Google streaming may emit thoughtSignature without thought=true (including empty-text signature-only parts). Treat non-empty thoughtSignature as thinking to avoid leaking reasoning into normal text and retain signature across streaming deltas. Add unit test coverage.
Previously all openai-codex models had pricing set to 0, causing the
TUI to always show $0.00 for cost tracking.
Updated pricing based on OpenAI Standard tier rates:
- gpt-5.2/gpt-5.2-codex: $1.75/$14.00 per 1M tokens
- gpt-5.1/gpt-5.1-codex/gpt-5.1-codex-max: $1.25/$10.00 per 1M tokens
- gpt-5/gpt-5-codex: $1.25/$10.00 per 1M tokens
- codex-mini-latest: $1.50/$6.00 per 1M tokens
- gpt-5-mini/gpt-5.1-codex-mini/gpt-5-codex-mini: $0.25/$2.00 per 1M tokens
- gpt-5-nano: $0.05/$0.40 per 1M tokens
Source: https://platform.openai.com/docs/pricing
- Add sessionId to StreamOptions for providers that support session-based caching
- OpenAI Codex provider uses sessionId for prompt_cache_key and routing headers
- Agent class now accepts and forwards sessionId to stream functions
- coding-agent passes session ID from SessionManager and updates on session changes
- Update ai package README with table of contents, OpenAI Codex OAuth docs, and env vars table
- Increase Codex instructions cache TTL from 15 minutes to 24 hours
- Add tests for sessionId forwarding in ai and agent packages
- Remove per-thinking-level model variants (gpt-5.2-codex-high, etc.)
- Remove thinkingLevels from Model type
- Provider clamps reasoning effort internally
- Omit reasoning field when thinking is off
fixes#472
- Add LoginDialogComponent with proper borders (top/bottom DynamicBorder)
- Refactor all OAuth providers to use racing approach (browser callback vs manual paste)
- Add onEscape handler to Input component for cancellation
- Add abortable sleep for GitHub Copilot polling (instant cancel on Escape)
- Show OS-specific click hint (Cmd+click on macOS, Ctrl+click elsewhere)
- Clear content between login phases (fixes GitHub Copilot two-phase flow)
- Use InteractiveMode's showStatus/showError for result messages
- Reorder providers: Anthropic, ChatGPT, GitHub Copilot, Gemini CLI, Antigravity
Previously, users had to wait up to 60 seconds for the browser callback
to timeout before being prompted to paste the authorization code. This
was problematic for SSH/VPS sessions where the callback cannot work.
Now the paste input is shown immediately alongside the browser flow:
- Browser callback and manual paste race - whichever completes first wins
- Desktop users: browser callback succeeds, input is cleaned up
- SSH/VPS users: paste code immediately without waiting
Changes:
- Add cancelWait() to OAuth server for early termination of polling loop
- Add onManualCodeInput callback that races with browser callback
- Show paste input immediately in TUI for openai-codex provider
- Clean up input on success, error, or when browser callback wins
Co-authored-by: cc-vps <crcatala+vps@gmail.com>
Previously the system prompt was converted to an input message in convertMessages,
then stripped out by filterPiSystemPrompts. Now the system prompt is passed directly
to transformRequestBody and appended after CODEX_PI_BRIDGE in the bridge message.
- /tools command opens SettingsList-based selector for all loaded tools
- Space/Enter toggles individual tools between enabled/disabled
- Changes apply immediately on toggle (like /settings)
- Tool selection persisted to session via appendEntry()
- State restored from current branch on session_start, session_tree, session_branch
- Uses getBranch() to respect branch-specific tool configuration
- Export getSettingsListTheme and getSelectListTheme for hooks to use
Claude requires max_tokens > thinking.budget_tokens. When caller specifies
a small maxTokens (e.g. compaction with ~13k tokens) and reasoning is enabled
with high budget (16k tokens), the constraint was violated.
Fix: In mapOptionsForApi, add thinkingBudget on top of caller's maxTokens
(capped at model.maxTokens). If still not enough room, reduce thinkingBudget
to leave space for output.
Applied to both anthropic-messages and google-gemini-cli APIs.
Also adds test utilities for OAuth credential resolution and tests for
compaction with thinking models.
fixes#413
- Implement google-vertex provider in packages/ai
- Support ADC (Application Default Credentials) via @google/generative-ai
- Add Gemini model catalog for Vertex AI
- Update packages/coding-agent to handle google-vertex provider
- Add tree sidebar with search and filter (Default/All/Labels)
- Client-side markdown/syntax highlighting via vendored marked.js + highlight.js
- Base64 encode session data to avoid escaping issues
- Reuse theme.ts color tokens via getResolvedThemeColors()
- Sticky sidebar, responsive mobile layout with overlay
- Click tree node to scroll to message
- Keyboard shortcuts: Esc to reset, Ctrl/Cmd+F to search
- Add GoogleThinkingLevel type mirroring Google's ThinkingLevel enum
- Update GoogleGeminiCliOptions and GoogleOptions to use our type
- Cast to any when assigning to Google SDK's ThinkingConfig
- Check response.stopReason instead of catching errors for abort detection
- Return result object from _generateBranchSummary instead of throwing
- Fix summary attachment: attach to navigation target position, not old branch
- Support root-level summaries (parentId=null) in branchWithSummary
- Remove setTimeout hack for re-showing tree selector on abort
- Add TreeSelectorComponent with ASCII tree visualization
- Add AgentSession.navigateTree() for switching branches
- Add session_before_tree/session_tree hook events
- Add SessionManager.resetLeaf() for navigating to root
- Change leafId from string to string|null for consistency with parentId
- Support optional branch summarization when switching
- Update buildSessionContext() to handle null leafId
- Add /tree to slash commands in interactive mode
Hook commands now always execute immediately, even during streaming.
If a command needs to interact with the LLM, it uses pi.sendMessage()
which handles queueing automatically.
This simplifies the API and eliminates the issue of queued slash
commands being sent to the LLM instead of executing.
- Add pkce.ts with generatePKCE() using Web Crypto API
- Update anthropic.ts, google-gemini-cli.ts, google-antigravity.ts
- Replace Buffer.from() with atob() for base64 decoding
- Works in both Node.js 20+ and browsers
The OAuth modules still use Node.js http.createServer for callbacks,
so they only work in CLI environments, but they no longer crash on
import in browser bundles.
- Renamed AppMessage to AgentMessage throughout
- New agent-loop.ts with AgentLoopContext, AgentLoopConfig
- Removed transport abstraction, Agent now takes streamFn directly
- Extracted streamProxy to proxy.ts utility
- Removed agent-loop from pi-ai (now in agent package)
- Updated consumers (coding-agent, mom) for AgentMessage rename
- Tests updated but some consumers still need migration
Known issues:
- AgentTool, AgentToolResult not exported from pi-ai
- Attachment not exported from pi-agent-core
- ProviderTransport removed but still referenced
- messageTransformer -> convertToLlm migration incomplete
- CustomMessages declaration merging not working properly
HookMessage (role: hookMessage) was being passed directly to transport
without transformation. Now it's transformed via messageTransformer
which converts it to a proper user message for the LLM.
Cleans up the temporary emitLastMessage plumbing since we now use
Agent.prompt(AppMessage) for hook messages instead of appendMessage+continue.
- Remove emitLastMessage parameter from Agent.continue()
- Remove from transport interface and implementations
- Remove from agentLoopContinue()
Following the same pattern as BashExecutionMessage:
- HookAppMessage has role: 'hookMessage' with customType, content, display, details
- isHookAppMessage() type guard for checking message type
- messageTransformer converts to user message for LLM context
- TUI checks isHookAppMessage() for rendering as CustomMessageComponent
This makes the API clean for anyone building on AgentSession - they can
use the type guard instead of knowing about internal marker fields.
When calling continue() with emitLastMessage=true, the agent loop
emits message_start/message_end events for the last message in context.
This allows messages added outside the loop (e.g., hook messages via
sendHookMessage) to trigger proper TUI rendering.
Changes across packages:
- packages/ai: agentLoopContinue() accepts emitLastMessage parameter
- packages/agent: Agent.continue(), transports updated to pass flag
- packages/coding-agent: sendHookMessage passes true when triggerTurn
- getEnvApiKey: ANTHROPIC_OAUTH_TOKEN now takes precedence over ANTHROPIC_API_KEY
- findCutPoint: Stop scan-backwards loop at session header (was decrementing past it causing null preparation)
- generateSummary/generateTurnPrefixSummary: Throw on stopReason=error instead of returning empty string
- Test files: Fix API key priority order, use keepRecentTokens=1 for small test conversations
