Create a text content chunk (not finished).

Get the path to this session's metadata file

Read a plan from the session's plan.g3.md file.

Remove a research task (e.g., after it's been fully processed)

Execute a simple async operation with retry (for non-agent tasks) This is a simpler retry wrapper for operations like LLM API calls that don't involv

Save a session continuation artifact Save session continuation for potential resumption

Enable or disable plan mode (blocks file changes without approved plan)

Shorten any paths found within a shell command string. This replaces project paths with `<project_name>/`, workspace paths with `./`, and home paths w

Take all completed research tasks that haven't been injected yet Marks them as injected so they won't be returned again

Truncate a string value for display, respecting UTF-8 boundaries. Takes only the first line to avoid multi-line output in compact display.

Approve the current revision.

Check if two tool calls are duplicates (same tool and args)

Build the summary request messages from conversation history.

Calculate max_tokens for a summary request, ensuring it satisfies the thinking constraint. Returns (max_tokens, whether_fallback_is_needed) IMPORTANT

Compute a canonical YAML representation of the envelope facts. This produces a deterministic string by sorting keys, ensuring the same facts always p

Create a test agent in a temporary directory

Internal helper: thin context and track the event

Create the workspace directory if it doesn't exist

Execute the rehydrate tool. Loads a fragment from disk and returns its contents for the LLM to review.

Expand tilde in path to home directory.

Extract the last block from the response (for coach feedback in autonomous mode) This looks for the final_output content which is the last substantial

Find the start of the current "word" being typed, respecting quotes. Returns (word_start, word) where word_start is the byte index.

Format a duration for display

Format inline content with bold, italic, code, strikethrough, and links.

Format a str_replace result summary.

Format a write_file result for compact display. Parses the tool result which is in format: "wrote N lines | M chars" Returns a compact summary like "w

Get the set of dirty file paths from `git status --porcelain`. Returns an empty set if not a git repo or if the command fails. Each entry is the file

Get the path to the g3 binary (same directory as studio)

Get all language prompts for detected languages in the workspace. Returns formatted content ready for injection into the system prompt.

Get the configured max_tokens for a provider from config.

Get the path to the plan.g3.md file for a session.

Get the path to the todo.g3.md file for a specific session. Returns .g3/sessions/<session_id>/todo.g3.md

Get the path to the .g3/sessions directory (where all sessions are stored)

Get the status of a specific research task

Get the tool definitions for this toolset.

Check if the pipeline is complete

List all fragments for a session, ordered by creation time.

Load skills from a directory into the map. Each subdirectory should contain a SKILL.md file.

Get the path to new_requirements.md

Normalize Unicode space characters in a file path to regular ASCII spaces. macOS uses special Unicode space characters in certain filenames: - U+202F

Read an action envelope from the session's envelope.yaml file.

Recalculate token usage based on current conversation history.

Reset context window with a summary and optional ACD stub. Preserves the original system prompt as the first message. If stub is provided, it's added

Resolve the max_tokens to use for a given provider, applying fallbacks.

Create a MockResponse with specific cache statistics

Send an auto-memory reminder to the LLM if tools were called during the turn. This prompts the LLM to call the `remember` tool if it discovered any ke

Feed content character by character (worst case for streaming)

Convert the envelope to a YamlValue for selector evaluation.

Get the path to todo.g3.md

Truncate a line for tool output display

Truncate a string to approximately max_len characters, ending at a word boundary. This function attempts to break at a space character for cleaner di

Get a reference to the UI writer.

Validate that the check has required fields.

Verify a code location evidence (file exists, line numbers valid). # Arguments `file_path` - Path to the file (relative to working_dir or absolute) `

Update cumulative token usage (for streaming) when no provider usage data is available. NOTE: This only updates cumulative_tokens, not used_tokens.

Clean up all processes on shutdown

Clear project content from the system message. This is equivalent to calling `set_project_content(None)`.

Convert g3 messages to Gemini format Key differences: - Gemini uses "model" instead of "assistant" - System messages go in system_instruction, not co

Create the core tools that are always available

Helper function to create a minimal config for testing

Create tool definitions for native tool calling providers. Returns a vector of Tool definitions that describe the available tools and their input sch

Creates an Assistant message with a tool call in g3's JSON format.

Get the path to current_requirements.md

Delete pipeline state from disk

Display the pipeline with current stage highlighted

Ensure the planner_history.txt file exists, creating it if necessary

Ensure the new_requirements.md file has the ORIGINAL_REQUIREMENTS tag

Estimate token count from byte size (must match file_ops.rs)

Evaluate a single predicate against the fact lookup table.

Execute a task with retry logic for recoverable errors.

Execute a tool with an optional working directory (for discovery commands)

Expand tilde in path to home directory

Main entry point for exploring a codebase at the given path. Detects which languages are present and generates a comprehensive report.

Legacy method - extract the final_output content from the response Now just delegates to extract_last_block() for backwards compatibility

Extract text content from Gemini response parts

Extract tool name from assistant message content.

Mark a research task as failed

Filter JSON tool calls from streaming content for display. This is a UI concern - the raw content should be preserved for logging. Default implementat

Build the final response and prepare for return. This is the single canonical path for completing a streaming turn: 1. Finish streaming markdown 2. S

Convert accumulated tool calls map to final Vec<ToolCall>.

Check if text contains any stop sequence. Returns the truncation position if found.

Create a copy of the config for player mode in autonomous execution

Format a code_coverage result summary.

GLM-4 ChatGLM4 format: [gMASK]<sop><|role|>\ncontent

Format user input with markdown and special highlighting (ALL CAPS, quotes).

Format a read_file result summary (e.g., "✅ read 42 lines | 1.2k chars").

Format a rehydrate result summary.