- name
SKILL.md
技能檔案保留在發布者的儲存庫中。
在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add sickn33/agentic-awesome-skills --skill agent-harness-fault-injection -a cursor -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add sickn33/agentic-awesome-skills --skill agent-harness-fault-injection -a claude-code -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add sickn33/agentic-awesome-skills --skill agent-harness-fault-injection -a codex -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add sickn33/agentic-awesome-skills --skill agent-harness-fault-injection -a grok -y在 Grok Bot 聊天中傳送複製的訊息。Bot 會從 GitHub 取得這個 SKILL.md,並加入你所有 Bot 共用的技能中。
Install this skill: https://github.com/sickn33/agentic-awesome-skills/tree/HEAD/plugins/agentic-awesome-skills-claude/skills/agent-harness-fault-injection來自發布者
Use a deterministic, non-production fault schedule to test whether an agent workflow preserves state, budgets, safety boundaries, and evidence when a dependency fails. The output is a small fault matrix, an event timeline, and a verdict that distinguishes recovered, contained, unrecoverable, and inconclusive runs.
Do not use this skill against a production target, real user data, live credentials, or an unbounded external service. Convert those cases to a local simulator or an authorized staging harness first.
inconclusive.Write the invariant before injecting a fault. A useful contract names the state that must survive and the side effects that must not repeat:
After recovery, resume from the latest durable checkpoint, preserve the task
identity and safety policy, spend no more than the remaining retry/deadline
budget, and commit each externally visible effect at most once.
Model the workflow with explicit states. For example:
created -> running -> checkpointed -> waiting_for_tool
| |
v v
failed <--------- recovering -> resumed -> completed
For each transition, define the owner, durable fields, allowed retry count, and terminal behavior. In-memory values are not checkpoints unless the harness proves they survive the simulated restart.
Select the smallest set of faults that covers the new recovery logic. Do not randomize the schedule until a deterministic schedule has passed.
Use event numbers rather than wall-clock randomness. A schedule should be portable across harnesses:
{
"seed": "harness-fixture-07",
"faults": [
{"event": "tool.call", "ordinal": 2, "kind": "timeout", "tool": "search"},
{"event": "worker.start", "ordinal": 2, "kind": "restart"},
{"event": "branch.join", "ordinal": 1, "kind": "partial_failure", "branch": "summarize"}
]
}
The harness should emit the schedule, not merely the seed. Keep fault identity separate from the observed error so a wrapper cannot accidentally turn a timeout into a generic failure. Run the same schedule twice and compare the normalized timeline before trying a different schedule.
Represent each branch as its own child attempt. The join record must retain success, failure, timeout, and not-started states. Choose one predeclared join policy:
all_required: any required branch failure stops the join;best_effort: continue with an explicit degraded marker;compensate: run a bounded compensating action and then stop or resume.Never let a successful sibling erase a failed branch from the final ledger.
Clear only the ephemeral context named in the schedule. Rebuild from the
checkpoint and durable evidence, then check that the agent does not fabricate
missing user intent, tool output, or approval. If a required fact is absent,
the safe result is inconclusive or a human clarification state.
Track remaining attempts and remaining time after every event. Do not reset a budget on a worker restart or branch retry. Use these verdicts:
contained_failure is not autonomous success. Report it separately from
completed work and include the terminal reason.
Produce one machine-readable record and one concise human summary. Every event
should include run_id, monotonic seq, logical time, state_before,
state_after, actor, event, fault_id (when injected), attempt,
checkpoint_seq, retry_remaining, deadline_remaining_ms, and a redacted
evidence_ref.
{
"run_id": "fi-2026-08-19-07",
"verdict": "recovered",
"invariants": {"resume_from_checkpoint": "pass", "effect_at_most_once": "pass", "budget": "pass"},
"faults": [{"id": "f1", "kind": "tool_timeout", "at": "tool.call#2", "handled": true}],
"timeline": [
{"seq": 4, "event": "checkpoint.write", "checkpoint_seq": 3},
{"seq": 5, "event": "tool.timeout", "fault_id": "f1", "retry_remaining": 1},
{"seq": 8, "event": "workflow.completed", "checkpoint_seq": 4}
],
"limitations": ["Tool output was synthetic; no deployed MCP was exercised."]
}
The human summary should state the frozen contract, injected schedule, verdict, failed invariants, budget consumption, and the narrowest next verification. Redact prompts, tokens, private records, and tool payloads; stable references are enough for replay.
Fixture: checkout planner / seed harness-fixture-07
Schedule: search timeout on call 2; worker restart after checkpoint 3
Policy: one retry, 2s deadline, all_required branch join
Result: recovered
Proof: checkpoint 3 reloaded, search request key replayed once, no duplicate
commit, deadline remaining 640ms, final ledger contains both branch outcomes.
inconclusive, never as a passing recovery.inconclusive or unrecoverable with evidence.@agent-evaluation-reporting - Report autonomous, assisted, failed, timed-out, and invalid outcomes.@cross-platform-contract-propagation-audit - Trace recovery fields and status contracts across consumers.@multi-agent-patterns - Choose a multi-agent topology before testing its failure behavior.Use when an agent workflow needs deterministic recovery evidence for sandbox, MCP/tool, worker, checkpoint, memory, or orchestration failures.
---name: agent-harness-fault-injectiondescription: Use when an agent workflow needs deterministic recovery evidence for sandbox, MCP/tool, worker, checkpoint…---# plugins/agentic-awesome-skills-claude/skills/agent-harness-fault-injection/SKILL.md常見問題
從發布者下載 plugins/agentic-awesome-skills-claude/skills/agent-harness-fault-injection/SKILL.md,或複製路徑後開啟儲存庫。PlugMCP 只收錄這項技能,不代管檔案。
發布者將它存放在 https://github.com/sickn33/agentic-awesome-skills。
Claude、Cursor、Codex 和 Grok Build 可以讀取你從發布者儲存庫加入的 SKILL.md。在聊天中傳送技能的 GitHub 資料夾連結,Grok Bot 就會安裝它。