Skip to content

CHANGELOG.md

Latest commit

 

History

History
213 lines (180 loc) · 14.7 KB

CHANGELOG.md

File metadata and controls

213 lines (180 loc) · 14.7 KB

Changelog

All notable changes to GoClaw are documented here. For full documentation, see docs.goclaw.sh.

Unreleased

Added

  • Behavior UX sidecar delivery overrides — Adds sidecar-generated Quick Acknowledgement and Intermediate Replies with provider/model, timeout, token, and char caps. Effective config resolves Channel > Agent > Workspace, with agent overrides stored in other_config.delivery_behavior.

  • Built-in skill workspace-organizing — closes #71. Discipline skill that teaches agents to keep personal, team, and delegate workspaces tidy. Enforces a purpose-based folder convention with two modes: flat (notes/, data/, outputs/, scripts/, archive/) for ad-hoc work and project (projects/<slug>/{docs,assets,source,reports,research}/) for named multi-file work. Per-agent namespacing under shared/<agent_key>/ prevents collisions in team workspaces. Integrates pre-write discovery via vault_search, memory_search, and knowledge_graph_search to surface related files before writing and avoid duplicates; documents Vault scope mirroring and id-routing rules.

  • Skill agent manage grants — Adds per-agent skill edit/delete grants with backend checks, HTTP/WS support, SQLite and PostgreSQL schema updates, and web dashboard controls for granting and revoking manage access.

  • Packages Update Flow (Phase 2a: pip + npm) — closes #900 (Phase 2a). Extends Phase 1 update infrastructure to pip and npm package sources. /v1/packages/updates now returns mixed-source results with an availability: {github, pip, npm} map. Multi-source UI with per-source filter pills; unavailable sources (binary not on PATH or Lite edition) hidden automatically. apk deferred to Phase 2b. See docs/packages-pip-npm.md for command matrix, runbook, and min versions.

  • Packages Update Flow (Phase 1: GitHub binaries) — closes #900. Proactive "N updates available" badge + per-row [Update] + [Update All] on the Runtime & Packages page. Backend endpoints under /v1/packages/updates* (master-scope). ETag-aware polling (304 responses don't burn rate limit), stale-while-revalidate cache, atomic two-phase .bak swap with rollback. Pre-release detection via regex + GitHub API flag; semver ordering via golang.org/x/mod/semver; non-semver tags use string-inequality fallback with downgrade protection. WebSocket events package.update.* for owner clients. See docs/packages-github.md § "Updating Installed Packages".

Changed

  • Behavior UX simplification — Retires user-facing Tool Status Messages and deterministic tool-status channel text. Show Reasoning remains separate for debugging/testing, while Quick Acknowledgement and Intermediate Replies are delivery-only sidecar messages. Legacy block_reply config remains readable as an inherited Intermediate Replies default but is no longer exposed as a separate Web UI control.

  • ChatGPT Subscription (OAuth) — default model and backend-owned model catalog now prefer gpt-5.5, with reasoning metadata and context-window defaults updated for provider-first model selection.

Fixed

  • Quick Acknowledgement generated mode — Generated acknowledgements now use the sidecar delivery generator instead of always falling back to fixed templates. Sidecar failures stay non-blocking and fall back to templates.

  • Intermediate Replies are sidecar-generated — Tool-call progress no longer appends fixed "I'll use ..." text or relies on main-pipeline assistant content. Visible progress is generated from bounded delivery metadata and is kept out of session history.

  • Multi-attachment messages no longer trigger N agent replies (#63). Three coalescing surfaces hardened so a single user action produces ONE agent run regardless of how the platform delivers attachments:

    1. Bus debouncer — removed the media-bypass shortcut that fired immediately for any message with attachments; media now goes through the same per-(channel, chatID, senderID, agentID) silence window as text. Media-floor (max(configured, mediaFloor)) guarantees a minimum window when attachments are present so multi-file uploads coalesce. Dedup seed prevents the same MessageID from being buffered twice on bursty arrivals.
    2. Web Chat debouncer (internal/gateway/methods/chat_debounce.go) — parallel structure for /v1/chat/completions streaming: per-session buffer + media floor + Take/Discard semantics for flush/cancel control. Merges queued payloads at flush time (latest params win; text concatenated newline-separated).
    3. Telegram album aggregator (internal/channels/telegram/album_aggregator.go) — channel-layer coalescing for albums. Telegram delivers a media-group (multiple photos/videos shared as one user action) as N separate Message updates sharing a MediaGroupID. The aggregator buffers by (chatID, MediaGroupID) after all access gates pass, pins the sender on first arrival as a security tripwire, and dispatches ONE processResolvedMessage call with all members on a 500ms silence window. Stop() synchronously drains pending buffers before pollCancel so in-flight albums always publish.

    Cross-surface invariants (see CONTRIBUTING.md → "Multi-attachment coalescing"): no media bypass, media floor on every surface, drop-and-log dual caps, no time.Timer.Reset (use AfterFunc + Stop), sender pin on first arrival, post-stop pushes rejected with warn log.

  • Upstream critical security remediation — hardens gateway no-token fallback, Feishu/Lark and Pancake webhooks, sandbox path/write handling, tenant-admin checks for mutable HTTP surfaces, and Lite hook schema migration verification.

  • SecureCLI runtime npm binaries — binary discovery and credentialed exec now resolve tools installed under the GoClaw runtime directories, including {runtimeDir}/npm-global/bin, and support single-binary npm package aliases such as openrouter-cli exposing orc.

Breaking Changes

  • Context pruning now opt-in. Previously tool-result trimming ran by default for all providers; now requires explicit contextPruning.mode: "cache-ttl" in config.agents.defaults to enable. Matches upstream TS design and prevents silent prompt-cache invalidation on Anthropic.

    Migration — add to config.json5:

    agents: {
  defaults: {
    contextPruning: { mode: "cache-ttl" }
  }
}

New Features

  • Pancake private-reply (comment → DM). Enables a one-time DM to commenters after the public reply. Stateless on GoClaw side — no DB dedup table, no in-memory state:
    • Config: features.private_reply (bool) + private_reply_message (text).
    • Template variables {{commenter_name}} and {{post_title}} with literal-replace semantics (pre-sanitizes {{/}} from var values to prevent var-in-var substitution).
    • Empty private_reply_message → English fallback constant.
    • Dedup strategy: webhook-level comment_id dedup (already in comment_handler.go) + Facebook's per-comment idempotent private_replies endpoint handle duplicates platform-side. No GoClaw state required.
    • No DB migration.

Improvements

  • Context pruning cleanup. Removed redundant Pass 0 (per-result 30% guard), deduplicated double prune call per iteration, added SanitizeHistory to PruneStage for broken tool_use/tool_result pair cleanup.
  • Context pruning config backfill (migration). Agents with existing custom context_pruning config (e.g., softTrimRatio, keepLastAssistants) but missing a mode field get auto-backfilled with mode: "cache-ttl" to preserve their intent after the opt-in flip. Rows with NULL config stay NULL (new opt-in default applies). PG migration 51; SQLite schema v19.
  • Pancake channel metadata routing. Whitelist in internal/channels/routing_metadata.go now preserves post_id and display_name across the inbound → outbound hop so the private-reply template variables survive the agent pipeline round-trip.

Fixed

  • Skill grant tenant isolation. Agent skill grants now validate both the skill and agent tenant scope before insert, revoke, grant listing, or can-manage checks. Visibility auto-promote/auto-demote updates are scoped to the calling tenant or system skills so one tenant cannot mutate another tenant's skill.

  • Agent provider switching. Saving an agent after changing provider/model now handles cleared ChatGPT OAuth routing config without writing SQL NULL into NOT NULL JSON config columns.

Project Status

Implemented & Tested in Production

  • Agent management & configuration — Create, update, delete agents via API and web dashboard. Agent types (open / predefined), agent routing, and lazy resolution all tested.
  • Telegram channel — Full integration tested: message handling, streaming responses, rich formatting (HTML, tables, code blocks), reactions, media, chunked long messages.
  • Seed data & bootstrapping — Auto-onboard, DB seeding, migration pipeline tested end-to-end.
  • User-scope & content files — Per-user context files (user_context_files), agent-level context files (agent_context_files), virtual FS interceptors, per-user seeding (SeedUserFiles), and user-agent profile tracking all implemented and tested.
  • Core built-in tools — File system tools (read_file, write_file, edit_file, list_files, search, glob), shell execution (exec), web tools (web_search, web_fetch), and session management tools tested in real agent loops.
  • Memory system — Long-term memory with pgvector hybrid search (FTS + vector) implemented and tested with real conversations.
  • Agent loop — Think-act-observe cycle, tool use, session history, auto-summarization, and subagent spawning tested in production.
  • WebSocket RPC protocol (v3) — Connect handshake, chat streaming, event push all tested with web dashboard and integration tests.
  • Store layer (PostgreSQL) — All PG stores (sessions, agents, providers, skills, cron, pairing, tracing, memory, teams) implemented and running.
  • Browser automation — Rod/CDP integration for headless Chrome, tested in production agent workflows.
  • Lane-based scheduler — Main/subagent/team/cron lane isolation with concurrent execution tested. Group chats support up to 3 concurrent agent runs per session with adaptive throttle and deferred session writes for history isolation.
  • Security hardening — Rate limiting, prompt injection detection, CORS, shell deny patterns, SSRF protection, credential scrubbing all implemented and verified.
  • Web dashboard — Channel management, agent management, pairing approval, traces & spans viewer, skills, MCP, cron, sessions, teams, and config pages all implemented and working.
  • Prompt caching — Anthropic (explicit cache_control), OpenAI/MiniMax/OpenRouter (automatic). Cache metrics tracked in trace spans and displayed in web dashboard.
  • Agent delegation — Inter-agent task delegation with permission links, sync/async modes, per-user restrictions, concurrency limits, and hybrid agent search. Tested in production.
  • Agent teams — Team creation with lead/member roles, shared task board (create, claim, complete, search, blocked_by dependencies), team mailbox (send, broadcast, read). Tested in production.
  • Evaluate loop — Generator-evaluator feedback cycles with configurable max rounds and pass criteria. Tested in production.
  • Delegation history — Queryable audit trail of inter-agent delegations. Tested in production.
  • Skill system — BM25 search, ZIP upload, SKILL.md parsing, and embedding hybrid search. Tested in production.
  • MCP integration — stdio, SSE, and streamable-http transports with per-agent/per-user grants. Tested in production.
  • Cron scheduling — at, every, and cron expression scheduling. Tested in production.
  • Docker sandbox — Isolated code execution in containers. Tested in production.
  • Text-to-Speech — OpenAI, ElevenLabs, Edge, MiniMax providers. Tested in production.
  • HTTP API — /v1/chat/completions, /v1/agents, /v1/skills, etc. Tested in production. Interactive Swagger UI at /docs.
  • API key management — Multi-key auth with RBAC scopes, SHA-256 hashed storage, show-once pattern, optional expiry, revocation. HTTP + WebSocket CRUD. Web UI for management.
  • Hooks system — Event-driven hooks with command evaluators (shell exit code) and agent evaluators (delegate to reviewer). Blocking gates with auto-retry and recursion-safe evaluation.
  • Media tools — create_image (DashScope, MiniMax), create_audio (OpenAI, ElevenLabs, MiniMax, Suno), create_video (MiniMax, Veo), read_document (Gemini File API), read_image, read_audio, read_video. Persistent media storage with lazy-loaded MediaRef.
  • Additional provider modes — Claude CLI (Anthropic via stdio + MCP bridge), Codex (OpenAI gpt-5.3-codex via OAuth).
  • Google Cloud Vertex AI provider — Enterprise GCP integration via Vertex OpenAI-compatible endpoint. OAuth2 service account auth (inline JSON or file path) with automatic token refresh, plus Application Default Credentials (ADC) for GKE/Cloud Run/Compute Engine. Regional endpoints for data residency (e.g. asia-southeast1, us-central1). Addresses #576.
  • Knowledge graph — LLM-powered entity extraction, graph traversal, force-directed visualization, and knowledge_graph_search agent tool.
  • Memory management — Admin dashboard for memory documents (CRUD, semantic search, chunk/embedding details, bulk re-indexing).
  • Persistent pending messages — Channel messages persisted to PostgreSQL with auto-compaction (LLM summarization) and monitoring dashboard.
  • Heartbeat system — Periodic agent check-ins via HEARTBEAT.md checklists with suppress-on-OK, active hours, retry logic, and channel delivery.

Implemented but Not Fully Tested

  • Slack — Channel integration implemented, not yet validated with real users.
  • Other messaging channels — Discord, Zalo OA, Zalo Personal, Feishu/Lark, WhatsApp channel adapters are implemented but have not been tested end-to-end in production. Only Telegram has been validated with real users.
  • OpenTelemetry export — OTLP gRPC/HTTP exporter implemented (build-tag gated). In-app tracing works; external OTel export not validated in production.
  • Tailscale integration — tsnet listener implemented (build-tag gated). Not tested in a real deployment.
  • Redis cache — Optional distributed cache backend (build-tag gated). Not tested in production.
  • Browser pairing — Pairing code flow implemented with CLI and web UI approval. Basic flow tested but not validated at scale.
⚡