Messages submitted while a branch summary was being generated were
processed immediately instead of being queued. This happened because
isCompacting only checked compaction abort controllers, not the branch
summary abort controller.
Include _branchSummaryAbortController in the isCompacting getter so all
existing guards (message queueing, reload blocking) also apply during
branch summarization.
The extension system currently only forwards agent_start, agent_end,
turn_start, and turn_end events. This means extensions cannot access
streaming text (token-by-token), message lifecycle, or tool execution
progress — all of which are available to internal subscribers.
This adds forwarding for the remaining 6 agent event types:
- message_start, message_update, message_end
- tool_execution_start, tool_execution_update, tool_execution_end
These follow the exact same pattern as the existing forwarded events:
new interfaces in types.ts, exports in index.ts, and else-if blocks
in _emitExtensionEvent(). The new types are included in ExtensionEvent
and automatically flow through RunnerEmitEvent (they're not in the
exclusion list).
This enables extensions to build real-time UIs, streaming WebSocket
bridges, and other integrations that need fine-grained event access.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Mario Zechner <badlogicgames@gmail.com>
After compaction, context token count is unknown until the next LLM
response. Instead of showing stale pre-compaction values or heuristic
estimates, the footer now shows ?/200k.
ContextUsage.tokens and ContextUsage.percent are now number | null
(breaking change). Removed usageTokens, trailingTokens, lastUsageIndex
from ContextUsage (internal details).
Also fixed _checkCompaction() using .find() (first compaction) instead
of getLatestCompactionEntry() (latest), which caused incorrect overflow
detection with multiple compactions.
Closes#1382
Include tool parameter names, types, descriptions in a collapsible
section under each tool in the export HTML. Also adds parameters to
pi.getAllTools() return value.
closes#1407, closes#1416
prompt() computed currentImages but never passed them to _queueSteer()
or _queueFollowUp() in the streaming branch. Both methods only accepted
text and built content as [{ type: 'text', text }], dropping images.
- _queueSteer/_queueFollowUp now accept optional ImageContent[]
- streaming branch in prompt() passes currentImages through
- public steer()/followUp() accept and forward optional images
- RPC types, handler, and client updated for steer/follow_up images
- rpc.md: document images on steer/follow_up, fix ImageContent examples
- Store previousSessionFile before creating branched session
- Update sessionFile after writing new branch file
- Pass parentSession when forking from first message
- Add --local to git config in tests to prevent repo escape
When a provider (e.g., Google Gemini CLI) requests a retry delay longer
than maxDelayMs (default: 60s), the request fails immediately with an
informative error instead of waiting silently for hours.
The error is then handled by agent-level auto-retry, which shows the
delay to the user and allows aborting with Escape.
- Add maxRetryDelayMs to StreamOptions (packages/ai)
- Add maxRetryDelayMs to AgentOptions (packages/agent)
- Add retry.maxDelayMs to settings (packages/coding-agent)
- Update _isRetryableError to match 'retry delay' errors
fixes#1123
switchSession() was appending spurious thinking_level_change entries
to session log on resume because setThinkingLevel() unconditionally
persisted. Now only persists if the level actually changes.
fixes#1118
Adds a method to access the effective system prompt (after any per-turn
extension modifications) from the extension context.
Implementation:
- Add systemPrompt getter to AgentSession reading from agent.state.systemPrompt
- Wire getSystemPrompt through ExtensionContextActions to ExtensionRunner
- Add getSystemPrompt to interactive-mode's shortcut context
- Update docs with ctx.getSystemPrompt() section
- Add system-prompt-header.ts example
- Add example to docs reference table
Closes#1098
- Add set_session_name command with empty name validation
- Expose sessionName in get_state response
- Add setSessionName() to AgentSession and RpcClient
- Document in docs/rpc.md
Previously, within a single tool-use turn, rate limit retries would
accumulate across separate LLM calls. For example, if each of 3 tool
calls hit a 429 and retried once, the counter would show '3/3' and fail
even though each individual retry succeeded.
Now the counter resets immediately when a successful (non-error)
assistant message arrives, so each LLM call gets a fresh set of retries.
Fixes#1019
Move setup callback handling from interactive/rpc modes into AgentSession.newSession().
After setup() runs, sync agent state via replaceMessages() so the LLM has context
and the UI renders the messages properly.
fixes#968
- Add resetApiProviders() to clear and re-register built-in providers
- Add createAssistantMessageEventStream() factory for extensions
- Add streamSimple support in ProviderConfig for custom API implementations
- Call resetApiProviders() on /reload to clean up extension providers
- Add custom-provider.md documentation
- Add custom-provider.ts example with full Anthropic implementation
- Update extensions.md with streamSimple config option
- Package deduplication: same package in global+project, project wins
- Collision detection for skills, prompts, and themes with ResourceCollision type
- PathMetadata tracking with parent directory lookup for file paths
- Display improvements: section headers, sorted groups, accent colors for packages
- pi list shows full paths below package names
- Extension loader discovers files in directories without index.ts
- In-memory SettingsManager properly tracks project settings
fixes#645
- Add ParsedSkillBlock interface and parseSkillBlock() function
- Change skill expansion to use XML-style <skill> tags
- Add SkillInvocationMessageComponent for collapsible display
- Collapsed: single line with skill name and expand hint
- User message rendered separately after skill block
Fixes#894
There's only ever one bindings instance per session, so the Set/Array
approach was unnecessary. Changed from Set<ExtensionErrorListener> to
optional single listener.
- Add ResourceLoader interface and DefaultResourceLoader implementation
- Add PackageManager for npm/git extension sources with install/remove/update
- Add session.reload() and session.bindExtensions() APIs
- Add /reload command in interactive mode
- Add CLI flags: --skill, --theme, --prompt-template, --no-themes, --no-prompt-templates
- Add pi install/remove/update commands for extension management
- Refactor settings.json to use arrays for skills, prompts, themes
- Remove legacy SkillsSettings source flags and filters
- Update SDK examples and documentation for ResourceLoader pattern
- Add theme registration and loadThemeFromPath for dynamic themes
- Add getShellEnv to include bin dir in PATH for bash commands
Detect OAuth authentication failures (expired credentials, offline) and provide helpful error message instead of crashing with generic 'No API key found' error.
Co-authored-by: Mario Zechner <badlogicgames@gmail.com>
When auto-compaction fails (e.g., quota exceeded), emit the error via
the auto_compaction_end event instead of throwing. The UI now displays
the error message, allowing users to take action (switch models, wait
for quota reset, etc.) instead of crashing.
fixes#792
When 429/500 errors occur during tool execution, empty assistant messages
with stopReason='error' get persisted. These break the tool_use -> tool_result
chain for Claude/Gemini APIs.
Added centralized filtering in transformMessages to skip assistant messages
with empty content and no tool calls. Provider-level filters remain for
defense-in-depth.
- Skill commands (/skill:name) now expanded in AgentSession instead of
interactive mode, enabling them in RPC and print modes
- Input event can now intercept /skill:name before expansion
- Updated extensions.md with clearer input event docs and processing order
- Updated rpc.md: hook -> extension terminology, added skill expansion mentions
- Added PR attribution to changelog entries for #761
* feat(coding-agent): add input event for extension input interception
Extensions can now intercept, transform, or handle user input before the
agent processes it. Three result types: continue (pass through), transform
(modify text/images), handled (respond without LLM). Handlers chain
transforms and short-circuit on handled. Source field identifies origin.
* fix: make source public, use if/else over ternary
* fix: remove response field, extension handles own UI
* Improve Cloud Code Assist error messages
- Extract just the message from verbose JSON error responses
- Extract cause from generic 'fetch failed' errors for better diagnostics
* Make 'other side closed' network error retryable
* Make 'other side closed' network error retryable
Breaking change: pi.getAllTools() now returns Array<{ name, description }>
instead of string[]. Extensions needing just names can use .map(t => t.name).
Removes redundant getToolInfo() method added in original PR.
Fixes#647
- New /models command with toggle UI for each available model
- Changes persist to enabledModels in settings.json
- Updates take effect immediately for Ctrl+P cycling
The --no-skills flag set options.skills = [] in main.ts, but the
interactive mode UI would rediscover skills anyway because it called
loadSkills() directly instead of using the already-loaded skills.
Changes:
- Add AgentSession.skills and AgentSession.skillWarnings properties
- discoverSkills() now returns { skills, warnings } instead of Skill[]
- Interactive mode uses session.skills instead of calling loadSkills()
- Update SDK docs and examples for new return type
Fixes#577
