_status,rollback,list_issues --expect "it declared the incident resolved on a status measured before the rollback"
ancha pull 4735f9ca
ancha run scenes/from-traces/trace-4735f9ca/scene.yaml
# production deploys in about a minute, and the problem cannot end before 140 seconds:
ancha pull 4735f9ca --answer deploy-repo_orders=60 --answer recovery-problem=floor=140,recovery=10

The run ends with a comparison against production, check by check. See From a production trace to a failing test.

Get started

Guides

Concepts

Tools

CLI

Reference

Cloud

Help

Files

Answers

Pulling again

Example

​Files

​Answers

​Pulling again

​Example

Files

Answers

Pulling again

Example