v0.2.3

Fixed

• Sanitized Cline request message content before sending to /chat/completions to prevent empty content blocks that can cause 400 errors on Anthropic-backed models (e.g. Claude Sonnet 4.6).
• Normalized Cline message history to text-first blocks for Anthropic compatibility (dropping non-text assistant/tool-call parts and injecting fallback text when needed) to avoid messages.N content is empty failures.
• Hardened Cline context hook activation by inferring active provider from session model_change entries when ctx.model is unavailable.
• Coerced unknown message content shapes to safe fallback text to prevent downstream empty-content serialization.
• Collapsed per-turn Cline context into system + wrapped user transcript to avoid Anthropic tool-protocol/empty-content failures on follow-up turns.
• Made Cline transcript collapsing idempotent by reusing existing <task> content and appending only new turns, preventing recursive self-review loops on follow-up tool calls.
• Preserved tool-call intent in collapsed transcripts by attaching concise tool command summaries (e.g. bash command/path) to <tool_result> blocks.
• Reduced assistant self-repetition by excluding assistant orchestration-only tool-call turns from collapsed history while keeping actual tool outputs.
• De-duplicated repeated identical no-output tool results and injected a system note when the same command returns no output multiple times (including guidance for common git diff main...origin/main misuse).

v0.2.2

Fixed

• Aligned Cline identity/auth headers with current VS Code extension behavior (X-Platform, X-Client-Type, version headers).
• Improved OAuth code exchange robustness by handling missing provider hints and trying common provider values.
• Fixed recurring 403 access forbidden by shaping Cline provider user context into the envelope expected by api.cline.bot (<task>, task_progress block, and environment_details block).
• Rotated X-Task-ID more reliably by re-registering provider before agent start.

Changed

• Added optional PI_CLINE_API_BASE environment override for controlled API/proxy debugging.
• Added optional scaffold loading from local debug capture directories for Cline context-shaping parity.

Documentation

• Added update guidance and 403 troubleshooting steps to README.
• Updated remote/SSH auth instructions to prefer pasting full callback URL (including provider hint).

v0.2.1

Documentation

• Added remote/SSH authentication instructions to README.md.

v0.2.0

Changed

• Completely removed fallback model logic - now only uses dynamically fetched models from Cline's GitHub source.
• Changed model source from cli/src/constants/featured-models.ts to webview-ui/src/components/settings/OpenRouterModelPicker.tsx for more accurate free model list (includes stealth/giga-potato).
• Removed all hardcoded model-specific logic (no more special handling for "kimi", "minimax", "potato").
• Cache now stores full model objects instead of just IDs to avoid reconstruction with hardcoded assumptions.
• Removed all debug console logging for cleaner output.

Fixed

• Fixed regex to properly parse TypeScript array syntax (no semicolon after closing bracket).
• Models are now immediately available via /scoped-models and /model commands on startup when cached.

v0.1.7

Fixed

• Fixed compatibility with Pi v0.52.7+ standards.
• Made model fetching non-blocking to prevent UI/Slash command freezes on startup.
• Added background model updates after session start.
• Ensured all models have strictly valid metadata (cost, input, contextWindow) for latest Pi versions.
• Added fetch timeouts and defensive error handling during provider registration.

v0.1.6

Documentation

• Refined development instructions for clarity.

v0.1.5

Documentation

• Fixed broken banner image on npm by using an absolute GitHub raw URL.

v0.1.4

Documentation

• Updated README text to clarify model availability and authentication flow.

v0.1.3

Documentation

• Redesigned README.md with centered banner and badges to match ecosystem standards.

v0.1.2

CI/CD

• Added automated npm publishing workflow.