alex/g3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

memory update

Browse Source
This commit is contained in:
Dhanji R. Prasanna
2026-02-13 11:35:26 +11:00
parent fcb839e5fd
commit 41584e4479
1 changed files with 10 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
analysis/memory.md
View File

@@ -1,5 +1,5 @@
# Workspace Memory # Workspace Memory
> Updated: 2026-02-11T05:07:33Z | Size: 30.0k chars > Updated: 2026-02-12T05:23:04Z | Size: 31.1k chars
### Remember Tool Wiring ### Remember Tool Wiring
- `crates/g3-core/src/tools/memory.rs` [0..5000] - `execute_remember()`, `get_memory_path()`, `merge_memory()` - `crates/g3-core/src/tools/memory.rs` [0..5000] - `execute_remember()`, `get_memory_path()`, `merge_memory()`
@@ -459,4 +459,12 @@ Makes tool output responsive to terminal width - no line wrapping, with 4-char r
- **Root cause**: `estimate_tokens()` only counted `message.content` chars, ignoring `tool_calls[].input` JSON. When sent to API, `tool_use` blocks include full input, causing massive undercount. - **Root cause**: `estimate_tokens()` only counted `message.content` chars, ignoring `tool_calls[].input` JSON. When sent to API, `tool_use` blocks include full input, causing massive undercount.
- **Impact**: In a real session, 303k chars of tool input (101k tokens) were invisible to the tracker. Context showed 39% but actual was >100%. Compaction never triggered → API 400 error. - **Impact**: In a real session, 303k chars of tool input (101k tokens) were invisible to the tracker. Context showed 39% but actual was >100%. Compaction never triggered → API 400 error.
- **Fix**: Added `estimate_message_tokens(message)` that sums content tokens + per-tool-call input tokens (chars/3 * 1.1 + 20 overhead). Updated `add_message_with_tokens()`, `recalculate_tokens()`, `clear_conversation()` to use it. - **Fix**: Added `estimate_message_tokens(message)` that sums content tokens + per-tool-call input tokens (chars/3 * 1.1 + 20 overhead). Updated `add_message_with_tokens()`, `recalculate_tokens()`, `clear_conversation()` to use it.
- **Tests**: 7 unit tests in `context_window.rs`, 1 integration test in `mock_provider_integration_test.rs::test_tool_call_input_tokens_tracked_in_context_window` - **Tests**: 7 unit tests in `context_window.rs`, 1 integration test in `mock_provider_integration_test.rs::test_tool_call_input_tokens_tracked_in_context_window`
### Anthropic ToolResult Structured Content (2026-02-12)
- `crates/g3-providers/src/anthropic.rs` [1058..1110] - `ToolResultContent` enum (Text | Blocks) with custom Serialize/Deserialize
- `crates/g3-providers/src/anthropic.rs` [1078..1086] - `ToolResultBlock` enum (Image, Text) for structured content inside tool_result
- `crates/g3-providers/src/anthropic.rs` [283..330] - `convert_messages()` User branch: images from read_image nested inside ToolResult content array, not as top-level Image blocks
- **Root cause**: Anthropic API rejects top-level `Image` blocks mixed with `ToolResult` blocks in the same user message. Images must be inside `tool_result.content` array.
- **Bug pattern**: `read_image` → images attached to tool result message → `convert_messages()` put images as top-level blocks → API 400 "tool_use ids without tool_result"
- Tests: `test_tool_result_with_images_nested_inside`, `test_tool_result_without_images_uses_string_content`, `test_regular_user_message_with_images_uses_top_level_blocks`, `test_strip_orphaned_tool_use_works_with_structured_tool_result`