AgentM

Structured procedure to locate the root cause of a confirmed bug. Uses keyword search → focused reading → call-graph tracing. Outputs a specific file + line + hypothesis.

Step-by-step procedure to reproduce a bug. Use this FIRST before trying to fix anything. Produces a minimal script that demonstrates the failure.

AgentM SDK development guide — manifest-as-agent-unit philosophy, SDK programmatic invocation, dynamic workflow orchestration, atom contract, Operations abstraction, event system, service communication, CLI conventions, scenario authoring, logging, structured output, and config resolution. Use whenever writing, editing, or reviewing code under src/agentm/ (atoms, core, gateway), contrib/scenarios/ (manifests), or contrib/extensions/ (workspace-member atoms). Also trigger when creating new atoms, modifying MANIFEST declarations, registering tools or events, touching FileOperations / BashOperations / ResourceWriter, writing CLI subcommands, configuring model profiles, spawning child sessions, building multi-agent orchestrators, writing workflow scripts, or when a code change looks like it might bypass the SDK's existing abstractions. If you catch yourself about to write raw os.stat / open() / subprocess.run in an atom, importing openai/anthropic directly instead of going through the provider layer, or shelling

Guide for adding new benchmarks to the agentm-eval framework and running experiments through the unified `agentm eval` CLI. Covers the adapter interface, experiment lifecycle, result schema, ClickHouse linking, and output conventions. Use whenever creating a new eval adapter, modifying an existing benchmark integration, writing code under contrib/evals/src/agentm_eval/, discussing experiment management, or when the user mentions "add a benchmark", "new eval", "接入评估", "新benchmark", "实验管理", "eval adapter". Also trigger when reviewing eval-related code or debugging experiment output structure.

How to write workflow scripts that orchestrate multiple agents in AgentM. Covers agent(), parallel(), pipeline(), phase(), return values, atom_config data passing, and the journal-based resume / invalidation / partial re-run mechanism (workflow_lineage, workflow_invalidate). Trigger when writing a workflow script, choosing an orchestration pattern, recovering from a wrong node result, or when the user says 编排 / workflow / 多 agent / 并行 / 扇出 / 调度 / 写个 workflow / 失效 / 部分重跑 / 重算. Also trigger when you catch yourself building prompts inside a workflow script — pass data via atom_config instead and let the agent build its own prompt.

Analyze agentm session traces using `agentm trace` atomic commands. Use when inspecting trajectories, token economics, tool calls, or aggregating stats across a trace tree (parent + child sessions). Compose atomic commands via shell pipes — never parse OTLP JSONL directly.

Develop, test, and validate Terminal User Interface applications using tmux as your "headless browser" for terminals. Use when asked to run a TUI application, capture its output, verify its behavior, or debug rendering issues. This skill provides a structured workflow for TUI development, including session management, interaction patterns, state capture, and regression detection.

Develop, test, and validate Terminal User Interface applications using tmux as your "headless browser" for terminals.

AgentM deployment topology, configuration, and operations guide — deployment modes (standalone, gateway+feishu, multi-bot), config.toml structure, per-channel workspace routing, persona management, observability, and troubleshooting. Use when the user asks about deployment, adds/removes/modifies bots, sets up workspace isolation, debugs a silent/slow/misbehaving bot, or manages bot personas. Also trigger when diagnosing which deployment mode is active.

AgentM deployment topology, configuration, and operations guide — deployment modes (standalone, gateway+feishu, multi-bot), config.toml structure, per-channel workspace routing, persona management, observability, and troubleshooting. Use when the user asks about deployment, adds/removes/modifies bots, sets up workspace isolation, debugs a silent/slow/misbehaving bot, or manages bot personas. Also trigger when diagnosing which deployment mode is active.

Analyze agentm session traces using `agentm trace` atomic commands. Use when inspecting trajectories, token economics, tool calls, or aggregating stats across a trace tree (parent + child sessions). Compose atomic commands via shell pipes — never parse OTLP JSONL directly.

Inspect your own running session to debug yourself — see the terminal UI the user sees (tui_snapshot), read your own live trace, and mine historical traces. Use when the user reports a bug in how you behaved, a rendering/interaction glitch in the terminal client, or asks you to figure out what went wrong in this or a past session.

Review agent trajectory and harness logs to find behavioral anti-patterns, then produce actionable improvements (prompt edits, scenario config, atom additions). Use when asked to analyze why an agent underperformed, find patterns across multiple runs, audit tool usage efficiency, or generate improvement recommendations. Triggers on: 提升, 改进, anti-pattern, 自我检查, 为什么表现差, 优化 agent, review trajectory, improve agent, analyze runs, what went wrong across runs.

AgentM SDK development guide — manifest-as-agent-unit philosophy, SDK programmatic invocation, dynamic workflow orchestration, atom contract, Operations abstraction, event system, service communication, CLI conventions, scenario authoring, logging, structured output, and config resolution. Use whenever writing, editing, or reviewing code under src/agentm/ (atoms, core, gateway), contrib/scenarios/ (manifests), or contrib/extensions/ (workspace-member atoms). Also trigger when creating new atoms, modifying MANIFEST declarations, registering tools or events, touching FileOperations / BashOperations / ResourceWriter, writing CLI subcommands, configuring model profiles, spawning child sessions, building multi-agent orchestrators, writing workflow scripts, or when a code change looks like it might bypass the SDK's existing abstractions. If you catch yourself about to write raw os.stat / open() / subprocess.run in an atom, importing openai/anthropic directly instead of going through the provider layer, or shelling

Debug the AgentM single-process gateway by connecting a synthetic wire client and dumping the outbound frames it emits. Use when a chat client (terminal/feishu) "doesn't see" something — a slash command, a tool, status, a reply — or when diagnosing wire/session_ready/command-catalog issues. Shows exactly what the gateway puts on the wire, bypassing the TUI.

Review AgentM code for boundary isolation, pluggability-axis violations, atom contract breaches, and design-pattern intrusion. Use when asked to run a boundary review over a path, git range, current branch, or staged changes; also trigger on "review-boundary", "boundary reviewer", "boundary audit", "pluggability review", or "§11 contract review".

Review agent trajectory and harness logs to find behavioral anti-patterns, then produce actionable improvements (prompt edits, scenario config, atom additions). Use when asked to analyze why an agent underperformed, find patterns across multiple runs, audit tool usage efficiency, or generate improvement recommendations. Triggers on: 提升, 改进, anti-pattern, 自我检查, 为什么表现差, 优化 agent, review trajectory, improve agent, analyze runs, what went wrong across runs.

Archive experiment results to the Notes repo with a structured research log entry, experiment data, reproduction steps, and context updates. Use when asked to take a note, archive experiment results, record a run, document llm-as-harness results, or save an experiment report.

Use when an RCA scenario agent failed or is suspected to have failed, and Codex needs to run AgentM fork experiments with targeted reminders to help the agent notice reasoning gaps, missed evidence, wrong assumptions, premature narrowing, root/effect confusion, or output synthesis errors. This skill is for research on when to inject hints and what to hint: compare multiple fork points and reminder strategies, using only information available to the forked agent at that point, then evaluate whether later forks recover the correct RCA answer.

DuckDB query guide for RCA: ground rules (column quoting, duration units, delta pattern), common pitfalls, and index of signal-specific recipe collections for traces, logs, metrics, and cross-signal correlation.

Guide for designing closed-loop control in agent-driven DevOps pipelines — from spec writing through parallel development, test verification, code review, and merge. Use when designing a new agent workflow, adding quality gates between pipeline stages, debugging an agent that fails to converge, or when stages produce errors that cascade downstream. Also trigger on: 控制闭环, 反馈设计, 怎么让 agent 不跑偏, 纠偏, observer, 级联错误, 闭环设计, 质量门, pipeline 设计, spec 验收, 并行开发, 多次采样.

Use when a chatbot request needs current web search or clean page fetching through TinyFish. This skill uses the local tinyfish CLI via the existing bash tool, reusing credentials from `tinyfish auth login`.

Methodology for analyzing verifier batch results on the ops-lite / RCA dataset. Use whenever the user asks to check, review, or analyze verifier run results — including "看看结果", "分析下", "为什么seed没confirm", "这些case啥情况", or any request to interpret fpg_scenario.json / run_meta.json / run_summary.jsonl outputs. The core principle: never trust any single source (GT labels, seed verdicts, aggregate stats). Every judgment must be backed by your own SQL queries against the parquet data. Hasty conclusions from aggregate numbers are the #1 failure mode — always trace the causal chain.

How to write workflow scripts that orchestrate multiple agents in AgentM. Covers agent(), parallel(), pipeline(), phase(), return values, and atom_config data passing. Trigger when writing a workflow script, choosing an orchestration pattern, or when the user says 编排 / workflow / 多 agent / 并行 / 扇出 / 调度 / 写个 workflow. Also trigger when you catch yourself building prompts inside a workflow script — pass data via atom_config instead and let the agent build its own prompt.

Debug the AgentM single-process gateway by connecting a synthetic wire client and dumping the outbound frames it emits. Use when a chat client (terminal/feishu) "doesn't see" something — a slash command, a tool, status, a reply — or when diagnosing wire/session_ready/command-catalog issues. Shows exactly what the gateway puts on the wire, bypassing the TUI.

General coding guidelines for the devloop coder agent. Load this skill when writing implementation code, tests, or fixing failures. Covers code quality, testing discipline, and common pitfalls.

Ensure implementation matches the spec's interface contract. Load this skill when writing implementation code or fixing review findings. Covers contract fidelity, spec-implementation consistency, and avoiding test-implementation collusion.

How to write, evaluate, and evolve skills based on workflow execution experience. Load this skill after a workflow completes (success or failure) when you want to distill what happened into reusable knowledge.

How to diagnose workflow failures by reading structured results and traces. Load this skill when a workflow run fails or produces low-quality output and you need to identify the root cause.

Audit whether a verifier verdict (confirmed/rejected) is sound for a specific service in a fault-injection case. Trigger when the user asks to "check a verdict", "audit a rejection", "look at why X was rejected/confirmed", or wants to understand whether the verifier got it right on a particular edge. Also trigger proactively when investigating discrepancies between verifier output and GT labels.

Shared operational guide for RCA agents (orchestrator and critic): data sources, tooling quickstart, service profile protocol, tool discipline, and anomaly definition.

A mandatory step-by-step first-move protocol to prevent empty-trajectory failures by ensuring the agent always begins investigation with concrete tool calls.

After finding one root cause service, systematically check all related services for additional faults before concluding.

Operate Feishu / Lark resources (messages, docs, calendar, bitable, mail, wiki, tasks, OKR, attendance, ...) through the official `lark-cli` binary invoked via the `bash` tool. Trigger when the user asks the agent to read or write anything in their Feishu / Lark workspace, or mentions "飞书" / "Lark" alongside an action verb (read messages, send, create doc, schedule meeting, search wiki, draft mail reply, build dashboard, etc.).