ACD (Aggressive Context Dehydration) fixes:
- Fixed dehydrate_context() to extract turn summary from context window
instead of using the passed-in final_response (which contained only
the timing footer, not the actual LLM response)
- Removed final_response parameter from dehydrate_context() since it
now self-extracts the last assistant message as the summary
- This ensures the actual turn summary is preserved after dehydration,
not just the timing footer
New /dump command:
- Added /dump command to dump entire context window to tmp/ for debugging
- Shows message index, role, kind, content length, and full content
- Available in both console and machine modes
UTF-8 safety:
- Fixed truncate_to_word_boundary() to use character indices instead of
byte indices, preventing panics on multi-byte UTF-8 characters
- Added UTF-8 string slicing guidance to AGENTS.md
Agent: g3
Adds a new --auto-memory CLI flag that automatically sends a reminder
to the LLM after each turn where tools were called, prompting it to
call the remember tool if it discovered any key code locations.
Changes:
- Add auto_memory field and set_auto_memory() method to Agent
- Add tool_calls_this_turn tracking in execute_tool_in_dir()
- Add send_auto_memory_reminder() that sends reminder after tool use
- Add --auto-memory CLI flag and wire it up in console/machine modes
- Call send_auto_memory_reminder() in single-shot and interactive modes
- Add visible status messages for auto-memory actions
Fixes bug where tool calls were not being tracked when execute_tool_in_dir
was called directly with working_dir=None.
- Add description field to SessionContinuation struct
- Extract first user message (truncated to ~60 chars at word boundary)
- Display as quoted text instead of session ID hash
- Fall back to session ID if no description available
Example: [2 hours ago] 'when I call /resume it only shows me 2 sessions...'
Fix session detection:
- Add save_session_continuation() calls at all session exit points
- Sessions now properly create .g3/session symlink for resume detection
- Fixes issue where g3 wasn't offering to resume previous sessions
Add /resume command:
- New list_sessions_for_directory() to scan available sessions
- New switch_to_session() method to safely switch between sessions
- Shows numbered list with timestamps, context %, and TODO status
- Saves current session before switching (can be resumed later)
- Restores full context if <80% used, otherwise uses summary
- Machine mode supports /resume and /resume <number>
Documentation:
- Add /clear and /resume to CONTROL_COMMANDS.md
- Update /help output with new commands
The timing footer (e.g., ⏱️ 19.4s | 💭 4.7s) was being saved to the
conversation history as a separate assistant message. This happened
because stream_completion_with_tools returns the timing footer in
TaskResult.response for display, but the caller was also saving it
to context.
Fix: Strip the timing footer (identified by \n\n⏱️) before saving
to context window. The timing footer remains display-only.
Also includes:
- Research tool blank line fix: only add visual separator for research
tool output, not all tools
- Research tool webdriver propagation: pass parent's webdriver browser
choice (Safari vs Chrome headless) to scout subprocess
New tool that spawns a scout agent to perform web research and return
a structured research brief. The scout agent uses webdriver to browse
the web and returns a decision-ready report.
Changes:
- Added 'research' tool definition (12 core tools total)
- Added research tool dispatch in tool_dispatch.rs
- Created tools/research.rs implementation:
- Spawns 'g3 --agent scout <query>' as subprocess
- Captures stdout and extracts last line (report file path)
- Reads and returns the report file contents
- Added exclude_research flag to ToolConfig
- Scout agent (agent_name == 'scout') does NOT have access to research
tool to prevent infinite recursion
- Updated system prompts to describe when to use research tool
- Added scout.md agent prompt with research brief output contract
The research tool is preferred for complex research tasks (APIs, SDKs,
libraries, approaches, bugs). WebDriver can still be used directly for
simple lookups or fine-grained control.
The buffer truncation code was slicing at a raw byte offset which could
land in the middle of a multi-byte character (like emojis), causing a
panic. Fixed by using char_indices() to find valid character boundaries.
Also added stop_reason field to CompletionChunk initializers in tests
to complete the stop_reason feature addition.
- Fix byte boundary panic in filter_json.rs line 327
- Add test for multi-byte character handling
- Update test files with missing stop_reason field
When the LLM executes a tool and then outputs text (e.g., analysis after
reading images), the text was being displayed during streaming but never
saved to the context window. This caused:
1. The response to appear truncated in the session log
2. Loss of context for subsequent turns
3. The LLM losing track of what it had already said
The fix saves current_response to the context window before breaking
out of the streaming loop for auto-continue after tool execution.
Reproduction scenario:
- User asks LLM to read images and analyze them
- LLM calls read_image tool
- Tool executes successfully
- LLM outputs analysis text ("Now I can see the results...")
- Text was displayed but lost from session log
Now the text is properly persisted to the context window.
- Remove final_output from tool definitions, dispatch, and misc tools
- Update system prompts to request summaries as regular markdown text
- Remove print_final_output from UiWriter trait and all implementations
- Remove final_output handling from agent core logic
- Rename final_output_summary → summary in session continuation
- Delete final_output test files
- Update tool count tests (12→11, 27→26)
This allows LLM summaries to stream through the markdown formatter
for a more natural, responsive user experience instead of buffering
everything into a tool call.
This fixes a bug where the agent would stop responding abruptly without
calling final_output. The root cause was the allow_multiple_tool_calls
config option (default: false) which caused the agent to break out of
the streaming loop mid-stream after executing the first tool, losing
any subsequent content.
Changes:
- Remove allow_multiple_tool_calls config option entirely
- Always process all tool calls without breaking mid-stream
- Simplify system prompt generation (no longer needs boolean param)
- Let the stream complete fully before continuing to next iteration
- Change find_last_tool_call_start to find_first_tool_call_start
- Remove parser.reset() call on duplicate detection
Benefits:
- Simpler logic with less conditional branching
- No lost content after tool calls
- Consistent behavior for all users
- Reduced config complexity
Auto-continue was incorrectly triggering when the LLM asked questions
in interactive/chat mode. Now auto-continue only activates when
is_autonomous is true, allowing proper back-and-forth conversation
in interactive mode.
Agent: fowler
- Extract check_duplicate_in_previous_message() helper to reduce nesting
from 6+ levels to 2 levels in stream_completion_with_tools
- Create do_thin_context() and do_thin_context_all() helpers to centralize
context thinning with event tracking
- Use provider_config::parse_provider_ref() in additional call sites
- All 295 tests pass
This continues the refactoring to eliminate code-path aliasing and
reduce cyclomatic complexity in the Agent implementation.
- Add TODO completion check to final_output tool in autonomous mode only
- When incomplete TODO items exist, reject final_output and prompt LLM to continue
- Non-autonomous modes (interactive, chat) are unaffected
- Add 6 tests verifying behavior in both autonomous and non-autonomous modes
Fixes issue where LLM would call final_output after completing first phase,
causing agent to stop prematurely instead of continuing with remaining phases.
- Move tool_executed = true after duplicate check to prevent auto-continue
from triggering when only duplicate tools were detected
- Reset parser state when duplicate detected to clear any partial/polluted
state from LLM stuttering or example tool calls in markdown blocks
- Print └─ before images to break out of tool output box
- Print ┌─ after images to resume tool output box
- Remove │ prefix from image preview and info lines
- Info line uses single space prefix, dimmed text
- Only include error messages in tool result (success info printed via imgcat)
- Remove │ prefix before image preview, use single space instead
- Keep info line on its own line with │ prefix
- Keep blank line spacing between images
- Remove unused assignment to final_output_called (returns immediately after)
- Mark cache_config field as #[allow(dead_code)] (reserved for future use)
- Mark print_status_line method as #[allow(dead_code)] (reserved for future use)
TODO lists are now stored in .g3/sessions/<session_id>/todo.g3.md instead
of the workspace root. This prevents different g3 sessions from accidentally
picking up or overwriting each other's TODOs.
Changes:
- Add get_session_todo_path() function in paths.rs
- Update todo_read/todo_write handlers to use session-specific paths
- Remove TODO loading at Agent initialization (sessions start fresh)
- Update prompts to reflect session-scoped behavior
Fallback behavior preserved for planner mode (G3_TODO_PATH env var).
Adds a new tool that allows launching processes (like game servers) in the
background while g3 continues to operate. The process runs independently
with stdout/stderr captured to a log file.
Features:
- Named process tracking for easy reference
- Automatic log capture to logs/background_processes/
- Returns PID and log file path for use with shell tool
- Automatic cleanup on agent shutdown via Drop trait
Usage: Use shell tool to interact with the process:
- Read logs: tail -100 <logfile>
- Check status: ps -p <pid>
- Stop process: kill <pid>
Files:
- New: crates/g3-core/src/background_process.rs
- New: crates/g3-core/tests/background_process_demo_test.rs
- Modified: crates/g3-core/src/lib.rs (tool definition + handler)
- Modified: crates/g3-core/src/prompts.rs (documentation)
When the LLM outputs text containing tool call patterns (e.g., reading
log files, showing examples, or discussing tool calls), the parser's
has_unexecuted_tool_call() would detect these as real tool calls and
trigger auto-continue, leading to repeated empty responses.
The fix: mark the parser buffer as consumed when content is displayed.
This prevents tool-call-like patterns in displayed text from triggering
false positives later. The fix is safe because:
1. Only runs when no tool was detected (inside 'if !tool_executed')
2. Legitimate tool calls are detected first by process_chunk()
3. Matches existing pattern of calling mark_tool_calls_consumed()
after tool execution
The auto-continue logic was adding User continue prompts without first
adding an Assistant message when the LLM returned an empty response.
This caused consecutive User messages in the conversation history,
which confused the LLM and caused it to return more empty responses.
The fix ensures an Assistant message is always added before the continue
prompt, using '[empty response]' as a placeholder when the LLM returned
nothing substantive. This maintains proper User/Assistant alternation.
Reduce lib.rs from 7481 to 6557 lines (-12.4%) by extracting:
- paths.rs: Session/workspace path utilities (get_todo_path, get_logs_dir, etc.)
- streaming_parser.rs: StreamingToolParser for LLM response parsing
- utils.rs: Diff parsing and shell escaping utilities
- webdriver_session.rs: Unified Safari/Chrome WebDriver abstraction
All public APIs preserved via re-exports for backward compatibility.
Added 13 new unit tests across extracted modules.
All 225 tests pass.
Added a quality-of-life feature that displays:
- Tokens used in the current turn (from LLM response, not estimated)
- Current context window usage percentage
These are displayed dimmed after the timing info:
⏱️ 1.2s | 💭 0.3s 1234tk | 45% ctx
The token count comes directly from the LLM's usage response data,
not from any estimation. If no usage data is available from the LLM,
only the context percentage is shown.
Added 13 tests to verify that duplicate detection only catches
IMMEDIATELY SEQUENTIAL duplicates:
- test_find_complete_json_object_end_* - Tests for JSON parsing helper
- test_same_tool_with_text_between_not_duplicate - Key test ensuring
tool calls separated by text are NOT duplicates
- test_different_tools_back_to_back_not_duplicate
- test_same_tool_different_args_not_duplicate
- test_identical_tool_calls_back_to_back_are_duplicates
- test_has_text_after_tool_call - Tests text detection logic
- test_tool_call_with_newlines_between
- test_tool_call_with_whitespace_text_between
- test_tool_call_in_middle_of_text
- test_multiple_different_tool_calls_with_text
Also made find_complete_json_object_end public for testing.
1. Set tool_executed=true when a tool call is detected, even if skipped
as a duplicate. This prevents the raw JSON from being printed to screen
when a tool call is detected but not executed.
2. Remove session-level duplicate detection entirely. All tools should be
allowed to be called multiple times in a session.
3. Fix sequential duplicate detection to only catch IMMEDIATELY sequential
duplicates:
- DUP IN CHUNK: Now only checks if the PREVIOUS tool call in the chunk
is the same (not any tool call in the chunk)
- DUP IN MSG: Now only checks if the LAST tool call in the previous
message matches AND there's no text after it. If there's any
non-whitespace text between tool calls, they're not considered
duplicates.
This allows legitimate re-use of tools while still catching cases where
the LLM stutters and outputs the same tool call twice in a row.
The bug was in the chunk.finished block inside stream_completion_with_tools.
When no tool was executed in the CURRENT iteration (!tool_executed), the code
would return early without checking if tools were executed in PREVIOUS iterations
(any_tool_executed) and final_output was never called.
This caused the agent to terminate prematurely after executing tools like
todo_read when the LLM responded with text instead of calling final_output.
The fix adds a check: if any_tool_executed && !final_output_called, we break
to let the outer loop's auto-continue logic prompt the LLM to continue.
Also fixed missing debug! import in g3-console/src/main.rs.
Converted ~77 info! macro calls to debug! across the codebase to prevent
log messages from interrupting the CLI experience during normal operation.
Users can still see these logs by setting RUST_LOG=debug if needed.
Affected crates:
- g3-cli
- g3-computer-control
- g3-console
- g3-core
- g3-ensembles
- g3-execution
- g3-providers
