Save pipeline state to disk

Write a "GIT COMMIT" entry

Helper to create a chunk

Helper to create a test agent with mock provider

Deduplicate tool calls, detecting sequential duplicates within a chunk and duplicates against the previous message. Returns each tool call paired wit

Generate the XML section for available skills. Returns an empty string if no skills are available. The XML format follows the Agent Skills specificat

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

Get the timeout duration for a tool (extracted for testing)

Prepare user input for plan mode, prepending "Create a plan: " if this is the first message. Returns the (possibly modified) input and whether the fla

Print a message and flush stdout (for prompts)

Append an entry to planner_history.txt. This function opens the file in append mode, writes a single line, and explicitly flushes the buffer to ensur

Combine AGENTS.md and memory content into a single string for project context. Returns None if all inputs are None, otherwise joins non-None parts wi

Test helper: Convert messages using the same logic as GeminiProvider This mirrors the convert_messages function behavior

Helper to create an agent with a mock provider

Extract shell commands from the LLM response. Looks for {{CODE EXPLORATION COMMANDS}} section and extracts commands from code blocks.

Format a timestamp for planner_history.txt entries Format: YYYY-MM-DD HH:MM:SS (ISO 8601 for readability)

Format the timing footer with optional token usage info

Reset the context window with a summary. Preserves the original system prompt as the first message.

Check if file changes occurred without an approved plan. This function should be called after each tool execution when in plan mode. It checks `git s

Make a git commit with the given summary and description

Create a new worktree for a session

Helper to create an agent with a mock provider

Ensure the session directory exists for a specific session ID. Creates .g3/sessions/<session_id>/ and subdirectories.

Evaluate a rulespec against an action envelope.

Helper to create a fragment and extract its topics via stub generation

Find the end byte-index of a complete JSON object starting at the beginning of `text`. Tracks brace nesting and string escaping to find the matching

Get detailed context statistics

Get a toolset by name. Returns `Ok(Toolset)` if found, or `Err` with a helpful message listing available toolsets if not found.

Handle a tool parsing hint - this is the core logic extracted for use in callbacks

Check if the plan has been approved.

Load and validate a project from a path string. This is the shared logic used by both `--project` CLI flag and `/project` command. It handles: - Tild

Parse a SKILL.md file from its content and path. The file must have YAML frontmatter delimited by `---` lines, with at least `name` and `description`

Parse an evidence string into an EvidenceType.

Read a line of user input

Save this fragment to disk.

Shorten a path string for display by: 1. Replacing project directory prefix with `<project_name>/` (if project is active) 2. Replacing workspace direc

Simulates the should_auto_continue logic from lib.rs After removing final_output, the logic is simpler: - Continue if there's an incomplete tool call

Check if we should trigger compaction (at 80% capacity or 150k tokens).

Start a new background process # Arguments `name` - A unique name for this process (used to reference it later) `command` - The shell command to exec

Feed content in chunks of specified size

Update token usage from provider response. NOTE: This only updates cumulative_tokens (total API usage tracking). Calibrates `used_tokens` from the pr

Add a message directly to the context window. Used for injecting discovery messages before the first LLM turn.

Get the git branch name for this session

Calculate the maximum bytes we should read based on context window state.

Clean LLM-specific tokens from content

Helper to create an agent with a mock provider

Extract a balanced JSON object/array from a string

Extract the research report from scout output. Looks for content between SCOUT_REPORT_START and SCOUT_REPORT_END markers. Preserves ANSI escape codes

Filter --accept from g3_args and return (filtered_args, was_accept_present) This handles the case where --accept is passed after positional args, whic

Find all code fence ranges in text. Returns (start, end) byte positions. Each range represents content INSIDE a fence (between ``` markers).

System prompt for providers with native tool calling (Anthropic, OpenAI, etc.) Uses include_str! to embed the prompt at compile time.

System prompt for providers without native tool calling (embedded models)

Get the thinking budget tokens for Anthropic provider, if configured.

Check if a partial JSON tool call has been invalidated. Invalidation cases: 1. Unescaped newline inside a JSON string (invalid JSON) 2. Newline follo

Load the latest session continuation artifact if it exists

Create a final completion chunk with tool calls and usage.

Mark the current stage as complete and advance

Mark the current stage as failed

Mark the current stage as running

Read include prompt content from a specified file path. Returns formatted content with emoji prefix, or None if path is None or file doesn't exist.

Helper to create a test git repository

Checks if a provider reference should be registered. A provider should be registered if: - Its full reference (e.g., "openai.default") is in the list

Helper function to truncate strings for logging

Truncate output if it exceeds the threshold, saving full content to a file. If the output is larger than OUTPUT_TRUNCATE_THRESHOLD: 1. Saves the full

Try to read an embedded skill by path. Recognizes paths like: - `<embedded:research>/SKILL.md` - `<embedded:skill-name>/SKILL.md` - `&lt;embedded:res

Set custom max retries

Add a response to the queue

Add multiple responses to the queue

Clear all stored tool state after output is complete.

Compress a file path to fit within the given width. Preserves the filename and as much of the path as possible. Truncates parent directories from the

Helper to create a test agent in agent mode

Platform-specific constructor

Detect languages present in the workspace by scanning for file extensions. Returns a list of detected language names.

Change to the workspace directory

Extract YAML content from a markdown file with ```yaml code block.

Find an incomplete agent session for the given agent name. Returns the most recent session that: 1. Was running in agent mode with the matching agent

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

Format an action envelope as human-readable markdown. This produces a rich, readable format suitable for tool output, showing the facts recorded abou

Generate a session ID based on description and optional agent name. For agent mode, uses agent name as prefix. For regular mode, uses first 5 words o

Get the fragments directory for a session (for ACD dehydrated context). Returns .g3/sessions/<session_id>/fragments/

Get the base .g3 directory path. This is the root for all g3 session data in the current workspace.

System prompt for providers with native tool calling, with skills support.

Log the error context with ERROR level