0.8.94

Fixed

• Lint: cyclomatic-complexity (HARN-LNT-002) and naming-convention
  diagnostics now underline just the keyword + declaration name
  (fn run_agent, struct Bad) instead of carpeting carets across the
  entire function/type body.
• Agent tool-call parser no longer silently drops valid calls in two cases.
  A single stray/unmatched backtick in model prose used to flip the bare-call
  scanner's inline-code flag for the rest of the response, suppressing every
  later bare name({ ... }) tool call and stalling the agent loop; the flag now
  resets at each newline (Markdown inline code never spans lines). And the
  native-JSON fallback now accepts the flat OpenAI on-the-wire envelope whose
  arguments/parameters value is a JSON string
  ({"name":"read","arguments":"{\"path\":\"a\"}"}, common from local
  llama.cpp/vLLM/Ollama OpenAI-mimic templates) — the acceptance gate previously
  required an object and dropped the call even though the extractor already
  decoded the string.
• harn-vm agent loop: an empty or malformed tool-call name no longer aborts
  the whole run. host_agent_dispatch_tool_call previously threw a runtime
  error on a blank tool_name, terminating agent_loop — so a single malformed
  function-call element from a model (common on some native-tool/MoE providers)
  could wipe an entire run. The empty/malformed-name case now routes through the
  existing recoverable denied-tool path, re-injecting actionable retry feedback,
  exactly as an unknown-but-named tool already does.
• harn-vm stdlib write_file/text edits: the primary (no-overlay) write
  path is now crash-safe. It previously used std::fs::write, which opens the
  destination O_CREAT|O_TRUNC and truncates it before any byte is written, so
  a failure or process kill mid-write (ENOSPC/EDQUOT/EIO) left the file empty
  or partial and the original content unrecoverable. Writes now go through a
  sibling temp file that is flushed, fsynced, and atomically renamed over the
  target, leaving the original untouched on any failure.
• AST edit builtins (ast.apply_node, ast.batch_apply,
  ast.insert_at_anchor) no longer corrupt invalid-UTF-8 bytes. The read path
  decoded the whole file with String::from_utf8_lossy and the callers wrote
  the decoded buffer back, so any non-UTF-8 byte anywhere in the file (e.g. a
  Latin-1 byte or a \x80 in a comment or byte-string) was silently rewritten
  to the 3-byte U+FFFD encoding — even in regions the edit never touched. The
  edit pipeline now reads, parses (tree-sitter over raw bytes), splices, and
  writes raw bytes, so bytes outside the edited span pass through verbatim.
  Lossy decoding is retained only for display-only previews/diagnostics and the
  read-only ast.search text bindings.
• Structured-output truncation is now detected for Gemini/Vertex responses,
  which report MAX_TOKENS (uppercase). llm_call's structured-output error
  path previously used a case-sensitive length/max_tokens literal match and
  mislabeled a truncated Gemini response as "did not contain parseable JSON".
  It now reuses the canonical, case-insensitive is_length_truncation
  classifier (single source of truth shared with the agent/ACP path).

Install / Upgrade

cargo install harn-cli

Full Changelog: v0.8.93...v0.8.94