v1.3.187

Overview

48 commits since v1.3.186. Major improvements in subagent UX, system prompt performance, provider resilience, and TUI stability.

Preserve spaces in reasoning chunks — DeepSeek reasoning tokens were having spaces stripped between words (every10sevenwhentheappisidle). Root cause: strings.TrimSpace() on every streaming chunk. Fixed: LLM output is now passed verbatim, never trimmed.
Never trim whitespace from LLM output — Extended fix to all text/reasoning paths: empty checks use == "" instead of strings.TrimSpace().
Reasoning event coalescing — extended text coalescing to AgentEventReasoning so hundreds of tiny reasoning chunks don't overflow the event buffer.
Subagent retry display fix — use Contains instead of HasPrefix for retry detection, preventing false matches.
Text content preservation during eviction — when all-text events are evicted, dropped text is prepended to remaining events to prevent word boundary corruption.
Text event coalescing tuning — increased event buffer and tuned merge interval for better streaming feel.
Subagent system message naming — prefix system messages with agent name for clarity.
Model display in named agent labels — show model name in use_namedagent tool call labels and follow strip.

Background section collector — system prompt sections (modified files, package symbols, project overview, toolchain) are pre-computed in a background goroutine every 10s, eliminating synchronous I/O on every prompt rebuild. Applied to TUI, daemon, and desktop.
modifiedFilesSection caching — eliminates per-submit latency by caching the section between collector ticks.
Follow panel rebuild skipping — skip expensive list rebuild when event count and status haven't changed.
Post-edit LSP diagnostics timeout reduction — faster file edit feedback.

Provider/Model failover — transparently switches to a backup LLM provider when the primary fails with permanent errors (quota/auth) or sustained transient errors. Configurable via fallback: section in ggcode.yaml.
Fail-fast on quota/auth errors — isRetryable now immediately returns false for quota and authentication errors, preventing wasteful 20x retry loops.

Smart file path suggestions — when read_file/edit_file fails on a non-existent path, returns fuzzy-matched suggestions ("did you mean: path/to/correct_file?"), saving 1-3 agent iterations per incorrect guess.
Session trace export — new /export-trace [session-id] command exports structured JSON execution traces (TTFT, tool latency, token usage, cache stats) for observability and offline analysis.
Cross-file LSP diagnostics — SiblingDiagnostics detects errors in related files after edits.
Configurable completion notifications — cross-platform desktop notifications when the agent completes a task.

os.Stderr fd-level redirect — pion/turn's "Fail to refresh permissions" error was corrupting the TUI by writing directly to stderr. Fixed: os.Stderr is now replaced with a pipe that routes all writes to debug.Log, catching both log.New(os.Stderr, ...) and fmt.Fprint(os.Stderr, ...).
Desktop stderr redirect — same fix applied to desktop main.go.
Output style persistence — /style and Ctrl+O now persist the change to config.

Auto-inject modified file names into system prompt — agent sees which files were recently modified without explicit reads.
Adaptive reasoning effort allocation — adjusts reasoning depth per turn based on task complexity.
Multi-language delimiter balance validation — post-edit integrity checks for bracket/paren matching across multiple languages.
Workspace file protection guardrails — prevents accidental overwrites of critical files.
Placeholder/stub code detection — write integrity checks detect incomplete implementations.