fix(context_compressor): harden tool-call previews in summaries

[sauerdaniel]

What does this PR do?

Hardens context-compaction summary rendering for assistant tool calls.

Compaction summaries currently render tool arguments as truncated function-call-shaped pseudo examples such as execute_code({ ... }). Those summaries are injected into later model context. This change renders bounded, JSON-safe descriptive previews as name args={preview}, reducing ambiguity while preserving useful argument context.

Related Issue

Fixes #12586

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• agent/context_compressor.py
  • Add _preview_tool_args_for_summary() for bounded, JSON-safe argument previews.
  • Bound string values, collection sizes, nesting depth, dictionary-key lengths, and total serialized size.
  • Preserve distinct values when shortening dictionary keys would otherwise collide.
  • Fall back to a compact valid-JSON envelope when a branching object exceeds the total preview budget.
  • Change summary serialization from name(args) pseudo calls to name args={preview}.
  • Preserve current sensitive-text redaction before preview construction.
• tests/agent/test_context_compressor.py
  • Cover short JSON, long/multiline values, long and colliding keys, total-size bounds, and non-call-shaped serializer output.

How to Test

1. Run scripts/run_tests.sh tests/agent/test_context_compressor.py -q.
2. Confirm long-key previews retain both values while bounding key length.
3. Confirm deeply branching arguments produce valid JSON within the configured budget.
4. Confirm summaries contain execute_code args= and not execute_code({.

Final focused verification after resolving the current-main redaction-helper rebase:

200 passed

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run the focused context-compressor suite on the final rebased head: 200 passed
• I've added tests for my changes
• I've tested on my platform: Fedora Linux x86_64

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — behavior is documented in helper docstrings and tests
• I've updated cli-config.yaml.example if I added/changed config keys — N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — pure serializer logic; no platform API
• I've updated tool descriptions/schemas if I changed tool behavior — N/A

Screenshots / Logs

/tmp/hermes-pr-venv/bin/python -m pytest tests/agent/test_context_compressor.py -q
200 passed in 8.39s

[teknium1]

Thanks for targeting a behavior that is still present on current main: agent/context_compressor.py:1506-1510 truncates raw arguments and emits name(args), and _generate_summary() consumes that serialization at agent/context_compressor.py:1778.

Problems

• The proposed helper's dictionary branch copies keys unchanged at agent/context_compressor.py:494 in the PR. max_items, max_depth, and _shorten_string() do not bound a very long key, so a valid arguments object can still make the supposedly bounded preview arbitrarily large.

Suggested changes

• Cap/replace overlong dictionary keys and add a regression test for that case, alongside the existing long-string-value coverage.

Automated hermes-sweeper review.

[sauerdaniel]

Rebased onto current main in 7747dbade. The reviewer-requested key bounds, collision preservation, collection/depth limits, and total serialized-preview cap remain covered. Verification: focused context-compressor suite 182 passed; full canonical suite at -j 32 passed 2,026 files / 41,622 tests; Ruff and git diff --check passed.

[sauerdaniel]

Rebased onto current main. The bounded JSON-safe preview, key-collision handling, depth/item limits, total serialized cap, and non-call-shaped name args= rendering remain intact. Focused context-compressor suite: 187/187; Ruff passed.

[sauerdaniel]

Final refresh onto the latest main: the focused context-compressor suite now passes 187/187; Ruff passes. The branch is one commit and zero commits behind upstream.

[andrexibiza]

Interlock completion for the context-compressor shard campaign:

• Current shard PR: refactor(context_compressor): extract durable session guards mixin (LB6) #81074
• Lock issue: Shard agent/context_compressor.py (god-file decomposition) #78645
• Campaign epic: [COMPLETE] Large-file decomposition: 20/20 done #78647
• This node is part of the same coordination graph; no issue is being closed by this refactor.

The full related set is explicitly recorded in #81074, including the prior compressor shards, same-surface colliders/fixers, and their referenced issue nodes. The residual agent/context_compressor.py remains alive until the 2k Law is satisfied.