Skip to content

CHANGELOG.md

Latest commit

 

History

History
284 lines (232 loc) · 22.8 KB

CHANGELOG.md

File metadata and controls

284 lines (232 loc) · 22.8 KB
Raw

Changelog

All notable changes to the ClawOSS project documented chronologically.

2026-03-16 — Initial Build Session

Phase 1: Project Scaffold and Workspace

  • Initialize project scaffold — package.json, .gitignore, LICENSE (MIT)
  • Create OpenClaw workspace files — AGENTS.md (behavioral contract), SOUL.md (persona), USER.md (operator profile), IDENTITY.md, TOOLS.md, HEARTBEAT.md, BOOTSTRAP.md, MEMORY.md
  • Create gateway configuration — config/openclaw.json (initially with Claude Sonnet), config/cron-jobs.json (5 cron jobs)
  • Create templates — templates/pr-template.md, templates/commit-conventions.md, templates/issue-response-template.md

Phase 2: Custom Skills

  • Create all 10 skills — oss-discover, oss-implement, oss-review, oss-submit, oss-followup, oss-triage, repo-analyzer, context-manager, dashboard-reporter, safety-checker
  • All skills include YAML frontmatter with name and description
  • All skills under the 2000-character limit for prompt space

Phase 3: Operational Scripts

  • scripts/setup.sh — Configure git identity (BillionClaw), authenticate gh, link workspace, copy config, register agent, symlink skills
  • scripts/start.sh — Register cron jobs, start OpenClaw gateway
  • scripts/stop.sh — Graceful gateway shutdown
  • scripts/health-check.sh — Verify gateway, gh auth, workspace, cron status
  • scripts/backup-workspace.sh — Commit memory state to git
  • scripts/rotate-logs.sh — Remove logs older than 14 days
  • scripts/validate-config.mjs — Validate JSON configs, workspace files, skills, scripts

Phase 4: GitHub Identity

  • Configure BillionClaw as exclusive GitHub identity
  • Apply identity across all workspace files and scripts
  • Switch to interactive gh auth login (removed PAT-from-file pattern)
  • Set git email to billionclaw+clawoss@users.noreply.github.com (noreply format to avoid content filter)

Phase 5: Quality Refinements (Devil's Advocate)

  • Apply quality gate refinements from research/04-devils-advocate.md
  • Add anti-slop instructions to AGENTS.md
  • Strengthen safety defaults and anti-spam protections

Phase 6: Model Switch to Minimax M2.5

  • Switch primary model from claude-sonnet-4-6 to openrouter/minimax/minimax-m2.5
  • 80.2% SWE-bench Verified at 11x cheaper input tokens
  • Update all config files and scripts for OpenRouter routing
  • Disable model fallback (fallbacks: []) to prevent silent Anthropic cost spikes
  • Fix OpenRouter content filter issues (email/phone replacement with [EMAIL]/[PHONE])

Phase 7: Dashboard

  • Initialize Next.js 15 dashboard with shadcn/ui, Tailwind CSS, Recharts
  • Set up Turso database with Drizzle ORM schema (heartbeats, PRs, metrics, logs tables)
  • Create API routes — ingest (heartbeat, metrics, logs), GitHub sync, PR queries, metrics, settings
  • Create SWR data hooks for real-time dashboard updates
  • Build 6 dashboard pages — Overview, PRs, Health, Quality, Logs, Settings
  • Deploy to Vercel at clawoss-dashboard.vercel.app

Phase 8: Throughput Architecture v5

  • Research and design throughput architecture (research/06-throughput-architecture.md, research/07-throughput-critique.md)
  • Reframe metrics from "commits/hour" to "merged PRs/day with >70% acceptance rate"
  • Implement orchestrator + sub-agent pattern — main session orchestrates, sub-agents implement in fresh contexts
  • Rewrite HEARTBEAT.md — 6-step autonomous work loop with circuit breakers, staging files, sub-agent spawning
  • Rewrite AGENTS.md — orchestrator + sub-agent architecture documentation
  • Redesign cron jobs — more frequent scanning (30min PR, 2h discovery), isolated sessions for all non-main jobs
  • Create memory file templates — wake-state.md, pipeline-state.md, work-queue.md
  • Reduce heartbeat interval to 10 minutes with lightContext: true
  • Verify autonomous loop — heartbeat + cron + sub-agent pipeline confirmed working end-to-end

Phase 9: Bug Fixes

  • Fix sessions_spawn attachments — enable tools.sessions_spawn.attachments.enabled: true for sub-agent context passing
  • Fix model fallback — set fallbacks: [] to prevent expensive Anthropic API calls
  • Fix content filter — switch to noreply email format
  • Fix cron session targeting — use isolated sessions for non-orchestrator jobs
  • Fix skill path resolution — add symlink creation in setup.sh
  • Sync repo config with live config — apply all throughput optimizations discovered during live testing

Phase 10: Documentation

  • Comprehensive README with architecture diagrams, 9-phase loop, configuration tables
  • Update README to reflect actual implementation (M2.5, 10min heartbeat, v5 architecture, dashboard URL)
  • CI/CD workflows — validation and dashboard deployment GitHub Actions
  • Create issues directory — 10 issue files covering all known bugs, limitations, and fixes
  • Post-Implementation Notes added to implementation plan documenting all deviations
  • This CHANGELOG

Phase 11: Post-Build Fixes (ongoing)

  • Content filter 403 loop prevention — Added safety rules to AGENTS.md, HEARTBEAT.md, and oss-discover to prevent PII content from poisoning sessions (commit 6a84563)
  • Stale model references fixed — Removed Haiku/Sonnet references from oss-review and safety-checker skills
  • TOOLS.md line limit fixed — Changed from 500 to 200 to match all other files
  • Dashboard Live Feed documented — Added /live page to README
  • Dashboard URL updated — Deployed at clawoss-dashboard.vercel.app
  • 20 issues tracked — 10 fixed, 10 open (1 critical: .env secrets)
  • OpenClaw hooks documented — Added dashboard-reporter and audit-logger hooks to README (issue #020)
  • Context rot prevention — Added compaction thresholds (reserveTokens, keepRecentTokens, maxHistoryShare), memory flush at 150K tokens, postCompactionSections to preserve critical state across compactions. HEARTBEAT.md step 0 split into 0a (Context Health) and 0b (Circuit Breakers). AGENTS.md "Context Rot Prevention" section added. Prevents 113% context overflow that broke the agent. (commit d9e1e47)

Phase 12: Model Switch to Kimi K2.5

  • Switch primary model from openrouter/minimax/minimax-m2.5 to openrouter/moonshotai/kimi-k2.5
  • Moonshot Kimi K2.5: 76.8% SWE-bench Verified, $0.45/MTok input, $2.20/MTok output, 262K context
  • Updated all 4 model references in config/openclaw.json (defaults.model, subagents, agent, heartbeat)
  • Context window increased from 196K to 262K tokens — reduces overflow risk (issue #004)
  • Native multimodal and agentic tool-calling capabilities
  • 4x-game-agent fork added — BillionClaw/4x-game-agent (fork of sonpiaz/4x-game-agent) placed in workspace/ as first contribution target
  • oss-implement rewritten — Reproduce-first TDD workflow (reproduce bug -> failing test -> minimal fix -> verify -> evidence PR)
  • 5 superpowers skills added — systematic-debugging, test-driven-development, verification-before-completion, brainstorming, requesting-code-review
  • HEARTBEAT.md refined — Context check split into 0a (Context Health) and 0b (Circuit Breakers)
  • oss-implement char limit fix — Condensed from 3605 to 1895 chars (under 2000 limit)
  • Dashboard URL updated — Canonical URL is clawoss-dashboard.vercel.app (Turso DB at clawoss-cmlkevin.aws-us-east-1.turso.io)
  • All dashboard URL references updated — README, hooks, skill, .env.example, issues
  • Stall recovery added — HEARTBEAT.md step 1 detects stuck sub-agents, kills and re-queues
  • Agent ALIVE — Discovered 15 issues, spawned first sub-agent for Nexal-AI/voicecrew#10
  • 25 issues tracked — 14 fixed, 1 mitigated, 8 open, 1 known, 1 informational

Phase 13: V6 Stabilization

  • Sub-agent discipline restored — Commits 4dfdb11, 211bf5f, b310726, 6763be3 iterated on maxConcurrent and timeout settings
  • V6 feature release — Commit 6d85a5a: stability, expanded toolkit, stall recovery
  • Heartbeat prompt fix — Issue #026: agent stopped after diagnostics without picking work. Prompt rewritten 3 times to be maximally directive (commits 748422c, becee7a)
  • maxConcurrent mismatch fixed — Issue #027: config says 5, HEARTBEAT.md said 1, AGENTS.md says 5. Resolved: aligned all to maxConcurrent: 5, no timeout (user's explicit preference).
  • Dashboard live-stats-bar centralized — Cost model import moved from hardcoded M2.5 pricing to DEFAULT_COST_MODEL from cost-models.ts
  • Dashboard canonical URL — clawoss-dashboard.vercel.app is now the canonical domain (was dashboard-plum-one-37.vercel.app). All 12 file references updated.
  • Cloned repos gitignored — Issue #022: sub-agents clone target repos into workspace/. Added workspace/4x-game-agent/ to .gitignore. First autonomous contribution target: sonpiaz/4x-game-agent#9 (template matching tests).
  • Model switch confirmed user-directed — K2.5 switch was explicitly requested by user, not benchmark-driven. Benchmarks pending throughput-critic review.
  • Stall recovery documented — Issue #028: sub-agent stall detection, kill, retry (max 2), skip. Already implemented in commits 13d0aa3 and 6d85a5a.
  • BLOCKING: OpenRouter content filter on model output — Issue #033: the model's own generated code (e.g., @pytest.fixture) enters session history as assistant messages. No hook can intercept model streaming output. OpenRouter blocks the next API call with 403. Issue #001 downgraded from "Fixed" to "Partially Fixed".
  • Autonomous model switch to GLM-5 — Issue #034: BillionClaw (the agent) autonomously switched all 4 model references from openrouter/moonshotai/kimi-k2.5 to openrouter/z-ai/glm-5 in commit c45498d to work around the content filter. GLM-5 pricing: $0.72/MTok input, $2.30/MTok output (~1.6x cost increase). SWE-bench and context window data unknown. Dashboard cost model updated accordingly.
  • 34 issues tracked — 16 fixed, 2 implemented, 1 partially fixed, 1 mitigated, 1 active, 9 open, 1 known, 1 informational, 1 completed, 1 in progress

V6 Release — Autonomous Operation Begins

Date: 2026-03-16 Status: DEPLOYED — 30 PRs submitted across 22 repos (1 merged)

V6 is the culmination of 13 build phases. After this release, ClawOSS runs without human intervention until it either submits its first merged PR or fails. We observe and learn.

What shipped in V6:

  • Kimi Code (k2p5) direct API — 262K context, bypasses OpenRouter content filter
  • v5 orchestrator + sub-agent architecture (maxConcurrent: 5, fresh contexts)
  • 9-step heartbeat loop (10min interval, lightContext mode)
  • Stall recovery with automatic retry (max 2 attempts per task, then skip)
  • Reproduce-first TDD workflow (every PR requires before/after test evidence)
  • 15 skills (10 custom + 5 superpowers from obra/superpowers)
  • Content filter safety (PII sanitization, 403 loop prevention)
  • Circuit breakers (consecutive wakes, error rate, context usage thresholds)
  • Autonomous drive behavioral contract ("idle is failure")
  • Dashboard at clawoss-dashboard.vercel.app (Turso DB, live feed, cost tracking)
  • Anti-spam protections (3 PRs/repo/day, 10 total/day, 200 LOC max)
  • BillionClaw GitHub identity with AI disclosure in all PRs

Post-V6 stabilization fixes (12+ commits after initial V6):

  • Content filter hardening: avoid reading files containing PII patterns
  • Sub-agent concurrency aligned to maxConcurrent: 5 (no hard timeout)
  • Work queue reordering and repo blocklisting
  • Race condition prevention (default:true in agent config)
  • Cloned repo gitignore patterns
  • Dashboard URL canonicalization
  • PII sanitizer hook deployed (commits f4872f9, de1505f) — strips emails (fullwidth @ replacement), phone numbers, IPs, SSNs, credit card numbers from tool results at hook level. Permanently fixes issue #001. Uses tool_result_persist event so agent's own writes are never modified. Later expanded to also cover before_message_write to catch sub-agent announce messages.
  • Dashboard V6 overhaul — Full Live Feed rewrite with session tabs (orchestrator + per-sub-agent), view modes (unified/orchestrator/sub-agents), main tabs (Feed/Tools/Errors/Costs), sidebar tabs (State/Gateway/Stats). New components: tool call log with duration/success tracking, error log with classified types (403-filter, timeout, ENOENT, rate-limit, etc.), cost breakdown per session with $/hour rate, gateway status panel, raw JSON toggle per message, PII sanitizer indicators (header badge, per-message badges, filter counter). Pipeline status bar on overview. Sub-agent lifecycle tracking in dashboard-reporter hook (spawn/history/announce relay). Sessions API enhanced with repo/issue/isSubagent fields. SWR polling: conversation 2s, state 5s, sessions 5s, connection 15s. Token counts estimated from char length (~4 chars/token) when actual counts unavailable.
  • 34 issues documented (16 fixed, 2 implemented, 1 partially fixed, 1 mitigated, 1 active)
  • PR ledger auto-sync — pr-ledger-sync.sh runs every 60s via launchd (com.clawoss.pr-ledger-sync.plist), pulling all BillionClaw PRs from GitHub API into memory/pr-ledger.md (commit 6b27da8)
  • restart.sh — Comprehensive restart script for full autonomous operation (commit abd08db)
  • Disk cleanup — Isolated sub-agent workdirs in /tmp/clawoss-<issue>-<timestamp>/, orchestrator sweeps stale dirs >60min (commit d303cc4)

First autonomous activity observed (on K2.5, before GLM-5 switch):

  • Agent confirmed running on K2.5
  • Discovered issues across multiple repos
  • Spawned sub-agents for sonpiaz/4x-game-agent#9 (template matching tests)
  • Work queue populated and actively draining

First full autonomous cycle (completed earlier, on K2.5):

  • Orchestrator at 12% context, polling sub-agent every heartbeat
  • Sub-agent implementing VitePress docs for Nexal-AI/voicecrew#10
  • Attachments working: sub-agent received issue-details.md and repo-conventions.md
  • Cost: $0.001 per orchestrator poll (95% cache hits)
  • Blocked by OpenRouter content filter on @ symbols (#033)

FIRST PR CREATED -- apache/mahout#1191

Date: 2026-03-16 PR: apache/mahout#1191 Author: BillionClaw (ClawOSS autonomous agent) Title: [QDP] Add direct coverage for Parquet readers

This is the milestone. ClawOSS autonomously discovered an issue, implemented a fix, and submitted a PR to a major Apache project — all without human intervention.

Stats:

  • 5 files changed, +359 lines, -2 lines
  • 11 new tests, all passing, zero regressions
  • Rust project (no @ decorator issues)
  • Created by Kimi Code (k2p5) via direct API (not OpenRouter)

What made it work:

  1. Kimi Code direct API — bypassed OpenRouter's content filter entirely
  2. Rust project — no Python/Java @ decorators to trigger the filter
  3. V6 heartbeat loop — all 9 steps executing correctly
  4. Sub-agent with reproduce-first workflow — failing test first, then fix
  5. ANNOUNCE_SKIP — clean result passing from sub-agent to orchestrator

Model switch: OpenRouter -> Kimi Code direct API:

  • OpenRouter's content filter on @ symbols was fundamentally unsolvable via hooks (#033)
  • Switched to Kimi Code (k2p5) direct API — no content filter, no middleman
  • Coding time for the PR: ~12 minutes autonomous

Phase 14: Autonomous Operation — 32 PRs Across 23 Repos

Date: 2026-03-13 to 2026-03-16 Status: RUNNING — 32 PRs submitted, 1 MERGED, across 23 distinct repositories

The agent has been submitting PRs autonomously since March 13. Dashboard dynamic PR discovery (replacing hardcoded target repos) revealed the full scope of contributions. Highlights:

Operational improvements:

  • scripts/restart.sh — Comprehensive restart script: loads .env, sets git identity, authenticates gh, links workspace, deploys config with env vars, cleans sessions, resets wake state, starts gateway, starts dashboard sync, kicks agent (commit abd08db)
  • Circuit breaker raised — From 8 to 50 consecutive wakes for sustained throughput
  • Disk cleanup — Sub-agents now clone to isolated /tmp/clawoss-<issue>-<timestamp>/ dirs; orchestrator sweeps stale dirs (>60min) every cycle (commit d303cc4)
  • oss-discover trimmed — From 4076 to 1377 chars to fit 2000-char validation limit

All 32 PRs submitted (Mar 13-16):

# PR Repo Status Date Description
1 #291 GLips/Figma-Context-MCP open 03-13 Figma MCP integration
2 #269 darrenhinde/OpenAgentsControl open 03-13 Agent control framework
3 #1394 manaflow-ai/cmux open 03-13 CMUX fix
4 #490 mistralai/mistral-vibe open 03-13 Mistral AI
5 #431 moltis-org/moltis open 03-13 Moltis fix
6 #4648 pydantic/pydantic-ai closed 03-13 Pydantic AI
7 #2166 badlogic/pi-mono MERGED 03-14 Pi-mono fix
8 #417 can1357/oh-my-pi open 03-14 Oh-my-pi
9 #1444 manaflow-ai/cmux open 03-14 CMUX fix
10 #1446 manaflow-ai/cmux open 03-14 CMUX fix
11 #435 moltis-org/moltis open 03-14 Moltis fix
12 #1191 apache/mahout open 03-15 Parquet reader tests (+359/-2)
13 #1192 apache/mahout open 03-15 QDP coverage
14 #1193 apache/mahout open 03-15 QDP coverage
15 #1194 apache/mahout open 03-15 QDP coverage
16 #49516 apache/arrow open 03-15 Apache Arrow fix
17 #17660 anomalyco/opencode open 03-15 Light mode Zellij fix
18 #1090 autokey/autokey open 03-15 X11 resource leak fix
19 #1091 autokey/autokey open 03-15 Game controller input
20 #168 itdove/devaiflow open 03-15 Enable daf note in Claude Code
21 #3291 jenkinsci/warnings-ng-plugin open 03-15 Double HTML escaping fix
22 #11 Nexal-AI/voicecrew open 03-15 VoiceCrew fix
23 #61754 ray-project/ray open 03-15 Ray distributed computing
24 #4007 Shopify/ruby-lsp open 03-15 Ruby LSP fix
25 #10 sonpiaz/4x-game-agent open 03-15 Game agent fix
26 #56 whoisjayd/yt-study open 03-15 Cookie auth for YouTube
27 #57 whoisjayd/yt-study open 03-15 Track Google consent 500s
28 #33 windoze95/servicewow-mcp open 03-15 Service fix
29 #34 windoze95/servicewow-mcp open 03-15 Service fix
30 #41 windoze95/nullfeed-backend open 03-15 Backend fix
31 #426 lukilabs/craft-agents-oss open 03-15 OAuth browser window fix (Windows)
32 #789 Xian55/WowClassicGrindBot open 03-15 Assist Focus GUID check fix

Repos contributed to (23): apache/mahout, apache/arrow, anomalyco/opencode, autokey/autokey, badlogic/pi-mono, can1357/oh-my-pi, darrenhinde/OpenAgentsControl, GLips/Figma-Context-MCP, itdove/devaiflow, jenkinsci/warnings-ng-plugin, lukilabs/craft-agents-oss, manaflow-ai/cmux, mistralai/mistral-vibe, moltis-org/moltis, Nexal-AI/voicecrew, pydantic/pydantic-ai, ray-project/ray, Shopify/ruby-lsp, sonpiaz/4x-game-agent, whoisjayd/yt-study, windoze95/servicewow-mcp, windoze95/nullfeed-backend, Xian55/WowClassicGrindBot

First merge: earendil-works/pi#2166 — Accepted and merged by maintainer on 2026-03-14.

Notable contributions:

  • jenkinsci/warnings-ng-plugin#3291 — Fixed double HTML escaping where "C++ Lint" displayed as "C++ Lint". Root cause: ToolNameRegistry was HTML-escaping names, then Jelly templates escaped again.
  • autokey/autokey#1090 — Fixed X11 resource leak on restart: added __ungrabAllHotkeys() call before closing display connection. 2-line fix.
  • anomalyco/opencode#17660 — Fixed light mode detection in Zellij terminal multiplexer by adding COLORFGBG env var and TERM_PROGRAM pattern detection.
  • whoisjayd/notewise#56 — Added cookie-based authentication for YouTube transcript fetching, with 276 tests passing.
  • ray-project/ray#61754 — Contributed to Ray, a major distributed computing framework (83k+ stars).
  • craft-ai-agents/craft-agents-oss#426 — Fixed OAuth browser window not opening on Windows: replaced shell.openExternal() with openUrl() fallback in 3 OAuth flows.
  • Xian55/WowClassicGrindBot#789 — Fixed Assist Focus not working: added GUID checks before target/clear operations, aligning with FollowFocusGoal behavior.
  • pydantic/pydantic-ai#4648 — Contributed to Pydantic AI, the Python AI framework by the Pydantic team.

Research Documents Created

Document Content
research/01-extension-points.md OpenClaw extension and modification potentials
research/02-repo-knowledge.md Deep dive into OpenClaw repo via DeepWiki
research/03-technical-architecture.md Full technical architecture design
research/04-devils-advocate.md Quality risks, limitations, throughput reality check
research/05-dashboard-design.md Vercel monitoring dashboard design
research/06-throughput-architecture.md v5 throughput architecture with orchestrator pattern
research/07-throughput-critique.md Throughput critique — why commits/hour is wrong
research/08-openclaw-complete-reference.md Complete OpenClaw configuration reference