When partial JSON tool call patterns appear in LLM output (e.g., from
quoting file content), the parser would incorrectly report them as
"incomplete tool calls", triggering auto-continue loops.
Fix: Added is_json_invalidated() to detect when partial JSON has been
invalidated by subsequent content that cannot be valid JSON:
- Unescaped newline inside a string (invalid JSON)
- Newline followed by prose text outside a string
The check is only applied to incomplete JSON - complete tool calls
with trailing text are still correctly detected.
Added 6 new tests covering:
- Tool results with partial JSON patterns
- LLM quoting file content inline vs on own line
- Comment prefixes (// # -- etc) with partial patterns
- Real incomplete tool calls (should still be detected)
The streaming parser was incorrectly detecting tool call patterns that
appeared inline in prose (e.g., when explaining the format), causing
g3 to return control mid-task.
Fix: Modified find_first_tool_call_start() and find_last_tool_call_start()
to only recognize patterns that appear on their own line (at start of
buffer or after newline with only whitespace before the pattern).
Changes:
- Added is_on_own_line() helper to check line-boundary conditions
- Updated detection methods to skip inline patterns
- Removed sanitize_inline_tool_patterns() and LBRACE_HOMOGLYPH (no longer needed)
- Rewrote tests for new behavior
- Added streaming_repro tests that use process_chunk() to verify the exact bug scenario
28 tests covering: streaming repro, line boundaries, Unicode, code contexts, edge cases
Rust-specific readability guidance for the carmack agent including:
- let...else example for shallow control flow
- Async: don't block the runtime (tokio::fs, spawn_blocking, Send)
- Visibility: prefer pub(crate), private fields with accessors
- Generics: impl Trait over explicit params, avoid complex where clauses
- Improved iterator guidance: if you need a comment, use a loop
- UTF-8 string slicing warnings
- Ownership/lifetime pragmatism
- Anti-patterns: no macros/typestate/proc-macros unless already in repo
Also adds Rust detection to LANGUAGE_PROMPTS (empty base prompt,
agent-specific prompts handle the guidance).
Automatically accept the session after g3 completes successfully,
but only if there are commits on the branch.
Changes:
- Add --accept flag to Run command (stripped, not passed to g3)
- Add has_commits_on_branch() helper using git rev-list --count
- Auto-accept triggers merge to main and cleanup when:
1. g3 exits successfully (exit code 0)
2. Branch has commits ahead of main
- Show warning if --accept set but no commits exist
Usage: studio run --agent carmack --accept
When running in agent mode (e.g., --agent carmack) in a workspace with
detected languages, inject agent+language-specific prompts from
prompts/langs/<agent>.<lang>.md at the end of the system prompt.
Changes:
- Add AGENT_LANGUAGE_PROMPTS static array for compile-time embedding
- Add get_agent_language_prompt() to look up specific agent+lang combos
- Add get_agent_language_prompts_for_workspace_with_langs() that returns
both content and matched languages for display
- Update agent_mode.rs to inject prompts and show which languages loaded
- Display format: '✓ carmack: racket language guidance'
- Add tests for new functionality
Uses the same detect_languages() mechanism as regular language prompts
to avoid code-path aliasing.
Use try_init() instead of init() for tracing subscriber setup to
gracefully handle cases where a global subscriber is already set.
This fixes a panic in the scout agent subprocess when spawned by the
research tool, where a dependency may have already initialized tracing.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Move initialize_logging() call to run immediately after CLI parsing,
before any mode checks. This ensures the --verbose flag works correctly
in planning mode, which previously bypassed logging initialization.
Previously, planning mode would return early before initialize_logging()
was called, causing verbose output to be silently ignored.
- Add language_prompts module that auto-detects programming languages in workspace
- Scan for language files with depth limit (2) to inject relevant toolchain prompts
- Add prompts/langs/ directory for language-specific markdown files
- Include Racket/raco toolchain guidance as first language prompt
- Update combine_project_content() to accept language_content parameter
- Integrate language detection into main CLI flow and agent mode
- Update project memory with new feature documentation
LLMs were prefixing shell commands with `cd <workspace> &&` unnecessarily,
wasting tokens and cluttering CLI display. Added clear guidance in the
shell tool description that commands already execute in the working directory.
• Agent prompts are now embedded within the g3 binary
• README.md - Added new "Agent Mode" section documenting:
• All 7 built-in agents with their focus areas
• Usage examples (--list-agents, --agent <name>)
• How to create custom workspace agents
Behavior
1. Workspace agents take priority - If agents/<name>.md exists in the workspace, it's used
2. Embedded fallback - If no workspace agent exists, the embedded version is used
3. Portability - g3 binary now works on any repo without needing the agents/ directory
4. Discoverability - g3 --list-agents shows all available agents and their source
When users explicitly pass --new-session, they want a fresh session.
Previously g3 would still prompt to resume an existing session.
Now the resume check is skipped entirely when the flag is set.
- Rename take_screenshot -> screenshot, code_coverage -> coverage (shorter names)
- Align | character across all compact tools (pad to 11 chars for str_replace)
- Make code_search a compact tool with summary display
- Show language and search name in code_search output (e.g., rust:"find structs")
- Add format_code_search_summary() to extract match/file counts from JSON response
Remove dead code - constructor variants that had no callers:
- new_with_readme()
- new_autonomous_with_readme()
- new_with_quiet()
These were thin wrappers around new_with_mode_and_readme() that were
never used externally. All 5 remaining constructors have verified callers.
Results:
- lib.rs reduced from 2817 to 2797 lines (-20 lines)
- Eliminated code-path aliasing: 8 constructors → 5 constructors
- All g3-core tests pass
- Full workspace compiles cleanly
Agent: fowler
- Updated memory reminder prompt with per-symbol char ranges
- Added two few-shot examples: Session Continuation (feature) + UTF-8 Safe Slicing (pattern)
- Updated system prompt Memory Format section to match
- Format: file -> nested symbols with [start..end] ranges and descriptions
- Enables direct read_file navigation to specific functions
- webdriver flag now defaults to true (tools always available)
- chrome_headless flag now defaults to true (Chrome is default browser)
- Use --safari flag to override and use Safari instead
- Updated README documentation to reflect new defaults
Extract the get_stats() function (158 lines) from lib.rs to a new stats.rs module.
Changes:
- Create stats.rs with AgentStatsSnapshot struct for capturing agent state
- Replace inline formatting logic with delegation to snapshot.format()
- Add unit tests for stats formatting (empty and populated states)
- Reduce lib.rs from 2961 to 2818 lines (-143 lines)
The new module improves:
- Testability: Stats formatting can now be unit tested in isolation
- Separation of concerns: Formatting logic is decoupled from Agent struct
- Readability: lib.rs is more focused on core agent behavior
All 271 workspace tests pass.
Agent: fowler
- Remove unused total_lines variable in file_ops.rs
- Fix UTF-8 boundary panic in utils.rs when generating diff error preview
The code was slicing at byte index 200 which could land inside a
multi-byte character (e.g., box-drawing chars like ─). Now uses
character-based slicing with chars().take() instead.
Studio can now run g3 without specifying an agent:
# Agent mode (existing)
studio run --agent carmack "fix the bug"
# One-shot mode (new)
studio run "fix the bug"
When no agent is specified, sessions are created under the 'single'
directory in .worktrees/sessions/single/<session-id>/
This makes Studio a complete replacement for Flock mode.
Standardize project name to lowercase 'g3' throughout documentation,
comments, and configuration files. Environment variables (G3_*) are
unchanged as they follow the uppercase convention.
Improve readability of stream_completion_with_tools (~1000 line function):
- Add deduplicate_tool_calls() helper with closure for previous-message check
- Add should_auto_continue() with AutoContinueReason enum for clearer control flow
- Replace inline deduplication loop with helper call (-19 lines)
- Replace complex auto-continue conditional with match on reason enum (-13 lines)
- Add section comments for major phases (State Init, Pre-loop, Main Loop, Auto-Continue, Post-Loop)
- Add comprehensive tests for new helpers
Net reduction: 82 deletions, behavior unchanged (172+ tests pass)
Agent: carmack
When the streaming parser encountered fragments of JSON that looked like
partial tool calls (e.g., {"tool":) embedded in inline text (like code
examples or prose), it would incorrectly enter JSON parsing mode and
poison the parser state, causing control to be returned to the user
mid-task.
This fix:
- Adds sanitize_inline_tool_patterns() to detect tool-call patterns that
are NOT on their own line and replace the opening brace with a Unicode
homoglyph (fullwidth left curly bracket U+FF5B)
- Integrates sanitization into process_chunk() before text is buffered
- Updates system prompts to instruct LLMs to use homoglyphs when showing
example tool call JSON in prose
- Adds comprehensive tests for the sanitization logic
Real tool calls from LLMs always appear on their own line, so those are
left untouched. Only inline patterns (with non-whitespace before them)
are sanitized.
When the LLM reads the same file multiple times in sequence (scrolling
through a large file), instead of showing each as a separate line:
● read_file | path [0..2000] | 50 lines | 100 ◉ 5ms
● read_file | path [2000..4000] | 50 lines | 100 ◉ 5ms
● read_file | path [4000..6000] | 50 lines | 100 ◉ 5ms
Now shows a cleaner continuation format:
● read_file | path [0..2000] | 50 lines | 100 ◉ 5ms
└─ reading further [2000..4000] | 50 lines | 100 ◉ 5ms
└─ reading further [4000..6000] | 50 lines | 100 ◉ 5ms
This makes it visually clear that the agent is scrolling through
a single file rather than reading multiple different files.
Implementation:
- Added last_read_file_path field to ConsoleUiWriter
- Detect when consecutive read_file calls target the same file
- Print continuation format for subsequent reads
- Reset tracking when:
- A different tool is executed (shell, write_file, etc.)
- A different file is read
- Text is output between tool calls
Agent: hopper
Adds 56 new integration tests covering the observable end-of-turn
behaviors in the streaming module:
- Timing footer formatting (5 tests): verifies user-facing timing display
with various durations, token counts, and context percentages
- Tool call duplicate detection (6 tests): ensures identical sequential
tool calls are detected while different tools/args are not
- Empty response detection (9 tests): validates detection of empty,
whitespace-only, and timing-only responses that trigger auto-continue
- Connection error classification (5 tests): verifies EOF, connection,
chunk, and body errors are correctly identified for graceful recovery
- Tool output summary formatting (17 tests): covers read_file, write_file,
str_replace, remember, screenshot, coverage, and rehydrate summaries
- Duration formatting (4 tests): milliseconds, seconds, minutes, zero
- Text truncation (4 tests): short/long strings, multiline, flag behavior
- LLM token cleaning (3 tests): removal of stop tokens like <|im_end|>
- Edge cases (4 tests): empty inputs, unicode handling, large numbers
All tests are blackbox/characterization style - they test observable
outputs through stable public interfaces without encoding internal
implementation details. Tests remain stable under refactoring that
preserves behavior.
The truncate_for_display() function now takes only the first line
of input before truncating. This prevents multi-line error messages
(like str_replace failures) from breaking the compact single-line
format.
Added tests for multi-line input handling.
1. architecture.md: Fixed diagram to show 'studio' instead of 'g3-console'
(the crate was renamed during development)
2. analysis/memory.md: Removed reference to non-existent machine_ui_writer.rs
3. theme.rs: Clarified that 'retro' is a theme option (the default theme),
not a separate TUI mode. No --retro CLI flag exists.
The write_file compact display was showing 1 line because it was
counting lines in the success message, not the actual written content.
Now parses the tool result (e.g. '✅ wrote 150 lines | 4.2k chars')
to extract and display the correct counts.
Added format_write_file_result() to parse the tool output.
