feat: rule-based tool input compaction for auto-capture (~10x storage reduction)

[lg320531124]

Summary

Add rule-based tool input compaction to the Claude Code auto-capture pipeline. compactToolInput() reduces Write/Edit/Bash/TaskCreate/TaskUpdate tool inputs to structural summaries instead of full file contents.

Problem

Sampling 16 CC sessions shows 79% of session storage volume is full code content from Write (avg 11,575 chars) and Edit (avg 2,525 chars) tool inputs. This causes:

1. Storage bloat — sessions are ~100KB when ~20KB would suffice
2. Vector signal dilution — bge-large-zh-v1.5 embeddings are dominated by code tokens instead of semantic content
3. VLM extraction failure — 1081 sessions with 0 memories extracted, likely because content is too large for the VLM ReAct loop to process effectively

Solution

No LLM involved — purely rule-based compaction:

Tool	Before	After	Ratio
Write	Full file content (~7KB)	file_path + "100 lines, 7200 chars" + 200-char preview (~280 chars)	~25x
Edit	old_string + new_string (~5KB)	file_path + length summary + 150-char previews (~400 chars)	~12x
Bash	command + description	command only	~1.5x
TaskCreate/Update	All fields	subject + status + taskId	~3x
Read/Glob/Grep/LSP	Full preservation	Unchanged	1x

Measured: 10.6x compression on realistic mixed session data.

Changes

• auto-capture.mjs: Add compactToolInput() with per-tool-type policies (TOOL_INPUT_POLICIES); modify harvestContent() to use compaction when enabled
• config.mjs: Add toolInputCompaction (default: true) and toolInputMaxChars (default: 2000) config options with env var overrides
• __tests__/auto-capture-compaction.test.mjs: 21 regression tests covering all tool types, truncation, edge cases, and compression ratio

Configuration

Env Var	Default	Description
OPENVIKING_TOOL_INPUT_COMPACTION	true	Enable/disable compaction
OPENVIKING_TOOL_INPUT_MAX_CHARS	2000	Global truncation cap (0 = disabled)

Set OPENVIKING_TOOL_INPUT_COMPACTION=false to restore the old verbatim behavior.

Test Plan

• 21 regression tests pass
• 10.6x compression ratio on realistic mixed tool call data
• JSON round-trip validation for all summary-mode tools
• Edge cases: empty content, empty old_string, maxChars=0, string input pass-through
• Manual verification: run a CC session with compaction enabled, check messages.jsonl size reduction
• VLM extraction comparison: same session before/after compaction, compare memory extraction success

Closes #2875

[itxaiohanglover]

Nice implementation! Rule-based tool input compaction without LLM involvement is a smart approach — reduces cost and latency for auto-capture. The inline test functions avoid coupling to the full module. Tool input policies (full vs summary) are well-designed.

[lg320531124]

The 06. API & CLI Integration Tests job fails here too with the identical tally — 3 failed, 25 passed, 7 skipped, 11 errors, all CONFLICT: Resource is busy on the session-scoped cli_test_4f5946d2 fixture.

I posted the full diagnosis on sibling PR #2900 (rebased onto latest main, head b00e336, same failure → rules out base-skew; two PRs touching disjoint code producing byte-identical failure counts = shared infra flake, not a code defect in either branch). Root cause points at the TreeLock-release path in openviking/utils/resource_processor.py:441 — retryable=True but the 15×@10s retry never recovers within the session fixture's lifetime.

This PR (feat/tool-input-compaction) touches tool-input compaction, not add-resource/locks/conftest, so the failing setup step is never reached by my code.

I don't have admin rights to rerun from a fork — could someone with access rerun the 06. API & CLI Integration Tests job on this branch? If it passes on retry that confirms the flake. (No code changes from my side — happy to act if it points back at the branch.)

[lg320531124]

Opened #2916 to track the Resource is busy TreeLock flake at the root-cause level (CI-infra, not this PR). This PR and #2900 hit the identical tally on the same run despite touching disjoint modules — evidence chain and the rerun ask are in the issue.