Exists / NotExists: check whether the claim has any values. `expect_present = true` → Exists; `false` → NotExists.

MinLength / MaxLength: compare the `__length` fact against a threshold.

Contains / NotContains: check whether a specific value is among the claim's facts. `expect_present = true` → Contains; `false` → NotContains.

GreaterThan / LessThan: parse the first claim value as f64 and compare.

AnyOf / NoneOf: parse the expected value as a bracketed set and check membership. `expect_in_set = true` → AnyOf (pass if value in set); `false` → Non

Contains / NotContains: search selected values using `value_contains`.

Exists / NotExists: any non-null value counts as "present".

MinLength / MaxLength: find the first Sequence and compare its length.

GreaterThan / LessThan: find the first Number and compare.

AnyOf / NoneOf: check if selected values are in (or not in) a set.

Execute Bash code

Execute bash command with streaming output in a specific directory

Execute user input with template processing and auto-memory reminder. This is the common path for both single-line and multiline input.

Execute an agent task with retry logic This function handles: - Error classification (timeout, rate limit, server error, etc.) - Exponential backoff

Expand tilde (~) in a path to the user's home directory

Expand tilde to home directory

Extract file path from tool call content.

Extract inference parameters from a completion request.

Extract the summary section from the LLM response

Extracts token data from an OAuth 2.0 token response.

Extract a summary of tool calls from messages.

Extract a brief topic from text.

Extract topic hints from messages using heuristics.

Filter out empty or very short responses (likely just tool calls).

Find the end of a JSON object (matching braces).

Helper to find the last assistant message in history

Find the position of a marker in text that may contain ANSI codes. Searches by stripping ANSI codes character by character to find the true position.

Find a tool call pattern on its own line. If `find_last`, search backwards.

Create a copy of the config for planner mode

Format datalog execution results for display. This is used for shadow/dry-run output - printed to console but not injected into the context window.

Llama/CodeLlama format: [INST] <<SYS>>\nsystem<</SYS>>\n\nuser [/INST]

Format the loaded files status message (e.g., "✓ brief.md ✓ status.md")

Mistral Instruct format: <s>[INST] ... [/INST] response</s>

Qwen ChatML format: <|im_start|>role\ncontent<|im_end|>

Format a summary of pending research for display

Format a tool argument value for display (truncated for readability). Special handling for shell commands (show first line only if multiline).

Generate a session ID based on the initial prompt

Generate the prompt section describing available toolsets. This is injected into the system prompt so the agent knows what toolsets are available to

Get agent-specific language prompts for detected languages in the workspace. Returns both the formatted content and the list of languages that had mat

Generate a system prompt for agent mode by combining the agent's custom prompt with the full G3 system prompt (including plan tools, code search, webd

Generate a system prompt for agent mode with skills support. The agent_prompt replaces only the G3 identity line at the start of the prompt. Everythi

Get the path to the context summary file for a session. Returns .g3/sessions/<session_id>/context_summary.txt

Get all embedded skills.

Get the current HEAD SHA

Get image dimensions from raw bytes.

Get the prompt content for a specific language.

Get the provider reference for planner mode

Get the cache control config for the current provider (if Anthropic with cache enabled).

Get the path to a specific session's directory

Get the configured temperature for a provider from config.

Get the thinned content directory for a session. Returns .g3/sessions/<session_id>/thinned/

Get the path to the global verification key: `~/.g3/verification.key`

Check if a branch has commits ahead of main

Helper to check for consecutive user messages

Check if the pipeline has any failures

Check if this session has incomplete TODO items

Initialize the planning directory structure

Get the task result if successful

Check if element is displayed

Check if this item is terminal (done or blocked).

List all tracked background processes

Get all available embedded agent names.

List session IDs from .g3/sessions/, sorted newest-first, with optional limit.

List all available toolset names.

Load a session continuation by session ID (full or partial prefix match). This function searches for sessions matching the given ID: - First looks fo

Load a single skill file and add to the map.

Create a final completion chunk with stop reason.

Merge a branch to main

Get the next response from the queue (or default)

Perform context compaction by summarizing conversation history. This is the unified implementation used by both: - `force_compact()` - manual compact

Pre-flight check to validate and adjust max_tokens for the thinking.budget_tokens constraint. Returns the adjusted max_tokens that satisfies: max_toke

Get the thinking budget tokens for Anthropic provider, if configured. Pre-flight check to validate max_tokens for thinking.budget_tokens constraint.

Prepare the inference context: create context, tokenize prompt, decode initial batch.

Print a status line showing what project files were loaded. Format: " ✓ README ✓ AGENTS.md ✓ Memory"

Print provider diagnostics through the UiWriter for visibility

Print the workspace path in a consistent format.

Read AGENTS.md configuration from the workspace directory. Returns formatted content with emoji prefix, or None if not found.

Read a specific byte range from a file using seek (O(1) seek + O(n) read where n = range size). Handles UTF-8 boundary issues by extending the read sl

Read a rulespec from `analysis/rulespec.yaml` relative to the working directory.

Read workspace memory from analysis/memory.md in the workspace directory. Returns formatted content with emoji prefix and size info, or None if not fo

Helper to recalculate max_tokens after context reduction.

Get a human-readable name for a recoverable error type.