v1.3.186

v1.3.186

Overview

31 commits since v1.3.185. Major improvements in agent robustness, error handling, and performance.

Agent Intelligence & Safety (16 features)

• Session-level token budget — progressive warnings at 75%/90%/100% of context window, with actionable guidance to /compact or /clear
• Reasoning block compaction tier — preserves thinking blocks longer than tool results, compacts them only when budget pressure is high
• Semantic scope drift detection — warns when the agent drifts from the original task scope, preventing scope creep
• Consecutive edit failure recovery — after 3 consecutive edit failures, injects strategic recovery guidance (re-read file, use different anchor)
• Batch edit conflict detection — detects overlapping edits to the same file within a single multi_edit_file call
• Stale-read detection — detects when files have been modified externally (mtime check) between read and edit
• Unread-file edit guard — warns when editing a file that hasn't been read in the current session
• Mid-run stale todo detection — flags todo items that have been pending for too long, suggesting completion or removal
• Post-write file integrity validation — validates file content growth, detects null bytes and content loss after writes
• Debug statement detection — catches debug print/log statements left in code across 11 file types
• Proactive Go import analysis — detects unused/missing imports at write time, not waiting for build failure
• Secret detection and redaction — scans tool outputs for API keys, tokens, and passwords (12 regex patterns)
• Leftover artifact detection — catches merge conflict markers and suspicious content growth ratios
• Empty search spiral detection — breaks the cycle of repeated empty-result searches
• Recurring build/test error fingerprinting — detects repeated identical build errors
• Per-tool latency baseline outlier detection — flags tools that are unusually slow vs their historical baseline

Tool & UX Improvements (6 features)

• Smart commit workflow — conventional commit analysis with scope cohesion scoring and size recommendations
• Schema-guided parameter validation — validates required parameters before tool execution, catches missing args early
• Schema-aware argument coercion — auto-coerces unknown tool name suggestions and type mismatches
• Deterministic command result caching — caches identical build/test commands to skip redundant execution
• Consecutive-line compression — compresses repetitive tool output (e.g., 100 identical log lines) for better readability
• code_execution whitelist expanded — from 16 to 35 read-only tools for batch analysis

Performance (3 improvements)

• Go AST parsing deduplication — reduced from 3x to 1x parser.ParseFile per .go file write (saves 10-40ms per write)
• BM25 index memory/CPU optimization — reduced index memory footprint and CPU overhead
• scanContent truncation — CPU-protective truncation for regex operations on large tool outputs

Error Handling & Reliability (4 fixes)

• Vendor-specific rate limit and quota error patterns — added support for Anthropic HTTP 529 (overloaded), Gemini RESOURCE_EXHAUSTED, Kimi exceeded_current_quota_error and engine_overloaded_error, OpenAI insufficient_quota. These were previously misclassified, causing either no retry (transient errors) or wasteful 20x retry loops (permanent errors)
• Swarm permanent error loop fix — teammate tasks that fail with permanent errors (quota/auth) are now marked as completed-with-error instead of reverting to pending, preventing infinite retry loops across teammates
• Content growth ratio off-by-one — trailing newline caused false positive in 5x growth detection
• Memory eviction — count-based cap with category-priority eviction, preventing unbounded memory growth

Metrics

• 31 commits since v1.3.185
• 64 files changed, +9289 lines
• All tests passing: Go unit tests, Desktop tests, 113 Vitest frontend tests