Flush final text before closing each AI SDK text block, surface event-processing failures to chat callers, and clear the remaining Companion OS check blockers.
fixes#273
Co-authored-by: Codex <noreply@openai.com>
Restore the bedrock provider import through an explicit workspace path mapping, align pi-grind package metadata with the existing package manifests, remove the unused controller-failure state path, and parse bare ISO dates in local time.
Co-authored-by: Codex <noreply@openai.com>
Replace the old project-scoped file memory plumbing with runtime-native conversational memory and remove obsolete pi-memory-md shipping/wiring.
Co-authored-by: Codex <noreply@openai.com>
Remove shell-based memory tree inspection and pre-sync before memory pushes.
Switch sync update detection to git HEAD comparisons instead of localized output parsing.
Co-authored-by: Codex <noreply@openai.com>
Add a first-class browser tool backed by agent-browser and enable it by default in coding sessions.
Include CLI and system-prompt wiring plus focused coverage for the new tool.
Co-authored-by: Codex <noreply@openai.com>
Bash tool had no default timeout, so commands like `npx serve` would
block the agent loop forever. This caused the randomtrends sandbox
thread to die - the LLM started a dev server, the shell never exited,
and the entire session hung with no way to recover.
Two safeguards:
- Default 120s timeout when LLM does not specify one
- 30s no-output timeout that kills processes that go silent
Both are configurable via BashToolOptions and can be disabled by
setting to 0. The LLM can still pass explicit higher timeouts for
long builds.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move the gateway runtime files into a dedicated core/gateway module and fix follow-up issues in session deletion, history import batching, message IDs, and legacy thread parsing.
Fixes#253
Co-authored-by: Codex <noreply@openai.com>
Add a Convex-backed durable chat layer for companion threads, recover gateway sessions from persisted Pi session files after eviction, and start splitting gateway runtime internals into focused modules.
Co-authored-by: Codex <noreply@openai.com>
- Copy all pi-mono source into apps/companion-os/
- Update Dockerfile to COPY pre-built binary instead of downloading from GitHub Releases
- Update deploy-staging.yml to build pi from source (bun compile) before Docker build
- Add apps/companion-os/** to path triggers
- No more cross-repo dispatch needed
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
