Replopencode v1.18.32-11

Stops apply_patch from running a sandbox out of memory when it deletes a large binary file (#71, REPL-34133). Deleting a binary file, or any file over 1 MiB, now records only its size in tool metadata instead of a full content diff. Before, a GPT agent deleting a 30 MiB video produced about 347 MiB of JSON per copy of the diff and was OOM-killed at the 4 GiB sandbox limit. Small text deletes keep their full diff. The TUI no longer shows "-0 lines" for these size-only deletes.

Replopencode v1.18.32-10

Stops a turn from dying with The final block in an assistant message cannot be thinking`` after the provider cuts a step short (#70). When a step finished tool-calls with no tool call and no text, carrying only Anthropic's "This part of the response was interrupted before it finished." reasoning placeholder, the loop replayed that message as a thinking-final prefill, which every provider rejects with a deterministic 400, so the chat showed "Something went wrong with this turn" even though the agent's edits had landed. Such a cut-off step is now removed and run again; a second cut-off in the same turn fails the assistant message with `MessageOutputLengthError` and publishes `session.error`, so the client shows the turn as failed instead of looping.

Replopencode v1.18.32-9

Stops a repeated tool call id from breaking a session for good (#68). A late tool input event for a call that already settled no longer creates a duplicate tool part, and history replay sends each call id once per message, so sessions that already hold a duplicate recover on their next turn.

Replopencode v1.18.32-repl32473.2 (PR #69)

Opt-in test build for REPL-32473, from PR #69.

Preserves saved unanswered questions across restarts and settles reply, dismissal, and Stop correctly. This revision simplifies the implementation by 47 production lines, adds negative coverage for stale/terminal question state, and removes 15 lines overall.

Validated with prompt, recovery, question lifecycle, and HTTP regression tests; package and monorepo typechecks; four binary builds; SDK generation; and repeated hard-restart HTTP smoke tests.

This is a prerelease for explicit installation only. It does not change the default stable release or deploy to the Replo fleet.

Replopencode v1.18.32-repl32473.1 (PR #69)

Opt-in validation build of #69 for REPL-32473.

Preserves unanswered persisted questions across process restarts and settles reply, dismissal, and Stop correctly. Includes Linux x64, baseline x64, ARM64, macOS ARM64, SHA256SUMS, and the SDK built from the same source.

Verified with regression tests, monorepo typechecks, and compiled-binary HTTP smoke tests spanning repeated SIGKILL/restart cycles. This prerelease does not change the latest stable version.

Replopencode v1.18.32-8

Code Mode: a plugin can rank execute's catalog (pinned tools) and index the rest by category (#66). Without plugin options the catalog is unchanged.

Replopencode v1.18.32-pr28290.6

Test build for replohq/andytown#28290: Gabe's prompt-cancel change (#67, same commits as #65) on top of current dev, which includes #66. Supersedes v1.18.32-pr28290.5, which predates #66.

Replopencode v1.18.32-pr28290.5 (PR #65)

Preserve prompts admitted at normal completion and protect explicit resumes from stale prompt aborts. Preserve ordinary idle notifications. Opt-in build for andytown PR anomalyco#28290; stable latest unchanged.

Replopencode v1.18.32-pr28290.4 (PR #65)

Retain cancellation ownership for every queued prompt. Preserve canceled-message exclusions across replacement handoff and superseded preparation. Opt-in build for andytown PR anomalyco#28290; stable latest is unchanged.

Replopencode v1.18.32-pr28290.3 (PR #65)

Isolate prompt cancellation from unrelated active model/shell work, resume admitted replacements, clean up settled ownership, and preserve direct processor interruption notifications. 149 tests passed, one existing skip. Opt-in prerelease; stable latest unchanged.