[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6DST8A0S1H4B3DK89YZT002":3,"skill-same:01M6DST8A0S1H4B3DK89YZT002":109},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":23,"publishedAt":43,"linkOnly":44,"markdownContent":45,"markdownBody":46,"frontMatter":47,"refName":48,"commitSha":49,"skillDirectory":50,"forksCount":51,"repositoryHtmlUrl":52,"repositoryHomepageUrl":53,"ownerHtmlUrl":54,"ownerBlogUrl":55,"ownerDescription":6,"updatedAt":56,"install":57,"resources":70},"01M6DST8A0S1H4B3DK89YZT002","oacp",null,"Use the OACP (Open Agent Coordination Protocol) CLI and message protocol to coordinate multi-agent work — send messages between agents, sign and verify them, watch inboxes, run review loops, follow OACP conventions. Trigger whenever the user mentions OACP, runs `oacp` commands, references `.oacp` workspace markers, dispatches messages between agents (notifications, task requests, review requests), or coordinates work across runtimes — even if they don't say \"OACP\" explicitly.","kiloloop\u002Foacp-skills","kiloloop","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F264514437?v=4","01M6DSS1K2JKZD3YRMXRH54968","Kiloloop","01M6DSS1SMM810F7327W2YX7ZZ",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Fskills\u002Foacp\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[5,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42],"say","trigger","watch","commands","across","message","task","coordinate","references","review","use","work","multi-agent","coordination","open","messages","notifications","run","cli","2026-11-10 10:12:30",false,"---\nname: oacp\ndescription: Use the OACP (Open Agent Coordination Protocol) CLI and message protocol to coordinate multi-agent work — send messages between agents, sign and verify them, watch inboxes, run review loops, follow OACP conventions. Trigger whenever the user mentions OACP, runs `oacp` commands, references `.oacp` workspace markers, dispatches messages between agents (notifications, task requests, review requests), or coordinates work across runtimes — even if they don't say \"OACP\" explicitly.\n---\n\n# \u002Foacp — OACP Protocol & CLI\n\nOACP is a file-based coordination protocol for multi-agent workflows. Each project lives under `$OACP_HOME\u002Fprojects\u002F\u003Cproject>\u002F`, each agent has an `inbox\u002F` and `outbox\u002F`, and messages are YAML files. This skill teaches you how to operate within that protocol — verify the environment, send messages, watch inboxes, and follow review-loop semantics.\n\nFor runtime-agnostic protocol concepts (message schema, lifecycle, type semantics) read `shared\u002FINTENT.md`.\n\n## 1. Version check (run this first)\n\nBefore any OACP operation, verify the CLI is installed and meets the version contract declared in `skill.yaml` (`requires_oacp: \">=0.5.3\"`):\n\n```bash\noacp --version\n```\n\nIf the command is missing or the version is below `0.5.3`, fail loudly and ask the user to install or upgrade:\n\n```bash\npip install --upgrade 'oacp-cli[crypto]>=0.5.3'   # [crypto] extra enables message signing + verification\n# or, if installed via uv:\nuv tool install --upgrade 'oacp-cli[crypto]'\n```\n\nWhy this matters: protocol fields and CLI flags evolve. Running this skill against an older CLI produces silent schema drift — messages may write but parse incorrectly, or `oacp watch` flags may not exist. Verify first; fix before continuing.\n\n## 2. Workspace orientation\n\nOACP workspaces live outside the project repo. Three things must exist:\n\n- **`$OACP_HOME`** — environment variable. Defaults to `$HOME\u002Foacp`. Set it in your shell profile if your workspace is elsewhere.\n- **`.oacp` marker** — file in the project repo root. Either a JSON file with `{\"project_name\": \"\u003Cname>\"}` or a symlink into `$OACP_HOME\u002Fprojects\u002F\u003Cname>\u002F`.\n- **Project workspace** — `$OACP_HOME\u002Fprojects\u002F\u003Cproject>\u002Fagents\u002F\u003Cagent>\u002F{inbox,outbox}\u002F`.\n\nTo create a fresh workspace:\n\n```bash\noacp init \u003Cproject> --agents claude,codex      # create project + agent inboxes\noacp setup claude --project \u003Cproject>          # write .oacp marker + runtime config in CWD\noacp doctor --project \u003Cproject>                # verify workspace health\n```\n\nTo add an agent later:\n\n```bash\noacp agent add \u003Cproject> \u003Cagent> --runtime \u003Cclaude|codex|...>\n```\n\n## 3. Sending a message — the canonical workflow\n\n`oacp send` composes and writes a protocol-compliant message to the recipient's inbox and the sender's outbox in one call. Use it as the only sanctioned way to send — never hand-write inbox YAML.\n\n```bash\noacp send \u003Cproject> \\\n  --from \u003Csender_agent> \\\n  --to \u003Crecipient_agent> \\\n  --type \u003Ctype> \\\n  --subject \"\u003Cone-line>\" \\\n  --body \"\u003Cmarkdown body>\" \\\n  --priority P1 \\\n  --parent-message-id \u003Cid-of-message-this-replies-to>   # only on a reply\n```\n\nUseful optional flags:\n\n- `--body-file \u003Cpath>` — read body from a file (preferred for multi-paragraph bodies)\n- `--related-pr \u003CN>` — link to a GitHub PR (used by `review_request` \u002F `review_feedback`)\n- `--related-packet \u003Cid>` — link to a design packet (review for specs, not PRs)\n- `--in-reply-to \u003Cid>` \u002F `--conversation-id \u003Cid>` — thread linkage\n- `--oacp-dir \u003Cpath>` — override `$OACP_HOME` (rarely needed)\n- `--dry-run` — print the YAML without writing\n- `--json` — machine-readable output\n\n**Verify the send landed.** `oacp send` prints the inbox path it wrote to. Confirm the file exists:\n\n```bash\ntest -f \"\u003Cinbox_path_from_oacp_send_output>\" && echo OK\n```\n\nDo **not** grep for the `msg-id` — the on-disk filename hash differs from the in-body id, so an id grep returns false negatives.\n\n### Picking a message type\n\n| Type | When to use |\n|------|-------------|\n| `notification` | Inform the recipient of something — no reply required. |\n| `task_request` | Ask the recipient to do work. They reply with `notification` when done. |\n| `question` | Ask for information. They reply with `notification`. |\n| `review_request` | Request review of a PR (`--related-pr`) or design packet (`--related-packet`). |\n| `review_feedback` | Reviewer → author: changes requested. |\n| `review_lgtm` | Reviewer → author: approval. |\n| `review_addressed` | Author → reviewer: feedback addressed. Informational — does **not** trigger re-review on its own; send a fresh `review_request` to start round N+1. |\n| `handoff` | Transfer context for the recipient to continue. They reply with `handoff_complete`. |\n| `handoff_complete` | Acknowledge the handoff was received and acted on. |\n| `brainstorm_request` | Ask for unstructured exploration. |\n| `brainstorm_followup` | Continue a brainstorm thread with new constraints. |\n| `follow_up` | Generic follow-up on an earlier thread when no other type fits. |\n\nSee `shared\u002FINTENT.md` for the full lifecycle of each type.\n\n### Worked example — dispatch a task end-to-end\n\n```bash\n# 1. Compose the body in a file (cleaner for multi-paragraph bodies).\ncat > \u002Ftmp\u002Ftask-body.md \u003C\u003C'EOF'\n## Task\nAdd unit tests for the parser module.\n\n## Acceptance\n- All public functions covered\n- Tests pass under the existing pytest config\n\n## Done criteria\nPR merged with green CI.\nEOF\n\n# 2. Send the task_request.\noacp send my-project \\\n  --from claude --to codex \\\n  --type task_request \\\n  --priority P1 \\\n  --subject \"Add unit tests for parser module\" \\\n  --body-file \u002Ftmp\u002Ftask-body.md\n\n# 3. The recipient processes via \u002Fcheck-inbox and replies on completion:\n#    oacp send my-project --from codex --to claude --type notification \\\n#      --subject \"Re: Add unit tests for parser module\" \\\n#      --parent-message-id \u003Cyour_msg_id> \\\n#      --body \"Done — merged in \u003CPR-link>.\"\n\n# 4. You receive the reply via your inbox watcher (Section 4).\n```\n\n## 4. Reading the inbox\n\nTwo CLIs: one-shot listing vs. event-driven watching.\n\n**One-shot** — `oacp inbox \u003Cproject> --agent \u003Cagent>` lists pending messages without consuming them.\n\n**Event-driven** — `oacp watch` emits a line per new message, suitable for a Monitor process or a shell loop:\n\n```bash\noacp watch --project my-project --agent claude                       # human-readable\noacp watch --project my-project --agent claude --json                # JSON Lines\noacp watch --project my-project --agent claude --state-id watch-1a2b --since epoch\noacp watch --all-projects --agent claude                             # multi-project (coordinator pattern)\n```\n\nFlags:\n\n- `--state-id \u003Cid>` (v0.4.0+) — use a per-subscriber cursor file (`state\u002Fwatch\u002F\u003Cagent>.\u003Cstate-id>.json`) so multiple concurrent watchers each receive every event independently. Omit for the legacy shared cursor. With per-subscriber cursors, **duplicate delivery across sessions is the norm** — confirm a message file still exists before processing it.\n- `--since now` (default) suppresses replay of pre-existing inbox messages on first run; `--since epoch` replays the existing backlog once on the first scan (recommended with a fresh `--state-id` so messages predating the watcher still fire); `--since 5m` \u002F `2h` \u002F ISO-8601 for windows. Only applies before the cursor file exists.\n- `--show-archived` emits events when messages disappear (off by default — self-loops are noisy when the watching agent is also the inbox owner).\n\n**Don't process messages here.** This skill teaches the *plumbing*. To actually act on incoming messages by type, invoke the companion `\u002Fcheck-inbox` skill, which has the auto-execute rules. The recommended pattern: a Monitor running `oacp watch` emits `NEW_MESSAGE` lines as chat notifications; you respond to each by running `\u002Fcheck-inbox`.\n\n## 4b. Message signing & trust (v0.4.2+)\n\nOACP messages carry an optional detached-JWS **auth trailer** (Ed25519). Signing is opt-in per sender; verification posture is per receiver. Requires the `[crypto]` extra.\n\n```bash\noacp trust keygen --agent claude               # generate a signing key under $OACP_HOME\u002Fkeys\u002F\noacp trust keys                                # inspect local keys\noacp trust import \u003Ckid>.pub.json \\\n  --project my-project --agent codex           # pin claude's identity for receiver codex\noacp trust list --project my-project           # catalog entries + receiver pins\noacp send my-project --from claude --to codex --type notification \\\n  --subject \"signed hello\" --body \"...\" --sign # sign one message explicitly\noacp verify \u003Cmessage.yaml> \\\n  --project my-project --receiver codex        # check the trailer against codex's pins\n```\n\n- **Sender side**: `--sign` signs one message; setting the sender's `signing.sign_messages` config knob signs every send by default.\n- **Receiver side**: `signing.verify_mode` in `agents\u002F\u003Creceiver>\u002Fconfig.yaml` — `warn` reports verification results without rejecting; `enforce` makes the autonomy gate reject anything not `signed-verified` at intake (evidence copy quarantined to `dead_letter\u002F`, decision `intake_rejected`). Pin every peer **before** processing their first message — under enforce, an unpinned sender's messages cannot verify.\n- **Audit stamping**: after an autonomy decision record exists, `oacp verify \u003Cmsg> ... --attach-audit \u003Caudit.yaml>` writes the canonical `result.message_auth` block into it. Never hand-write that block.\n- **Policy signing**: `oacp trust sign-policy --project \u003Cp> --agent \u003Creceiver>` signs the receiver's own policy files so the gate can detect tampering (`policy_auth` block in every decision).\n\n## 5. Review loop — Author → Reviewer (direct)\n\nThe OACP review loop is **agent-to-agent direct dispatch**, not coordinator-mediated. When you open a PR and want review:\n\n1. **Author sends `review_request` directly to the reviewer's inbox** (`--to \u003Creviewer>`, `--related-pr \u003CN>`). Do not route through a coordinator.\n2. Reviewer's `\u002Fcheck-inbox` picks it up and dispatches `\u002Freview-loop-reviewer`.\n3. Reviewer replies with `review_lgtm` (approve) or `review_feedback` (blocking \u002F non-blocking findings).\n4. On feedback, author runs `\u002Freview-loop-author` to address it, pushes the fix, sends `review_addressed` (informational), then sends a **fresh `review_request`** to trigger round N+1. The reviewer is stateless — `review_addressed` alone does not invoke a new review.\n5. Loop until `review_lgtm`. Author merges.\n\nWhen a 2-LGTM gate is enforced, it typically counts: author's self-review (informational, since GitHub Apps cannot formally approve their own PRs) + reviewer's `review_lgtm`.\n\n**Continuation grants (v0.4.3+)**: receivers may enable `continuation_grants` in their autonomy config so that, after one human approval of a review loop, subsequent same-thread `review_request` rounds dispatch without a per-round confirm. The grant is run-authorization only — it never confers verdict authority, and reviewer-side guards apply identically on granted rounds. See the `\u002Fcheck-inbox` skill's autonomy reference for the full semantics.\n\n## 6. Health checks\n\nRun `oacp doctor` to verify environment, workspace, inbox health, schemas, and agent status:\n\n```bash\noacp doctor                                       # env + global checks\noacp doctor --project \u003Cproject>                   # full workspace + agent status\noacp doctor --project \u003Cproject> --fix             # auto-fix safe issues (missing inboxes, stale timestamps)\noacp doctor --project \u003Cproject> --fix --json      # structured output\n```\n\nFor deeper diagnostics or interactive triage, invoke the companion `\u002Fdoctor` skill.\n\n## 7. Org memory (cross-project facts)\n\nOACP supports an org-level memory layer at `$OACP_HOME\u002Forg-memory\u002F` for facts that span projects (decisions, rules, recent events). Its commands come from the separate [agent-memory-cli](https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Fagent-memory) package; oacp-cli no longer ships memory commands.\n\n```bash\nagent-memory org init --home \"$OACP_HOME\"                            # bootstrap org-memory\nagent-memory event write --home \"$OACP_HOME\" --agent claude \\\n  --project my-project --type decision --slug api-style \\\n  --body \"Use REST for public APIs\"                                   # publish a structured event\n```\n\nUse `agent-memory event write` for structured cross-project facts (decisions, incidents, releases). Use `oacp send` for agent-to-agent messages within a project. They are different primitives — don't conflate.\n\n## 8. Validating a message\n\nIf `oacp send` rejects a message, or you've inherited a hand-written YAML, validate it explicitly:\n\n```bash\noacp validate \u003Cpath-to-message.yaml>\n```\n\nThis checks the YAML against the protocol schema and reports the first failure with field path and reason.\n\n## 9. Conventions\n\n- **Archive, don't delete.** Recipients move fully-processed messages to `inbox\u002Farchive\u002F` with a no-clobber atomic move under the original filename — byte-preserving, so signed-message evidence survives. The legacy `processed\u002F` directory is not a protocol location, and deleting loses receiver-side evidence. Pending, held, or approval-gated messages stay in `inbox\u002F`.\n- **Subject prefixes:** `WIP:` for in-progress updates, `Re:` for replies (auto-applied when you set `--parent-message-id`).\n- **Priorities:** `P0` (blocker, page now) → `P3` (informational). Default to `P2`. Use `P1` for actionable-this-session.\n- **Notify on merge, not on PR-open.** When working a `task_request`, the final completion `notification` goes after merge, not at PR-open. Use a `WIP:` notification for progress updates if needed.\n- **One-message-one-task.** Don't pile multiple distinct asks into one message body — split into separate `task_request`s so each can be tracked and replied to independently.\n\n## 10. Companion skills\n\n| Skill | When to invoke |\n|-------|----------------|\n| `\u002Fcheck-inbox` | Process pending inbox messages by type. Pair with `oacp watch` for continuous monitoring. |\n| `\u002Fdoctor` | Run `oacp doctor` with auto-fix and report blockers. |\n| `\u002Freview-loop-reviewer` | You received a `review_request` — analyze the PR and reply with findings. |\n| `\u002Freview-loop-author` | You received `review_feedback` — apply fixes and reply with `review_addressed`. |\n| `\u002Fself-improve` | Audit and improve your own skills, memory, and config. |\n\n## Failure modes to recognize\n\n- **`oacp send` succeeds, recipient never sees it.** The send can write the outbox copy and silently fail the inbox write. Always verify with `test -f \"\u003Cinbox_path>\"` against the path `oacp send` printed.\n- **Filename hash ≠ `msg-id`.** A message with id `msg-20260430-claude-39dd` may land at `..._d453.yaml`. Verify by path, not by id grep.\n- **Sandbox blocks `oacp send`.** If your runtime sandboxes filesystem writes outside the project repo, `$OACP_HOME` writes will fail. Surface the error to the user — don't paper over it.\n- **Stale CLI.** If a flag listed here doesn't exist (`oacp watch --state-id`, `oacp verify --attach-audit`), re-run the version check from Section 1 — the installed CLI may be older than the contract.\n- **Message held as unverified.** Under a receiver's `enforce` posture, a message from an unpinned or badly-signed sender is quarantined to `dead_letter\u002F` instead of processing. That is the protocol working, not a bug — import the sender's key stub (`oacp trust import`) and have them re-send, or surface to the user if the signature is genuinely invalid.\n","# \u002Foacp — OACP Protocol & CLI\n\nOACP is a file-based coordination protocol for multi-agent workflows. Each project lives under `$OACP_HOME\u002Fprojects\u002F\u003Cproject>\u002F`, each agent has an `inbox\u002F` and `outbox\u002F`, and messages are YAML files. This skill teaches you how to operate within that protocol — verify the environment, send messages, watch inboxes, and follow review-loop semantics.\n\nFor runtime-agnostic protocol concepts (message schema, lifecycle, type semantics) read `shared\u002FINTENT.md`.\n\n## 1. Version check (run this first)\n\nBefore any OACP operation, verify the CLI is installed and meets the version contract declared in `skill.yaml` (`requires_oacp: \">=0.5.3\"`):\n\n```bash\noacp --version\n```\n\nIf the command is missing or the version is below `0.5.3`, fail loudly and ask the user to install or upgrade:\n\n```bash\npip install --upgrade 'oacp-cli[crypto]>=0.5.3'   # [crypto] extra enables message signing + verification\n# or, if installed via uv:\nuv tool install --upgrade 'oacp-cli[crypto]'\n```\n\nWhy this matters: protocol fields and CLI flags evolve. Running this skill against an older CLI produces silent schema drift — messages may write but parse incorrectly, or `oacp watch` flags may not exist. Verify first; fix before continuing.\n\n## 2. Workspace orientation\n\nOACP workspaces live outside the project repo. Three things must exist:\n\n- **`$OACP_HOME`** — environment variable. Defaults to `$HOME\u002Foacp`. Set it in your shell profile if your workspace is elsewhere.\n- **`.oacp` marker** — file in the project repo root. Either a JSON file with `{\"project_name\": \"\u003Cname>\"}` or a symlink into `$OACP_HOME\u002Fprojects\u002F\u003Cname>\u002F`.\n- **Project workspace** — `$OACP_HOME\u002Fprojects\u002F\u003Cproject>\u002Fagents\u002F\u003Cagent>\u002F{inbox,outbox}\u002F`.\n\nTo create a fresh workspace:\n\n```bash\noacp init \u003Cproject> --agents claude,codex      # create project + agent inboxes\noacp setup claude --project \u003Cproject>          # write .oacp marker + runtime config in CWD\noacp doctor --project \u003Cproject>                # verify workspace health\n```\n\nTo add an agent later:\n\n```bash\noacp agent add \u003Cproject> \u003Cagent> --runtime \u003Cclaude|codex|...>\n```\n\n## 3. Sending a message — the canonical workflow\n\n`oacp send` composes and writes a protocol-compliant message to the recipient's inbox and the sender's outbox in one call. Use it as the only sanctioned way to send — never hand-write inbox YAML.\n\n```bash\noacp send \u003Cproject> \\\n  --from \u003Csender_agent> \\\n  --to \u003Crecipient_agent> \\\n  --type \u003Ctype> \\\n  --subject \"\u003Cone-line>\" \\\n  --body \"\u003Cmarkdown body>\" \\\n  --priority P1 \\\n  --parent-message-id \u003Cid-of-message-this-replies-to>   # only on a reply\n```\n\nUseful optional flags:\n\n- `--body-file \u003Cpath>` — read body from a file (preferred for multi-paragraph bodies)\n- `--related-pr \u003CN>` — link to a GitHub PR (used by `review_request` \u002F `review_feedback`)\n- `--related-packet \u003Cid>` — link to a design packet (review for specs, not PRs)\n- `--in-reply-to \u003Cid>` \u002F `--conversation-id \u003Cid>` — thread linkage\n- `--oacp-dir \u003Cpath>` — override `$OACP_HOME` (rarely needed)\n- `--dry-run` — print the YAML without writing\n- `--json` — machine-readable output\n\n**Verify the send landed.** `oacp send` prints the inbox path it wrote to. Confirm the file exists:\n\n```bash\ntest -f \"\u003Cinbox_path_from_oacp_send_output>\" && echo OK\n```\n\nDo **not** grep for the `msg-id` — the on-disk filename hash differs from the in-body id, so an id grep returns false negatives.\n\n### Picking a message type\n\n| Type | When to use |\n|------|-------------|\n| `notification` | Inform the recipient of something — no reply required. |\n| `task_request` | Ask the recipient to do work. They reply with `notification` when done. |\n| `question` | Ask for information. They reply with `notification`. |\n| `review_request` | Request review of a PR (`--related-pr`) or design packet (`--related-packet`). |\n| `review_feedback` | Reviewer → author: changes requested. |\n| `review_lgtm` | Reviewer → author: approval. |\n| `review_addressed` | Author → reviewer: feedback addressed. Informational — does **not** trigger re-review on its own; send a fresh `review_request` to start round N+1. |\n| `handoff` | Transfer context for the recipient to continue. They reply with `handoff_complete`. |\n| `handoff_complete` | Acknowledge the handoff was received and acted on. |\n| `brainstorm_request` | Ask for unstructured exploration. |\n| `brainstorm_followup` | Continue a brainstorm thread with new constraints. |\n| `follow_up` | Generic follow-up on an earlier thread when no other type fits. |\n\nSee `shared\u002FINTENT.md` for the full lifecycle of each type.\n\n### Worked example — dispatch a task end-to-end\n\n```bash\n# 1. Compose the body in a file (cleaner for multi-paragraph bodies).\ncat > \u002Ftmp\u002Ftask-body.md \u003C\u003C'EOF'\n## Task\nAdd unit tests for the parser module.\n\n## Acceptance\n- All public functions covered\n- Tests pass under the existing pytest config\n\n## Done criteria\nPR merged with green CI.\nEOF\n\n# 2. Send the task_request.\noacp send my-project \\\n  --from claude --to codex \\\n  --type task_request \\\n  --priority P1 \\\n  --subject \"Add unit tests for parser module\" \\\n  --body-file \u002Ftmp\u002Ftask-body.md\n\n# 3. The recipient processes via \u002Fcheck-inbox and replies on completion:\n#    oacp send my-project --from codex --to claude --type notification \\\n#      --subject \"Re: Add unit tests for parser module\" \\\n#      --parent-message-id \u003Cyour_msg_id> \\\n#      --body \"Done — merged in \u003CPR-link>.\"\n\n# 4. You receive the reply via your inbox watcher (Section 4).\n```\n\n## 4. Reading the inbox\n\nTwo CLIs: one-shot listing vs. event-driven watching.\n\n**One-shot** — `oacp inbox \u003Cproject> --agent \u003Cagent>` lists pending messages without consuming them.\n\n**Event-driven** — `oacp watch` emits a line per new message, suitable for a Monitor process or a shell loop:\n\n```bash\noacp watch --project my-project --agent claude                       # human-readable\noacp watch --project my-project --agent claude --json                # JSON Lines\noacp watch --project my-project --agent claude --state-id watch-1a2b --since epoch\noacp watch --all-projects --agent claude                             # multi-project (coordinator pattern)\n```\n\nFlags:\n\n- `--state-id \u003Cid>` (v0.4.0+) — use a per-subscriber cursor file (`state\u002Fwatch\u002F\u003Cagent>.\u003Cstate-id>.json`) so multiple concurrent watchers each receive every event independently. Omit for the legacy shared cursor. With per-subscriber cursors, **duplicate delivery across sessions is the norm** — confirm a message file still exists before processing it.\n- `--since now` (default) suppresses replay of pre-existing inbox messages on first run; `--since epoch` replays the existing backlog once on the first scan (recommended with a fresh `--state-id` so messages predating the watcher still fire); `--since 5m` \u002F `2h` \u002F ISO-8601 for windows. Only applies before the cursor file exists.\n- `--show-archived` emits events when messages disappear (off by default — self-loops are noisy when the watching agent is also the inbox owner).\n\n**Don't process messages here.** This skill teaches the *plumbing*. To actually act on incoming messages by type, invoke the companion `\u002Fcheck-inbox` skill, which has the auto-execute rules. The recommended pattern: a Monitor running `oacp watch` emits `NEW_MESSAGE` lines as chat notifications; you respond to each by running `\u002Fcheck-inbox`.\n\n## 4b. Message signing & trust (v0.4.2+)\n\nOACP messages carry an optional detached-JWS **auth trailer** (Ed25519). Signing is opt-in per sender; verification posture is per receiver. Requires the `[crypto]` extra.\n\n```bash\noacp trust keygen --agent claude               # generate a signing key under $OACP_HOME\u002Fkeys\u002F\noacp trust keys                                # inspect local keys\noacp trust import \u003Ckid>.pub.json \\\n  --project my-project --agent codex           # pin claude's identity for receiver codex\noacp trust list --project my-project           # catalog entries + receiver pins\noacp send my-project --from claude --to codex --type notification \\\n  --subject \"signed hello\" --body \"...\" --sign # sign one message explicitly\noacp verify \u003Cmessage.yaml> \\\n  --project my-project --receiver codex        # check the trailer against codex's pins\n```\n\n- **Sender side**: `--sign` signs one message; setting the sender's `signing.sign_messages` config knob signs every send by default.\n- **Receiver side**: `signing.verify_mode` in `agents\u002F\u003Creceiver>\u002Fconfig.yaml` — `warn` reports verification results without rejecting; `enforce` makes the autonomy gate reject anything not `signed-verified` at intake (evidence copy quarantined to `dead_letter\u002F`, decision `intake_rejected`). Pin every peer **before** processing their first message — under enforce, an unpinned sender's messages cannot verify.\n- **Audit stamping**: after an autonomy decision record exists, `oacp verify \u003Cmsg> ... --attach-audit \u003Caudit.yaml>` writes the canonical `result.message_auth` block into it. Never hand-write that block.\n- **Policy signing**: `oacp trust sign-policy --project \u003Cp> --agent \u003Creceiver>` signs the receiver's own policy files so the gate can detect tampering (`policy_auth` block in every decision).\n\n## 5. Review loop — Author → Reviewer (direct)\n\nThe OACP review loop is **agent-to-agent direct dispatch**, not coordinator-mediated. When you open a PR and want review:\n\n1. **Author sends `review_request` directly to the reviewer's inbox** (`--to \u003Creviewer>`, `--related-pr \u003CN>`). Do not route through a coordinator.\n2. Reviewer's `\u002Fcheck-inbox` picks it up and dispatches `\u002Freview-loop-reviewer`.\n3. Reviewer replies with `review_lgtm` (approve) or `review_feedback` (blocking \u002F non-blocking findings).\n4. On feedback, author runs `\u002Freview-loop-author` to address it, pushes the fix, sends `review_addressed` (informational), then sends a **fresh `review_request`** to trigger round N+1. The reviewer is stateless — `review_addressed` alone does not invoke a new review.\n5. Loop until `review_lgtm`. Author merges.\n\nWhen a 2-LGTM gate is enforced, it typically counts: author's self-review (informational, since GitHub Apps cannot formally approve their own PRs) + reviewer's `review_lgtm`.\n\n**Continuation grants (v0.4.3+)**: receivers may enable `continuation_grants` in their autonomy config so that, after one human approval of a review loop, subsequent same-thread `review_request` rounds dispatch without a per-round confirm. The grant is run-authorization only — it never confers verdict authority, and reviewer-side guards apply identically on granted rounds. See the `\u002Fcheck-inbox` skill's autonomy reference for the full semantics.\n\n## 6. Health checks\n\nRun `oacp doctor` to verify environment, workspace, inbox health, schemas, and agent status:\n\n```bash\noacp doctor                                       # env + global checks\noacp doctor --project \u003Cproject>                   # full workspace + agent status\noacp doctor --project \u003Cproject> --fix             # auto-fix safe issues (missing inboxes, stale timestamps)\noacp doctor --project \u003Cproject> --fix --json      # structured output\n```\n\nFor deeper diagnostics or interactive triage, invoke the companion `\u002Fdoctor` skill.\n\n## 7. Org memory (cross-project facts)\n\nOACP supports an org-level memory layer at `$OACP_HOME\u002Forg-memory\u002F` for facts that span projects (decisions, rules, recent events). Its commands come from the separate [agent-memory-cli](https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Fagent-memory) package; oacp-cli no longer ships memory commands.\n\n```bash\nagent-memory org init --home \"$OACP_HOME\"                            # bootstrap org-memory\nagent-memory event write --home \"$OACP_HOME\" --agent claude \\\n  --project my-project --type decision --slug api-style \\\n  --body \"Use REST for public APIs\"                                   # publish a structured event\n```\n\nUse `agent-memory event write` for structured cross-project facts (decisions, incidents, releases). Use `oacp send` for agent-to-agent messages within a project. They are different primitives — don't conflate.\n\n## 8. Validating a message\n\nIf `oacp send` rejects a message, or you've inherited a hand-written YAML, validate it explicitly:\n\n```bash\noacp validate \u003Cpath-to-message.yaml>\n```\n\nThis checks the YAML against the protocol schema and reports the first failure with field path and reason.\n\n## 9. Conventions\n\n- **Archive, don't delete.** Recipients move fully-processed messages to `inbox\u002Farchive\u002F` with a no-clobber atomic move under the original filename — byte-preserving, so signed-message evidence survives. The legacy `processed\u002F` directory is not a protocol location, and deleting loses receiver-side evidence. Pending, held, or approval-gated messages stay in `inbox\u002F`.\n- **Subject prefixes:** `WIP:` for in-progress updates, `Re:` for replies (auto-applied when you set `--parent-message-id`).\n- **Priorities:** `P0` (blocker, page now) → `P3` (informational). Default to `P2`. Use `P1` for actionable-this-session.\n- **Notify on merge, not on PR-open.** When working a `task_request`, the final completion `notification` goes after merge, not at PR-open. Use a `WIP:` notification for progress updates if needed.\n- **One-message-one-task.** Don't pile multiple distinct asks into one message body — split into separate `task_request`s so each can be tracked and replied to independently.\n\n## 10. Companion skills\n\n| Skill | When to invoke |\n|-------|----------------|\n| `\u002Fcheck-inbox` | Process pending inbox messages by type. Pair with `oacp watch` for continuous monitoring. |\n| `\u002Fdoctor` | Run `oacp doctor` with auto-fix and report blockers. |\n| `\u002Freview-loop-reviewer` | You received a `review_request` — analyze the PR and reply with findings. |\n| `\u002Freview-loop-author` | You received `review_feedback` — apply fixes and reply with `review_addressed`. |\n| `\u002Fself-improve` | Audit and improve your own skills, memory, and config. |\n\n## Failure modes to recognize\n\n- **`oacp send` succeeds, recipient never sees it.** The send can write the outbox copy and silently fail the inbox write. Always verify with `test -f \"\u003Cinbox_path>\"` against the path `oacp send` printed.\n- **Filename hash ≠ `msg-id`.** A message with id `msg-20260430-claude-39dd` may land at `..._d453.yaml`. Verify by path, not by id grep.\n- **Sandbox blocks `oacp send`.** If your runtime sandboxes filesystem writes outside the project repo, `$OACP_HOME` writes will fail. Surface the error to the user — don't paper over it.\n- **Stale CLI.** If a flag listed here doesn't exist (`oacp watch --state-id`, `oacp verify --attach-audit`), re-run the version check from Section 1 — the installed CLI may be older than the contract.\n- **Message held as unverified.** Under a receiver's `enforce` posture, a message from an unpinned or badly-signed sender is quarantined to `dead_letter\u002F` instead of processing. That is the protocol working, not a bug — import the sender's key stub (`oacp trust import`) and have them re-send, or surface to the user if the signature is genuinely invalid.\n",{"name":5,"description":7},"main","ea78d5cce0208a15b59f6f2c49e729ba71a3c5af","skills\u002Foacp",1,"https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills","https:\u002F\u002Foacp.dev\u002Fskills\u002F","https:\u002F\u002Fgithub.com\u002Fkiloloop","https:\u002F\u002Fkiloloop.com","2026-10-31 14:39:52",{"source":8,"skillName":5,"commands":58,"prompt":68,"zipAvailable":69},[59,62,65],{"tool":60,"command":61},"npx","npx skills add kiloloop\u002Foacp-skills --skill oacp",{"tool":63,"command":64},"bunx","bunx skills add kiloloop\u002Foacp-skills --skill oacp",{"tool":66,"command":67},"pnpm","pnpm dlx skills add kiloloop\u002Foacp-skills --skill oacp","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Fskills\u002Foacp\u002FSKILL.md",true,[71,76,79,82,85,88,91,95,98,102,106],{"path":72,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":14,"sourceHtmlUrl":75},"algorithms.md","REFERENCE","参考文档","https:\u002F\u002Fgithub.com\u002Fhefarica\u002Farbitragex-v2\u002Fblob\u002Fmain\u002F.claude\u002Fagents\u002Fskills\u002Fbeam-search-for-high-frequency-route-pruning\u002Falgorithms.md",{"path":77,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":14,"sourceHtmlUrl":78},"failure_modes.md","https:\u002F\u002Fgithub.com\u002Fhefarica\u002Farbitragex-v2\u002Fblob\u002Fmain\u002F.claude\u002Fagents\u002Fskills\u002Fbeam-search-for-high-frequency-route-pruning\u002Ffailure_modes.md",{"path":80,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":14,"sourceHtmlUrl":81},"formulas.md","https:\u002F\u002Fgithub.com\u002Fhefarica\u002Farbitragex-v2\u002Fblob\u002Fmain\u002F.claude\u002Fagents\u002Fskills\u002Fbeam-search-for-high-frequency-route-pruning\u002Fformulas.md",{"path":83,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":14,"sourceHtmlUrl":84},"implementation_notes.md","https:\u002F\u002Fgithub.com\u002Fhefarica\u002Farbitragex-v2\u002Fblob\u002Fmain\u002F.claude\u002Fagents\u002Fskills\u002Fbeam-search-for-high-frequency-route-pruning\u002Fimplementation_notes.md",{"path":86,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":14,"sourceHtmlUrl":87},"integration_arbitragex.md","https:\u002F\u002Fgithub.com\u002Fhefarica\u002Farbitragex-v2\u002Fblob\u002Fmain\u002F.claude\u002Fagents\u002Fskills\u002Fbeam-search-for-high-frequency-route-pruning\u002Fintegration_arbitragex.md",{"path":89,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":14,"sourceHtmlUrl":90},"manifest.json","https:\u002F\u002Fgithub.com\u002Fhefarica\u002Farbitragex-v2\u002Fblob\u002Fmain\u002F.claude\u002Fagents\u002Fskills\u002Fbeam-search-for-high-frequency-route-pruning\u002Fmanifest.json",{"path":92,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":93,"sourceHtmlUrl":94},"README.md",2159,"https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Fskills\u002Foacp\u002FREADME.md",{"path":96,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":14,"sourceHtmlUrl":97},"references.json","https:\u002F\u002Fgithub.com\u002Fhefarica\u002Farbitragex-v2\u002Fblob\u002Fmain\u002F.claude\u002Fagents\u002Fskills\u002Fbeam-search-for-high-frequency-route-pruning\u002Freferences.json",{"path":99,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":100,"sourceHtmlUrl":101},"shared\u002FINTENT.md",6260,"https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Fskills\u002Foacp\u002Fshared\u002FINTENT.md",{"path":103,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":104,"sourceHtmlUrl":105},"skill.yaml",318,"https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Fskills\u002Foacp\u002Fskill.yaml",{"path":107,"resourceType":73,"resourceTypeDesc":74,"sizeBytes":14,"sourceHtmlUrl":108},"validation_tests.md","https:\u002F\u002Fgithub.com\u002Fhefarica\u002Farbitragex-v2\u002Fblob\u002Fmain\u002F.claude\u002Fagents\u002Fskills\u002Fbeam-search-for-high-frequency-route-pruning\u002Fvalidation_tests.md",[110,121,126,152,158,167],{"skillNo":111,"name":112,"nameZh":6,"description":113,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":114,"categories":115,"keywords":116,"publishedAt":43},"01M6DSWR7WDRG7C3D31R4QKW5T","\u003Cskill-name>","\u003Cone-line description>","https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Ftemplate\u002Fcodex\u002FSKILL.md",[],[117,118,119,120],"skill name","name","\u003Cskill name>","line",{"skillNo":122,"name":112,"nameZh":6,"description":113,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":123,"categories":124,"keywords":125,"publishedAt":43},"01M6DSWMRPCHQH4HJJF3MATTXZ","https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Ftemplate\u002Fclaude\u002FSKILL.md",[],[117,118,119,120],{"skillNo":127,"name":128,"nameZh":6,"description":129,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":130,"categories":131,"keywords":132,"publishedAt":43},"01M6DSWH83XKNGJSKX2XWQCRKJ","wrap-up","End-of-session cleanup for Codex — inspect stale artifacts, run debrief, write significant org-memory events, run a delta-first self-improve pass, then publish bounded approved memory and repository changes.","https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Fskills\u002Fwrap-up\u002Fcodex\u002FSKILL.md",[],[133,134,135,41,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151],"wrap up","up","wrap","memory","cleanup","artifacts","repository","publish","first","improve","write","session","stale","session cleanup","delta","self","end","debrief","codex",{"skillNo":153,"name":154,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":155,"categories":156,"keywords":157,"publishedAt":43},"01M6DSWEQBS2CE6F2H8T3VQNZY","claude","https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Fskills\u002Fwrap-up\u002Fclaude\u002FSKILL.md",[],[],{"skillNo":159,"name":128,"nameZh":6,"description":160,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":161,"categories":162,"keywords":163,"publishedAt":43},"01M6DSW99EPXDVAG8GHR785CC5","End-of-session cleanup, debrief, org-memory events, self-improve, commit, and push in one command.","https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Fskills\u002Fwrap-up\u002FSKILL.md",[],[133,134,135,148,136,164,137,165,142,166,144,146,149,150],"commit","push","command",{"skillNo":168,"name":169,"nameZh":6,"description":170,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":171,"categories":172,"keywords":173,"publishedAt":43},"01M6DSW4FRNRKRQ0TRMZAYNMKZ","self-improve","Audit and improve Codex operating guidance: skills, curated project memory, AGENTS.md, and relevant configuration. Use for explicit self-improve requests, routine wrap-up drift checks, stale or conflicting guidance, or an approved cleanup. Route release-wide OACP migrations and context-budget audits to audit-oacp-skills. Do not hand-edit generated Codex memories or installed plugin\u002Fsystem artifacts.","https:\u002F\u002Fgithub.com\u002Fkiloloop\u002Foacp-skills\u002Fblob\u002Fmain\u002Fskills\u002Fself-improve\u002Fcodex\u002FSKILL.md",[],[174,148,142,175,137,176,177,178,179,180,181,182,134,183,184,185,186,187,5,136],"self improve","project","release","audit","Budget","system","edit","context","generated","do","route","drift","memories","operating"]