abc/analysis_claude_code
1
0
Fork 0
mirror of https://github.com/shareAI-lab/analysis_claude_code.git synced 2026-09-20 12:13:38 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: preserve tool results during context compaction

Browse Source
This commit is contained in:
Haoran
2026-08-23 11:50:52 +08:00
parent 61c3e64595
commit 6d70d2e071
24 changed files with 843 additions and 497 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
s08_context_compact/README.md
View File

@@ -88,11 +88,11 @@ This step examines only the latest batch of tool results. The complete output re
## Step 2: snip_compact
Once the history exceeds 50 messages, `snip_compact` writes the complete history to `.transcripts/`, then keeps the first 3 and latest 47 messages. The marker records how many messages were removed and where to find the complete transcript.
Once the history exceeds 50 messages, `snip_compact` writes the complete history to `.transcripts/`, then keeps the first 3 and latest 46 messages. The archive marker occupies the remaining slot, records how many messages were removed, and points to the complete transcript.
```python
head_end = 3
tail_start = len(messages) - (max_messages - head_end)
tail_start = len(messages) - (max_messages - head_end - 1)
if self.has_tool_use(messages[head_end - 1]):
while (head_end < tail_start
@@ -117,37 +117,34 @@ This step controls the number of messages. Tool results inside the retained mess
## Step 3: micro_compact
After the first two steps, `prepare` estimates the remaining context size and runs `micro_compact` only when it is above `CONTEXT_CHAR_LIMIT`. `micro_compact` preserves every `tool_result` added after the most recent assistant response, so the model sees each new result in full once. Among results the model has already consumed, it keeps the latest 3 and shortens older results longer than 120 characters. Persisted results keep their file path; the rest become placeholders:
After the first two steps, `prepare` estimates the remaining context size and runs `micro_compact` only when it is above `CONTEXT_CHAR_LIMIT`. Among results the model has already consumed, `micro_compact` keeps the latest 3 and shortens older results longer than 120 characters until the context approaches 80% of the limit. Before replacing an old result, it writes the complete content to disk, so every replacement retains a recovery path:
![Replacing old results](images/micro-compact.en.svg)
![Replacing old results with recovery paths](images/micro-compact.en.svg)
```python
unseen = self.unseen_tool_result_positions(messages)
consumed = [entry for entry in results if entry[:2] not in unseen]
for _, _, block in consumed[:-self.KEEP_RECENT_RESULTS]:
if self.estimate_chars(messages) <= target_chars:
break
content = str(block.get("content", ""))
if len(content) <= 120:
continue
saved_path = next(
(line.removeprefix("Full output: ") for line in content.splitlines()
if line.startswith("Full output: ")),
None,
)
block["content"] = (
f"[Earlier tool result saved at {saved_path}]"
if saved_path else "[Earlier tool result omitted.]"
)
saved_path = self.persisted_output_path(content)
if not saved_path:
saved_path = self.save_output(block["tool_use_id"], content)
block["content"] = f"[Earlier tool result saved at {saved_path}]"
```
An old result that was not persisted keeps only a placeholder. Results saved in Step 1 retain the path to their complete output.
New results normally stay complete until the model consumes them. If an unseen batch alone is too large for the context, `fit_tool_results` persists its largest results and keeps a 1,000-character preview plus the full-output path. This avoids summarizing the entire history before the model can inspect the new result.
The first two steps run every round. Step 3 runs only when the context is above the limit. All three are deterministic text and structure operations; they do not add API calls.
The first two steps run every round. Step 3 runs only when the context is above the limit. All three are deterministic and recoverable text and structure operations; they do not add API calls.
## Step 4: compact_history
After `micro_compact`, the code estimates the context again with `estimate_chars(messages)`:
After `micro_compact` and `fit_tool_results`, the code estimates the context again with `estimate_chars(messages)`:
```python
CONTEXT_CHAR_LIMIT = 50000
@@ -181,13 +178,16 @@ This lesson uses character count as its trigger, and all related thresholds use
## Why the Order Is Fixed
The pipeline uses this order and only enters the lossy steps when necessary:
The pipeline uses this order and only enters the lossy summary step when necessary:
```python
messages = self.tool_result_budget(messages)
messages = self.snip_compact(messages)
if self.estimate_chars(messages) > self.CONTEXT_CHAR_LIMIT:
messages = self.micro_compact(messages)
target = int(self.CONTEXT_CHAR_LIMIT * 0.8)
messages = self.micro_compact(messages, target)
if self.estimate_chars(messages) > self.CONTEXT_CHAR_LIMIT:
messages = self.fit_tool_results(messages, target)
if self.estimate_chars(messages) > self.CONTEXT_CHAR_LIMIT:
messages = self.compact_history(messages, active_request)
```
@@ -195,7 +195,7 @@ if self.estimate_chars(messages) > self.CONTEXT_CHAR_LIMIT:
This order satisfies two constraints:
1. Steps 1 and 2 run every round. Step 3 runs only above the limit, and only Step 4 adds an API request.
2. `tool_result_budget` must run before `micro_compact`. Large results need to reach disk before older results can become placeholders.
2. Every shortened tool result keeps a trusted path inside `.task_outputs/tool-results/`; only a remaining overflow reaches model-generated history summarization.
Each round therefore starts with the lowest-cost operation whose information is easiest to recover.
@@ -310,7 +310,7 @@ Read the README.md files from s01_agent_loop through s05_todo_write.
Compare their top-level headings and summarize the naming pattern.
```
This task produces at least 5 file results. Every result remains complete until the model sees it once. On later turns, the latest 3 consumed results remain complete while older long results become `[Earlier tool result omitted.]`. A persisted result retains its saved path.
This task produces at least 5 file results. New results normally remain complete until the model sees them once; an oversized unseen result keeps a preview and recovery path instead. On later turns, the latest 3 consumed results remain complete while older long results become `[Earlier tool result saved at ...]` references.
### Experiment 2: Persist a Large Result