v3.0.3

v3.0.3

Changes

• Detect incomplete paginated GitLab diff responses from merge request changes_count and mark AI Review coverage as incomplete instead of treating truncated data as complete.
• Prevent oversized single diff lines from entering AI Review batches, including lines that overflow after a normal context line has already started the chunk.
• Preserve final diff-only coverage metadata when later AI Review batches fail, so synthetic chunk headers are not counted as reviewed source bytes.
• Show the single_diff_line_too_large incomplete coverage reason in Dashboard review coverage details.

Assets

• Windows x86_64: gitlab-work-runner-v3.0.3-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v3.0.3-linux-x86_64.tar.gz

v3.0.2

v3.0.2

Changes

• Split oversized single-file AI Review diffs by unified diff hunks/chunks instead of truncating the file-level diff.
• Added [ai_review].max_tool_rounds; the default is 0 for unlimited context-tool rounds.
• Added clean diff-only finalization fallback when a non-zero tool-round/call/byte budget forces finalization but the model still requests context tools.
• Stopped search_code and list_files repository traversal as soon as their result limits are reached.
• Excluded empty GitLab diff entries from AI Review coverage so they are not reported as unreviewed files.
• Improved Dashboard running progress with current tool round/call counts, moved tool stats below the batch title, and added total run duration to the Review detail page.
• Formatted GitLab manual review summary duration as hours/minutes/seconds instead of raw seconds.
• Added scripts/linux-background.sh to Linux release packages for nohup start/stop/restart/status management of runner and dashboard.

Configuration

• max_tool_rounds = 0 means unlimited context-tool rounds.
• Existing rules.toml files remain compatible. max_tool_calls, max_tool_result_bytes, and max_tool_total_bytes keep their existing meanings.

Assets

• Windows x86_64: gitlab-work-runner-v3.0.2-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v3.0.2-linux-x86_64.tar.gz

v3.0.1

v3.0.1

Changes

• Reduced unnecessary AI context exploration by requiring a concrete candidate defect before repository lookup.
• Added one-based inclusive line ranges to read_file, with a maximum of 250 lines per ranged call.
• Added exact context-tool call deduplication and the per-batch max_tool_total_bytes cumulative result budget.
• Forced final findings submission after a cache hit or when tool-call/result-byte budgets are exhausted.
• Changed context follow-up 504 retries to finalize without context tools; a repeated 504 now enters the existing diff-only fallback.

Configuration

• Added [ai_review].max_tool_total_bytes, defaulting to 40000; set it to 0 for unlimited cumulative tool-result bytes per batch.
• Existing rules.toml files remain compatible. The existing defaults for max_tool_calls and max_tool_result_bytes are unchanged.

Assets

• Windows x86_64: gitlab-work-runner-v3.0.1-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v3.0.1-linux-x86_64.tar.gz

v3.0.0

v3.0.0

Breaking changes

• Removed script tasks. Existing [[script_tasks]] configuration blocks are now rejected and must be removed when upgrading.
• Removed second_pass_on_clean; clean AI Review results no longer trigger an automatic second confirmation pass.

Changes

• Added diff-only fallback when repository archive safety limits are exceeded. Context tools are not exposed in that fallback.
• Added one-shot diff-only fallback for eligible context-assisted AI Review timeout failures.
• Added review execution metadata for context mode, diff-only fallback mode, fallback reason, and context/fallback durations.
• Improved Dashboard Review coverage progress with execution mode badges, batch progress, and clearer running-state display.
• Hid running task finding counts and duration totals until final values are available.

Assets

• Windows x86_64: gitlab-work-runner-v3.0.0-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v3.0.0-linux-x86_64.tar.gz

v2.0.17

v2.0.17

Changes

• Simplified the manual MR review summary by removing batch and context-tool runtime counters from the GitLab comment.
• Hid the user preference section when no manual review preference was provided.

Assets

• Windows x86_64: gitlab-work-runner-v2.0.17-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v2.0.17-linux-x86_64.tar.gz

v2.0.16

v2.0.16

Changes

• Fixed a clippy failure in the skipped-AI-review summary path so release builds pass cargo clippy --all-targets --all-features -- -D warnings.

Assets

• Windows x86_64: gitlab-work-runner-v2.0.16-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v2.0.16-linux-x86_64.tar.gz

v2.0.14

v2.0.14

Changes

• AI review prompts now separate fixed system policy, administrator extra_instructions, low-privilege MR comment scope preferences, and untrusted MR diff content.
• Tool-call mode and JSON fallback mode now generate distinct final-output instructions, avoiding mixed output paths in the same prompt.
• Batched AI review prompts now include the current batch index/count and file count, and MR diff content is wrapped with generated untrusted-data boundaries instead of Markdown fences.
• Removed the configurable system_prompt field; extra_instructions is the only administrator prompt extension and is appended to the system prompt.
• README, English README, and rules.example.toml now match the current configuration surface, with extra_instructions = "" by default.

Assets

• Windows x86_64: gitlab-work-runner-v2.0.14-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v2.0.14-linux-x86_64.tar.gz

v2.0.13

v2.0.13

Changes

• The built-in AI review system prompt is now a strict, low-false-positive reviewer policy covering untrusted MR input, context-tool usage, high-confidence finding criteria, added-line location rules, and final output requirements.
• Configured [ai_review].system_prompt no longer replaces the built-in reviewer policy; it is appended as extra system-level constraints so project configuration cannot remove core safety and output rules.
• README and rules.example.toml now document that the built-in system prompt always applies.

Assets

• Windows x86_64: gitlab-work-runner-v2.0.13-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v2.0.13-linux-x86_64.tar.gz

v2.0.12

v2.0.12

Changes

• Dashboard tool-call usage now reports the peak tool calls used by any single AI review batch against the configured per-batch max_tool_calls limit.
• This avoids presenting the per-batch max_tool_calls setting as a whole-review limit when multiple batches are used.

Assets

• Windows x86_64: gitlab-work-runner-v2.0.12-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v2.0.12-linux-x86_64.tar.gz

v2.0.11

v2.0.11

Changes

• Dashboard run details now fall back to failed task errors when the overall review request has no stored failure message, so task-level AI failures are still visible in the main run detail panel.
• Partial AI review failure comments now include the structured failure code and a sanitized error detail for each failed AI review.
• Dashboard coverage output hides per-file coverage rows when a failed AI review produced zero effective coverage, while keeping the coverage field visible.
• Dashboard coverage now shows simplified runtime usage for AI batches and context tool calls as used / limit.
• AI review coverage records now include max_batches, tool_calls_used, and max_tool_calls in SQLite and the Dashboard API.
• Dashboard sidebar icons now use fixed-size inline SVGs for clearer rendering.

Assets

• Windows x86_64: gitlab-work-runner-v2.0.11-windows-x86_64.zip
• Ubuntu x86_64: gitlab-work-runner-v2.0.11-linux-x86_64.tar.gz