Skip to content

feat(compressor): smart tool output collapse during context pruning#9661

Closed
kshitijk4poor wants to merge 1 commit into
NousResearch:mainfrom
kshitijk4poor:feat/smart-tool-collapse
Closed

feat(compressor): smart tool output collapse during context pruning#9661
kshitijk4poor wants to merge 1 commit into
NousResearch:mainfrom
kshitijk4poor:feat/smart-tool-collapse

Conversation

@kshitijk4poor

@kshitijk4poor kshitijk4poor commented Apr 14, 2026

Copy link
Copy Markdown
Collaborator

Summary

Replace the generic [Old tool output cleared to save context space] placeholder with informative 1-line summaries during the pre-compression pruning pass.

Before:

[Old tool output cleared to save context space]

After:

[terminal] ran `npm test` -> exit 0, 47 lines output
[read_file] read src/auth.py from line 40 (1,203 chars)
[search_files] content search for 'parse_token' in src/ -> 4 matches
[patch] replace in src/auth.py (279 chars result)
[browser_navigate] https://docs.python.org/3/library/re.html (26,034 chars)
[delegate_task] 'Analyze auth middleware for vulnerabilities' (28,139 chars result)

How it works

  1. New _summarize_tool_result() function generates per-tool-type summaries by parsing the tool name + arguments from the preceding assistant message's tool_calls
  2. _prune_old_tool_results() builds a call_id -> (tool_name, args) index, then calls the summarizer instead of inserting the generic placeholder
  3. Covers all built-in tools with a generic fallback for unknown/plugin tools

Benchmark results

Metric OLD (placeholder) NEW (smart summary)
Space savings 99.3% 98.9%
Key facts preserved 0/24 (0%) 24/24 (100%)
Extra cost ~70 tokens per compression

The model now knows what each tool did (file path, command, exit code, match count) without re-reading the full output.

Test plan

  • All 63 existing compressor/engine/focus tests pass
  • No changes to public API — _prune_old_tool_results is internal

Part of #9666.

Replace the generic '[Old tool output cleared to save context space]'
placeholder with informative 1-line summaries that preserve the tool
name, key parameters, and outcome.

Before: [Old tool output cleared to save context space]
After:  [terminal] ran `npm test` -> exit 0, 47 lines output
        [read_file] read src/auth.py from line 40 (1,203 chars)
        [search_files] content search for 'parse_token' in src/ -> 4 matches

Covers all built-in tools (terminal, read_file, write_file, search_files,
patch, browser_*, web_search, web_extract, delegate_task, execute_code,
skill_view, memory, todo, etc.) with a generic fallback for unknown tools.

Space savings are near-identical (~98.9% vs ~99.3% for placeholder) since
both replace large content with short text, but the new summaries carry
actual information: the model knows WHAT each tool did without re-reading
the full output.
teknium1 pushed a commit that referenced this pull request Apr 15, 2026
…rade, hardening

Combined salvage of PRs #9661, #9663, #9674, #9677, #9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs #9665 (user msg preservation — duplication risk) and #9675 (dead code).
teknium1 pushed a commit that referenced this pull request Apr 15, 2026
…rade, hardening

Combined salvage of PRs #9661, #9663, #9674, #9677, #9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs #9665 (user msg preservation — duplication risk) and #9675 (dead code).
@teknium1

Copy link
Copy Markdown
Contributor

Merged via #10088. Your changes were salvaged onto current main with your authorship preserved. Thanks for the contribution! 🎉

@teknium1 teknium1 closed this Apr 15, 2026
ulasbilgen pushed a commit to ulasbilgen/hermes-adhd-agent that referenced this pull request May 1, 2026
…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).
aj-nt pushed a commit to aj-nt/hermes-agent that referenced this pull request May 1, 2026
…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).
02356abc pushed a commit to 02356abc/hermes-agent that referenced this pull request May 14, 2026
…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).
gweeteve pushed a commit to gweeteve/hermes-agent that referenced this pull request Jun 2, 2026
…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).
Egavasyug pushed a commit to Egavasyug/hermes-agent that referenced this pull request Jun 10, 2026
…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants