turnal-fork-history

name	turnal-fork-history
description	Rerun or branch from a task recorded by Turnal and compare the resulting attempts. Use when asked to retry a past agent turn, explore an alternative implementation from historical state, run A/B attempts, inspect fork readiness, compare or select Case attempts, or apply a chosen attempt. Covers source-turn discovery, isolated execution, Case/Attempt/Task identifiers, and exact-base application.

Fork Turnal history

Fork from a captured pre-turn checkpoint into a supervised isolated workspace. The child does not run in the source workspace, but execution writes durable Case and Attempt records to the Turnal store. Read-only history questions belong to $turnal-inspect-history; reach for a fork only when something must be rerun.

Find the source

1. Run turnal status and confirm the workspace has Turnal history.
2. Use turnal sessions --json, turnal log, or turnal search --json "<query>" to obtain a real session ID and turn number.
3. Inspect the source with turnal show <session>:<turn> and turnal diff <session>:<turn> when the user's intended base or task is unclear.
4. Copy every identifier from Turnal output.

Read references/fork-arguments.md before constructing the fork command. It defines accepted targets, the -- boundary, instruction replay, and JSON stream behavior.

Preview readiness

Always begin with a non-executing readiness check:

turnal fork <session>:<turn> --dry-run --json

Inspect the pre-turn base, captured file count, instruction status, fidelity, and limitations. When the instruction is missing or redacted, ask the user for a prompt — raw storage is not a recovery path for it. A dry run accepts exactly one target and no child command.

Execute an attempt

Run only after the user has asked to retry, branch, or experiment:

turnal fork <session>:<turn> -- codex exec

Place Turnal flags before -- and the complete child command after it. Bare codex and codex exec commands receive the captured instruction automatically. If the child command already includes an explicit prompt, Turnal leaves it unchanged. Use --no-replay-instruction only when the child should not receive the capture, and provide an explicit prompt when needed.

Use --keep only when the isolated workspace must survive for inspection. With --json, parse stdout as the Turnal result and treat child output on stderr separately.

Record result.case_id, result.attempt_id, and result.run_id from JSON, or the case:, attempt:, and run: lines from human output. Read references/cases-and-ids.md before creating sibling Cases or using durable IDs.

Compare and choose

1. Run turnal compare <case-id> --json to compare every completed Attempt with the same immutable base.
2. Request a full patch only for a candidate: turnal compare <case-id> --patch <attempt-id>.
3. Weigh command status, file changes, captured limitations, and frozen verifier results together; patch size alone decides nothing.
4. Record the choice with turnal select <case-id> <attempt-id> --json when the user asks to select an Attempt.

Apply only with explicit authorization

Applying changes the live workspace. Do not treat a request to compare, recommend, or select as permission to apply.

Preview first:

turnal apply <case-id> --dry-run --json

Run turnal apply <case-id> only after the user explicitly asks to apply the selected Attempt. Use --attempt <attempt-id> to override the recorded selection only when the user identifies that Attempt. Apply is exact-base only: it refuses if the captured workspace surface differs from the Case base, creates a safety checkpoint before restoring, and never performs a three-way merge.