Module tooling 

Tool-call construction + tool-result extraction. 1:1 port of upstream’s agent-sdk/src/bridge/tooling.ts (714 LoC).

Three groups of helpers:

1. Front-of-tool: create_tool_call, tool_title, normalize_tool_kind.
2. Result extraction: build_tool_result_fields per-tool branches (Bash / Read / Write / Edit / Agent / ReadMcpResource), normalize_tool_result_text, extract_text, persisted_output_first_line, sanitize_sdk_rejection_text, extract_tool_output_metadata.
3. Helpers: unwrap_tool_use_result.

Re-exports

use serde_json::Map; 🔒

use serde_json::Value; 🔒

use serde_json::json; 🔒

use forge_primitives::BashOutputMetadata; 🔒

use forge_primitives::ChunkContent; 🔒

use forge_primitives::ToolCall; 🔒

use forge_primitives::ToolCallContent; 🔒

use forge_primitives::ToolCallLocation; 🔒

use forge_primitives::ToolCallUpdateFields; 🔒

use forge_primitives::ToolOutputMetadata; 🔒

Structs

UnwrappedToolResult

Mirrors unwrapToolUseResult(rawResult). Walks the tool_result envelope: if the value is an object, surfaces the inner content / result / text field and the is_error / error flag.

Constants

CACHE_PREVIEW_LIMIT_BYTES 🔒

PERMISSION_DENIED_TOOL_USE_EXACT 🔒

PERMISSION_DENIED_TOOL_USE_PREFIX 🔒

PERSISTED_OUTPUT_CLOSE_TAG 🔒

PERSISTED_OUTPUT_OPEN_TAG 🔒

TOOL_RESULT_TYPES 🔒

Block types the CLI uses for tool results. Mirrors upstream’s TOOL_RESULT_TYPES set in tooling.ts:5.

USER_REJECTED_TOOL_USE_EXACT 🔒

USER_REJECTED_TOOL_USE_PREFIX 🔒

Functions

agent_title_from_agent_output 🔒

bash_background_message 🔒

build_bash_display_output 🔒

build_tool_result_fields

Mirrors buildToolResultFields(isError, rawContent, base?, rawResult?). The high-level entry that turns an arbitrary tool_result block into the structured ToolCallUpdateFields upstream’s TUI consumes.

create_tool_call

Mirrors createToolCall(toolUseId, name, input, parentToolUseId). Builds the rich ToolCall envelope upstream emits when the assistant first invokes a tool.

edit_diff_content 🔒

Mirrors editDiffContent(name, input) - initial diff content for Edit / Write tool cards (before the result lands). Empty Vec for other tools.

edit_diff_from_input 🔒

edit_diff_from_result 🔒

extract_text 🔒

Mirrors extractText(value) - flattens a tool_result content payload into a single String. Accepts string, array of { type: "text", text } blocks, or a single object carrying prose under one of plan / description / summary / message / text (priority order: plan-mode-specific first, generic last).

extract_tool_output_metadata 🔒

file_unchanged_result_text 🔒

find_bash_result_record 🔒

is_tool_result_block_type

mcp_resource_content_from_result 🔒

normalize_tool_kind 🔒

Mirrors normalizeToolKind. Maps tool name to the typed ToolKind consumed by a view’s tool-card renderer.

normalize_tool_result_text 🔒

parse_json_candidate 🔒

persisted_output_first_line 🔒

persisted_output_inner_text 🔒

preview_kilobyte_label 🔒

push_nested_records 🔒

push_record 🔒

push_records 🔒

push_structured_record_candidates 🔒

resolve_tool_name 🔒

result_indicates_background 🔒

True when a tool_result carries a background sentinel - the same fields Bash keys on (backgroundTaskId / assistantAutoBackgrounded / backgroundedByUser).

result_record_candidates 🔒

sanitize_sdk_rejection_text 🔒

tool_search_matches_from_content 🔒

Walk a tool_result.content array looking for the CLIENT ToolSearch result wire shape - [{type: "tool_reference", tool_name: "..."}, ...] - and return the matched tool names in source order. Anything not a tool_reference entry is skipped, so a mixed-content result still surfaces the references it carried.

tool_title 🔒

Mirrors toolTitle(name, input). Produces the human-friendly title shown on the tool card header (e.g. Bash → command, Glob → pattern + path, etc.).

unwrap_tool_use_result

write_diff_from_input 🔒

write_diff_from_result 🔒