Follow along: I share what breaks (and what fixes it) putting AI agents into production — @surlymochan on X · Contra
⭐ If evalid or swrm saves you time, a star is the best way to support the work — I'm pushing for 100 stars this month.
AI agents · Harness engineering · Reliable loops & graphs
I build tools to coordinate coding agents and verify what they deliver.
Claimed ≠ verified. I’m building agent systems where delivery is an evidence trail, not a status update.
| Project | What it does | Status |
|---|---|---|
| Phanes | Coordinates goals, distributed execution and verified acceptance. | Consolidating |
| volto | Planned roles, memory and knowledge for collaborating agents. | Planned relaunch |
| swrm | Dispatches coding agents with leases, isolated worktrees and limits. | Available |
| evalid | Seals test-run snapshots and replays them to verify results. | Available |
On six selected hard SWE-bench Verified issues, each arm used Pi with DeepSeek-V4.1-Flash, one run per issue, and a 20-minute cap. A solve means the official tests passed: swrm 5/6 · evalid 4/6 · plain agent 3/6 · Superpowers 2/6. This is a small project-reported experiment, not a general leaderboard; raw runs and methodology are in the swrm benchmark and evalid benchmark.
Opened and merged PRs across my projects, nibiru-labs and upstream repositories, including co-authored work. Monthly totals use Singapore time.
|
|
|
Recent merged contributions:
- senpi — fenced session-less wire-status reads and added per-session fence tests.
- senpi — priced Codex turns at the sent tier when the backend echoes auto/default.
- senpi — kept the pre-signal error when shutdown recovery throws.
- senpi — covered the shutdown recovery path per call site.
Currently contributing fixes to oh-my-openagent, OpenBot, and strands-agents/harness-sdk. Open pull requests are proposals, not merged features or maintainer roles.
Contribution history & open pull requests
Merged upstream pull requests: 66 across 17 repositories.
code-yeongyu/senpi
- senpi#6343 fix(coding-agent): estimate context after a model switch instead of showing stale usage
- senpi#6333 fix(anthropic-subscription): read the prompt-cache lifetime from loaded settings files (senpi#2988)
- senpi#6323 fix: Ultrafast follow-ups from #2975
- senpi#6301 fix(goal): recover a crashed goal-continuation stop on session start
- senpi#6295 fix(coding-agent): release the anthropic-subscription pump on abort (#2974)
- senpi#6284 fix(agent-session): refused engine-turn messages follow the transcript-failure contract
- senpi#6279 fix(coding-agent): hide console window for win32-reachable spawns (#3965)
- senpi#6275 fix(coding-agent): model-scoped 403 refusals no longer block the credential slot
- senpi#6274 fix(coding-agent): stop isolated CLI re-launch replaying the caller's eval entry mode
- senpi#6255 fix(session-backends): declare the complete bun:sqlite Database in the sqlite-bun package (#5996)
- senpi#6222 fix(interactive): footer names the serving account when the pin is blocked (#5983)
- senpi#6211 fix(rpc): retry a failed host-token mint before the cached token expires (#5919)
- senpi#6264 fix(prompt-preset): render the Haiku 5.5 search-grounding sentence for Anthropic-native search
- senpi#6232 docs(senpi-codemode): qualify process.execPath as a Bun launcher — standalone Bun only
- senpi#6169 fix(mcp): fence session-less wire-status reads; add per-session fence tests for #3001
- senpi#6156 fix(ai): price Codex turns at the sent tier when the backend echoes auto/default
- senpi#6149 fix(compaction): accept valid configurationUpdate messages in retained-suffix safety (senpi#8566)
- senpi#6146 test(coding-agent): cover the shutdown recovery path per call site (senpi#3067 L5)
- senpi#6144 fix(coding-agent): keep the pre-signal error when shutdown recovery throws (senpi#3067 L4)
- senpi#5997 test(kill-guard): stop owned sleepers on process exit, shorten TTL to 60s
- senpi#5908 test(coding-agent): drive user bash through default operations in kill-guard test
- senpi#5854 fix(codemode): probe kill-guard sleeper liveness without ps
- senpi#5700 fix(ask-user): restore dangling questions on startup opens
- senpi#5616 fix(claude-account): emit accounts_changed once per account add
- senpi#5614 test(coding-agent): run source compile cache teardown even when assertions fail
- senpi#5269 fix(retry-fallback): a 503 whose body says "Billing Scope" is not an account usage limit
- senpi#5203 fix(coding-agent): weight CJK text in compaction cut-point budgets
code-yeongyu/oh-my-openagent
- oh-my-openagent#12545 fix(memory-core): assert the timeout-diagnostics contract, not the host's snapshot
- oh-my-openagent#12544 fix(omo-senpi): name the onboarding skill instead of its install path
- oh-my-openagent#12568 fix(memory-sync): refuse non-regular leaf files on landing
- oh-my-openagent#12571 test(omo-senpi): rename the buildWidgetRows resident-member case left by #12160
- oh-my-openagent#12534 fix(omo-native): report broken omo wrapper and duplicate omo-ai installs in doctor
- oh-my-openagent#12529 test(omo-senpi): require the launcher attestation on initial-namespace Linux in sandbox-engine-trust.test.ts
- oh-my-openagent#12502 test(team-e2e): use a collision-proof fake root pid in taskkill cases
- oh-my-openagent#12501 test(senpi-task): create engine agent dirs with mode 0700
- oh-my-openagent#12040 Fix #11570: tokenize SVG foreignObject in fact verify
- oh-my-openagent#11909 feat(memory): name the parked-facts cause and run log in /doctor; report retry outcomes
- oh-my-openagent#11826 fix(memory-core): settle the relaunched cache before asserting on the store dir in cache-persist tests
- oh-my-openagent#11699 fix(memory): let update_description add frontmatter to a note that has none
- oh-my-openagent#11698 fix(omo-senpi): freeze the shared models.json parse and type it deeply readonly
- oh-my-openagent#11697 fix(omo-senpi): pin competing pathspec switches off for side panel git diffs
- oh-my-openagent#11237 feat(hooks): persist Momus review verdicts next to the plan
- oh-my-openagent#11228 fix(omo-senpi): make init-deep committed mode self-contained
Apache SkyWalking
- skywalking-satellite#292 Fix routing key order in OTLP metrics v1 forwarder
- skywalking-swck#232 Bump golang.org/x/net to v0.61.0 in adapter and operator
- skywalking-mcp#45 Bump golang.org/x/text and x/sys to clear govulncheck advisories
- skywalking-go#258 Bump grpc to v1.83.2 and x/net to v0.58.0 to fix known CVEs
- skywalking-cli#236 fix(deps): bump grpc, x/crypto and x/net for known CVEs
- skywalking-rover#222 Bump golang.org/x/net to v0.61.0 to fix HTTP/2 CVEs
- skywalking-infra-e2e#149 Bump vulnerable dependencies to patched versions
- skywalking-goapi#99 Bump grpc to v1.84.0 and x/net to v0.57.0
- skywalking-satellite#289 Archive 1.3.0 changelog and start 1.4.0 section
- skywalking-rover#220 Bump up grpc to 1.84.0 to fix CVE-2026-33186
OpenSpec
- OpenSpec#2120 fix(new-change): name the configured schema in the progress line
- OpenSpec#2116 perf(cli): load ora only when a spinner is shown
- OpenSpec#2114 fix(verify): match REMOVED baselines by exact normalized name
- OpenSpec#2103 fix(security): replace fast-glob with tinyglobby to drop braces
nightly-labs/openbot
- openbot#2577 Read inline triple-backtick code as text in Live Activities
- openbot#2576 Keep emoji whole when a Live Activity cuts a code span
- openbot#2578 Report an MCP sign-in that cannot start
comet-ml/opik
- opik#9278 issue-8590 SDK fix: accept repeated identical JSON objects in TS LLM-judge parser
- opik#9271 issue-8868 SDK fix: treat blank OPIK_API_KEY as unconfigured
infiniflow/ragflow
- ragflow#21110 fix(web): accept null table_column_* fields in dataset settings form
Jenqyang/Awesome-AI-Agents
- Awesome-AI-Agents#687 Add OpenSwarm to Multi-Agent Task Solver Projects
modelcontextprotocol/go-sdk
- go-sdk#1412 mcp: parse SSE field values per the SSE spec
openclaw/openclaw
- openclaw#177871 fix(ai): preserve compatible reasoning in follow-up requests
In review (101)
- ag-ui-protocol/ag-ui#3362 fix(adk-middleware): key text-less tool-call history messages by the first tool call id (#3254)
- ag-ui-protocol/ag-ui#3349 fix(adk-middleware): emit backend tool calls arriving with an LRO call (#3300)
- ag-ui-protocol/ag-ui#3341 fix(langgraph): translate CopilotKit emit events to AG-UI events (#1364)
- ag-ui-protocol/ag-ui#3340 fix(kotlin): publish deep state without recursive StateFlow equality
- ag-ui-protocol/ag-ui#3339 docs(langgraph): fix stale dojo run instructions
- ag-ui-protocol/ag-ui#3336 fix(dojo): make mobile chat drag-to-close reachable
- ag-ui-protocol/ag-ui#3335 issue-892 fix(langgraph): pass frontend tools to the model in create_agent() graphs
- ag-ui-protocol/ag-ui#3333 issue-3311 fix(aws-strands): close inline inner tool calls left open by a failing wrapper
- ag-ui-protocol/ag-ui#3330 issue-3213 docs: update CLAUDE.md Python SDK commands from poetry to uv
- ag-ui-protocol/ag-ui#3326 fix(adk-middleware): keep valid inline data when BinaryInputContent also carries id/url
- ag-ui-protocol/ag-ui#3325 issue-3077 adk-middleware fix: completed runs awaiting client tool results do not occupy concurrency slots
- ag-ui-protocol/ag-ui#3284 fix(adk-middleware): resume HITL confirmation in the sub-agent that asked for it (#3262)
- ag-ui-protocol/ag-ui#3265 fix(a2a): send agent.headers on A2A requests
- ag-ui-protocol/ag-ui#3258 fix(adk-middleware): filter rewound events in adk_events_to_messages
- ag-ui-protocol/ag-ui#3241 fix(client): give interrupts without expiresAt a bounded default expiry
- ag-ui-protocol/ag-ui#3235 fix(mastra): stream frontend tool answers under the stored message's continuation id
- ai-boost/awesome-harness-engineering#345 Add openflow to Verification & CI Integration
- ai-for-developers/awesome-ai-coding-tools#1006 Add OpenFlow to Testing and QA
- anomalyco/opencode#59369 feat(tui): add an opt-out for the Kitty keyboard protocol
- anomalyco/opencode#57237 fix(cli): expose retryable ACP plugin activation readiness errors
- anomalyco/opencode#57236 fix(core): retry truncated compaction with less history
- ARUNAGIRINATHAN-K/awesome-ai-agents-2026#494 Add OpenFlow to Coding Agents
- benchflow-ai/awesome-evals#134 Add openflow to eval frameworks and harnesses
- bradAGI/awesome-cli-coding-agents#516 Add openflow to Orchestrators & autonomous loops
- browser-use/browser-use#6132 fix(llm): keep Field descriptions when flattening $ref in optimized schemas
- caramaschiHG/awesome-ai-agents-2026#692 Add Openflow to Task and Workflow Agents (Automation)
- code-yeongyu/oh-my-openagent#13115 fix(senpi-task): name the designated models when the host cannot hold the chain
- code-yeongyu/oh-my-openagent#13088 fix(senpi-task): designated-chain follow-ups from #12761 review (#12778)
- code-yeongyu/oh-my-openagent#13075 fix(thread): send remote open_session on a short-lived connection
- code-yeongyu/oh-my-openagent#12802 fix(telemetry-core): compare spool file identity as bigints so a Windows swap is not read as the same file (#12462)
- code-yeongyu/oh-my-openagent#12793 fix(memory-core): pin core.hooksPath so global config cannot skip memory hooks
- code-yeongyu/oh-my-openagent#12791 fix(omo-senpi): keep the generated installer at mode 0644 and fail the freshness gate on mode drift
- code-yeongyu/oh-my-openagent#12742 fix(memory-core): persist the resolved NFD spelling in the render record (#12158 item 2)
- code-yeongyu/oh-my-openagent#12706 fix(omo-senpi): escalate the device-service bootstrap worker to SIGKILL after the deadline (#12267)
- code-yeongyu/oh-my-openagent#12623 fix(no-raw-tls): list the omo-native omowright runtime in KNOWN_BUILT_RUNTIMES (#12242)
- code-yeongyu/oh-my-openagent#12616 fix(utils): logger no longer captures globalThis timers at module load; close #12218 follow-ups
- code-yeongyu/oh-my-openagent#12604 fix(senpi-task): heal rpc key dir before task host ensure (#12194)
- code-yeongyu/oh-my-openagent#12531 fix(memory-core): pin the ps-lstart fallback to UTC so a live owner cannot read as dead under a different TZ
- code-yeongyu/oh-my-openagent#12111 fix(memory): treat an ESRCH-dead pid without a recorded start identity as dead
- code-yeongyu/oh-my-openagent#12070 fix(senpi-desktop-service): map unknown interventionWatch values to the unknown sentinel (#11731)
- code-yeongyu/oh-my-openagent#12069 fix(senpi-task): guard timer unref in host heartbeat teardown (#11714)
- code-yeongyu/oh-my-openagent#12041 fix(lsp): honor lsp bindings from oh-my-opencode.jsonc fallback configs
- code-yeongyu/oh-my-openagent#11745 fix(omo-opencode): session_list date filter uses metadata time.updated instead of reading every session (#10796)
- code-yeongyu/oh-my-openagent#11088 feat(mcp): add Volcano Engine as a websearch provider
- code-yeongyu/oh-my-openagent#11087 fix(omo-senpi): cover gateway-rules-extension in build-extension --check (#8263)
- code-yeongyu/oh-my-openagent#11085 fix(omo-senpi): run the init-deep advisor on the first real prompt, not session_start
- code-yeongyu/oh-my-openagent#11077 fix(omo-native): warn on broken catalog cache for proxy blocks and name keyed ids missing from the catalog
- code-yeongyu/oh-my-openagent#11075 fix(installer-mirror): retry transient GitHub 5xx asset downloads with backoff
- code-yeongyu/oh-my-openagent#10955 fix(config-migration): report legacy keys that config migrate does not import
- code-yeongyu/senpi#6330 fix(coding-agent): account stickiness follow-ups for anthropic subscription
- code-yeongyu/senpi#6320 fix(coding-agent): log rotation follow-ups from #2977
- code-yeongyu/senpi#6283 fix(ask-user): gate a partial answer to a multi-question required call
- code-yeongyu/senpi#6217 fix(rpc): refuse move_session copy ids the parser let through (#5930)
- code-yeongyu/senpi#6205 fix(rpc): clear the close grace timer when dispose throws synchronously; give router registry fakes a closeGraceMs (#5929)
- code-yeongyu/senpi#6155 fix(ai): renew Cursor logins through the IDE refresh grant (#2919)
- code-yeongyu/senpi#6129 fix(agent): skip a session file that vanishes before its header is read; fail listDir when its directory is removed mid-call
- code-yeongyu/senpi#6122 fix(rpc): verify_api_key takes a region so a Kimi Code key is checked on its regional host
- code-yeongyu/senpi#6118 fix(coding-agent): read process start times in UTC so DST fall-back hours keep live holders (senpi#3067 L1)
- code-yeongyu/senpi#6103 fix(fmt): resolve bun without a shell so spaced paths stay one argument; fix stale releasability comment (#5599)
- code-yeongyu/senpi#5914 test(coding-agent): skip grep search step when no grep engine is available
- code-yeongyu/senpi#5517 fix(hooks): skip the trust-state lock when the state file is absent
- comet-ml/opik#9498 issue-7848 SDK fix: fail loud on conflicting JSON objects in LLM-judge parsing
- comet-ml/opik#9320 issue-9069 SDK fix: keep function calls and all text in streamed Gemini traces
- comet-ml/opik#9270 NA SDK fix: aggregate Cohere streamed tool plan and tool calls
- crewAIInc/crewAI#8258 fix(schema): keep sibling keys of $ref nodes in resolve_refs
- cursor/plugins#1149 fix(pstack): pin poteto-mode script devDependencies
- cursor/plugins#1148 fix(pstack): satisfy strict typecheck in statechart run.ts and CLI tests
- e2b-dev/awesome-ai-agents#1855 Add openswarm
- filipecalegario/awesome-vibe-coding#392 Add OpenFlow
- huggingface/smolagents#3020 Only rewrite real final_answer assignments in fix_final_answer_code
- jamesmurdza/awesome-ai-devtools#1256 Add OpenFlow to Configuration & Context Management
- Jenqyang/Awesome-AI-Agents#688 Add OpenFlow to Tools
- mahseema/awesome-ai-tools#2543 Add Openflow to Code
- modelcontextprotocol/typescript-sdk#3050 fix(core): expand RFC 6570 list and multi-variable expressions correctly
- nightly-labs/openbot#4383 test(auth-api): refuse two account migrations with the same number
- nightly-labs/openbot#4382 fix(dev): run the pre-commit checks when OpenBot is in a subdirectory
- nightly-labs/openbot#4381 fix(dev): bind the dev stack to loopback unless --lan is passed
- nightly-labs/openbot#3786 fix(backend): re-read a spent usage reading before its reset
- nightly-labs/openbot#3686 Fail the Gemini turn when Google answers with an account verification prompt
- NipunaRanasinghe/awesome-ai-agents#257 Add OpenSwarm to Core Frameworks
- NousResearch/hermes-agent#141789 fix(agent): retry disabled Anthropic fast mode at standard speed
- NousResearch/hermes-agent#141743 fix(desktop): resolve remote previews without local filesystem state
- NousResearch/hermes-agent#141739 fix(sessions): omit hidden context from prompt-only exports
- NousResearch/hermes-agent#139572 fix(doctor): fail when gateway process or A2A port is down
- NousResearch/hermes-agent#135709 fix(dingtalk): ship python-socks so the stream websocket can use a SOCKS proxy
- obra/superpowers#2548 docs(skills): require RED output for batched TDD and fix-pass ledger
- obra/superpowers#2543 fix(task-brief): diagnose translated Task headings
- obra/superpowers#2538 fix(lint-shell): print full header comment for --help
- openclaw/openclaw#177754 fix(memory): report unexpected loss of remote file watching
- openclaw/openclaw#176558 fix(irc): SecretRef IRC passwords fail validation and service-only NickServ overrides skip the inherited ref
- openclaw/openclaw#175933 docs(macos): explain the Bun local-network and Git prompts on first launch
- RoggeOhta/awesome-codex-cli#487 Add openflow to Cross-Agent Tools
- rohitg00/awesome-claude-code-toolkit#874 Add openflow to the Ecosystem table
- run-llama/llama_index#24046 fix(core): pass callback_manager to response synthesizer in RetrieverQueryEngine.from_args
- run-llama/llama_index#24007 fix(core): return no nodes from SimplePropertyGraphStore.get for empty ids
- run-llama/llama_index#23992 fix(core): let as_retriever kwargs override index defaults
- run-llama/llama_index#23603 fix(core): use matching embeddings in EmbeddingRecencyPostprocessor
- slavakurilyak/awesome-ai-agents#875 Add OpenFlow (executable delivery workflows for AI coding agents)
- strands-agents/harness-sdk#5622 fix(tools): warn when a tool parameter named agent has a non-Agent annotation
- strands-agents/harness-sdk#5504 fix(models): port strict schema transform fixes from TS
- strands-agents/harness-sdk#5463 fix(cli): mirror the 2026-10 model-catalog context windows in FALLBACK_CONTEXT_WINDOWS
Generated from the GitHub search API by a daily workflow.
I publish benchmark methods and run records alongside my tools. Results are specific to the tasks, models, and configurations tested; small samples need careful interpretation.
More tools will appear here as their installation, documentation, and examples are ready for others to use.