Skip to content

v1.25.0

Choose a tag to compare

@github-actions github-actions released this 14 Nov 21:14
· 9 commits to main since this release
v1.25.0
018de75

This release keeps long agentic tasks in the REPL within the model's real context budget, tells the agent when files change under it, and fixes several REPL bugs found in live runs.

New

  • File change tracking. fuseraft keeps track of the files the agent has read or written. If one of them changes outside fuseraft, or while a tool runs, the tool's result names it. In the REPL, changes made while the agent waited for you are noted after your message. write_file refuses to overwrite a file that changed since the agent last saw it, until the agent reads it again. The new files.trackedFiles setting (default 50) sets how many files are tracked, and 0 turns tracking off, which can help on network shares. (#223)

Changed

  • In-turn budgets follow the provider's own token count. The REPL now scales its in-turn budget by the prompt size the provider reports, not just fuseraft's character estimate. On grok-4.7, billed prompts ran 1.5–2x that estimate, because xAI bills earlier rounds' hidden reasoning as input. Tool results also go to the model as plain text instead of escaped JSON. (#225)
  • grok models default to a 256K-token REPL budget (was 150K). When trimming is needed, older tool results are dropped 16 rounds at a time, so the prompt cache misses less often. (#225)
  • Long single turns can be compacted. Auto-compaction retries every turn while the context is over its threshold. It can now split one long turn at a tool round, keeping your request word for word. The summary is now worded as the same conversation, not a handoff to a new session. compact_context is no longer a default tool. (#225)
  • Reading guidance. The REPL prompt, read_file, and the fuseraft init templates now tell the agent to read a file whole when it fits in one call, and to reuse an unchanged earlier read instead of re-reading it in small slices. (#225, #226)

Fixed

  • The agent no longer writes context-trimming placeholders ([ELIDED …]) into files, todo lists or commits. todo_write and git_commit refuse them. (#225)
  • shell_run re-runs a repeated command if any other tool call came in between. Before, a test run after a fix could return the old failure. (#225)
  • A command that times out now returns what it printed before it stopped. A long truncated log keeps the failure lines from its cut middle. (#225)
  • A reply that repeats itself without end is cut off, and the text before the loop is kept. An answer repeated after a todo-list update is now kept only once. (#225)
  • A dropped reply no longer takes the turn's tool work with it, so its tools aren't run again on the retry. (#222)
  • An answer quoting a tool error in a code block is no longer dropped as leaked tool text. (#222)
  • Where-is questions work on Anthropic thinking models that refuse a forced tool call. (#222)
  • repl --vscode --model starts with a config file that has settings but no model. (#222)
  • The context indicator counts cached tokens on Anthropic. (#225)

Docs

  • The fuseraft init template list now includes greenfield, and each template's description matches what it generates. (#226)

What's Changed

  • fix(repl): keep a dropped reply's tool work; start and answer in more setups by @fuseraft in #222
  • feat(files): report files that change under the agent; refuse stale rewrites by @fuseraft in #223
  • fix(repl): keep long agentic turns inside the model's real budget by @fuseraft in #225
  • fix(init): read files whole when they fit; describe templates as generated by @fuseraft in #226

Full Changelog: v1.24.0...v1.25.0