inbox
Receiving side of /xm:toss — list, take, resolve, or drop cross-project bug reports that landed in this project's .xm/inbox/. Use when the user asks to check the inbox, "받은 편지함 봐줘", "what did other pr
SKILL.mdskills/3549-xm/14341-inbox/SKILL.md
GitHubx-inbox — Cross-Project Inbox
Overview
The receiving side of /xm:toss: .xm/inbox/<id>.json holds bug reports other
registered projects tossed at this one. /xm:inbox lists them, marks one as being
worked on (take), marks verified work complete (resolve), or dismisses one
that isn't relevant (drop). Every read
opportunistically archives resolved items older than 30 days locally (no network) —
neither this nor anything else in this CLI ever discards an unresolved item.
The CLI itself makes zero network calls. It cannot query mem-mesh pin state and
cannot recreate an expired pin — that requires an MCP session, which only you (the
skill, running inside Claude Code) have. Re-notifying an item whose pin already expired
is therefore something YOU do, using mcp__mem-mesh__pin_get / mcp__mem-mesh__pin_add
plus xm inbox record --scope inbox — see step 5 below.
When to Use
- "받은 편지함 봐줘" / "what's in the inbox" / "show what other projects tossed at us"
- Starting work and want to check for cross-project reports before diving in
- "이거 처리할게" / "I'll work on this one" →
take <id>(marks in-progress, returns the full repro/fix body to act on) - "수정 끝났어" / "done" →
resolve <id>after implementation and verification - "이건 아니다" / "not relevant, drop it" →
drop <id>
Do NOT Use When
- You want to SEND a report to another project — that is
/xm:toss. - You want to turn a taken item into a tracked x-build task —
takeonly returns the item's content; promoting it into a task/later item is a separate manual step (x-build tasks have no field that holds why/repro/fix_direction verbatim).
CLI Invocation
⚠ Call
xm inbox <command>directly. Claude Code's Bash tool starts a fresh shell on every invocation — shell functions (xinbox()) defined in one call do NOT persist to the next, causingcommand not found: xinbox. Never define a helper across calls; always use the dispatcher.Fallback (only when
xmis not in PATH — rare;${CLAUDE_PLUGIN_ROOT}is NOT exported to Bash subprocesses, so don't rely on it bare):XINBOX_CLI=$(ls -d ~/.claude/plugins/cache/xm/xm/*/lib/x-inbox-cli.mjs 2>/dev/null | sort -V | tail -1) node "$XINBOX_CLI" <command> [args]Forbidden:
XINBOX="node ..."; $XINBOX list— zsh treats the quoted string as a single command and fails.
Core Process
먼저 mem-mesh에서 수신 항목을 materialize합니다.
search에는 현재tags필터가 없으므로query="inbox"같은 넓은 검색으로 수신함을 판단하면 안 됩니다.pin_list에서status를 생략하면 모든 상태를 반환한다고 가정하지 마십시오. 현재 서버는 기본 상태를 선택할 수 있어 새in_progressdelivery pin이 빠질 수 있습니다. 다음 세 호출을 각각 실행하고, 반환된pins를id로 중복 제거해 하나의 배열로 합칩니다.limit은 각 호출에서 반드시 10 이하로 둡니다.mcp__mem-mesh__pin_list(project_id=<id>, tags=["inbox"], status="open", limit=10)mcp__mem-mesh__pin_list(project_id=<id>, tags=["inbox"], status="in_progress", limit=10)mcp__mem-mesh__pin_list(project_id=<id>, tags=["inbox"], status="completed", limit=10)
그 결과를 파일로 저장한 뒤
xm inbox reconcile에 먹여 무엇을 해야 하는지 판정합니다. 눈으로 대조하지 마십시오 — 이 명령이 pin 쪽과 원장 쪽 양방향 drift를 모두 계산합니다.# pin_list 결과(전체 envelope 또는 pins 배열)를 파일에 기록한 뒤 xm inbox reconcile --pins-file <path> --json출력의 action별로:
materialize— pin은 살아 있는데 로컬 원장에 항목이 없습니다. 아래 3단계로 가져옵니다.renotify— 미종결 항목의 pin이 죽었습니다(목록에 없거나completed). 7번 절차를 따릅니다.unmappable— toss id 없는 구 포맷 pin입니다. 추측하지 말고 pin id와 title을 보여 준 뒤 발신 측에 재전달을 요청합니다.none— 할 일 없음.
세 호출 중 하나라도
count == limit이면 목록이 잘렸을 수 있으므로--partial을 붙입니다. 그러면 "목록에 없음"을 pin이 죽었다는 증거로 쓰지 않아 잘못된 재알림을 막습니다. 세 결과를 합치지 않았거나 상태 하나라도 조회하지 못했을 때도--partial을 붙입니다.새 toss pin의
content는<toss id> — <title>형식입니다.materialize대상의 toss id로mcp__mem-mesh__search(query=<toss id>, project_id=<현재 프로젝트의 mem-mesh id>, limit=10)을 호출하고, 같은id를 가진 JSON 본문만 고릅니다. 그 결과의content를 아래 명령으로 materialize합니다. memory 본문은 파일로 먼저 쓰고--content-file로 넘깁니다 — 본문은 repro 명령과 캡처된 출력을 품은 JSON 문서라, 셸 인자로 넘기면 중첩 escape를 손으로 다시 맞춰야 하고 한 글자만 어긋나도 repro가 조용히 손상됩니다.# memory content를 그대로 파일에 기록한 뒤 xm inbox materialize --content-file <path> --memory-id <memory id> --pin-id <pin id> --json--content <json>인라인 형태도 여전히 동작하며, 둘 다 주면 인라인이 이깁니다. 기존 pin처럼 toss id가 없는 항목은 title 또는 일반어로 추측 검색하지 않습니다. pin id와 title을 보여 주고 발신 측에 재전달을 요청합니다. 이 CLI는 네트워크를 호출하지 않고 현재 cwd의.xm/inbox/<id>.json에만 기록합니다. 따라서 materialize 명령은 반드시 수신 프로젝트를 cwd로 하여 실행하고, 발신 프로젝트 경로·outbox·다른 checkout을 인자로 넘기지 않습니다. malformed JSON이나to_project가 현재 프로젝트와 다른 항목은 거부하고, 이미 같은 id가 있으면 로컬 상태를 바꾸지 않습니다. MCP 검색을 할 수 없으면 그 사실을 밝히고 기존 로컬 원장만 조회합니다.xm inbox list— materialize 뒤에 실행합니다. Prints unresolved items first, then in-progress, then resolved, then dismissed; add--jsonwhen you need to act on fields programmatically.Address items by
idonly, never by a remembered list position — re-runlistif unsure, since the sort order and the opportunistic archive sweep can shift between calls.xm inbox take <id>when starting work on one: relay the full returned body (why / repro command+output / fix direction) to the user, or use it directly as the starting point for a fix — don't re-derive it from scratch.takewritesstatus: "in_progress"; it never claims completion.xm inbox resolve <id> --summary "..." --verification "..." --jsononly after the implementation and relevant verification have completed.doneis an alias. Both flags are required and the command refuses the transition without them — a receipt is the only thing the sender ever sees, so one claiming "resolved" with no evidence closes the report against a fix that may not exist.--verificationmust name the concrete check you actually ran (command and its result, or the files/lines you confirmed); restating the summary is not verification. This writes the terminalstatus: "resolved", creates an immutable receipt (terminal state, time, summary and verification evidence), and returnsmcp_calls. Then, in order:- Call
mcp__mem-mesh__addwithmcp_calls.addverbatim, and record the returned memory id withxm inbox receipt record <id> --memory-id <id>. - If
mcp_calls.pin_completeis present, callmcp__mem-mesh__pin_completewith it. The delivery pin otherwise staysin_progressforever and every later session start announces a report that is already closed.
If MCP delivery fails or is unavailable, the terminal receipt remains locally
pending; do not claim the sender was notified. Usexm inbox receipt retry <id> --jsonin a later MCP-capable session to obtain the same payload (includingpin_complete), andreceipt status <id>to report the local transport state.- Call
xm inbox drop <id> --summary "..." --jsonwhen it doesn't need action.--summaryis required — say why this project is not acting on the report — but verification is not, since nothing was fixed. It follows the same receipt-delivery andpin_completeprocess as resolve. If it's ambiguous whether an item is relevant, confirm with the user before dropping — treat the drop as final in conversation even though dismissed items remain recoverable in the archive on disk.Re-notify a dead pin yourself, when relevant. For any
delivered/in_progressitem (plus legacyactioneditems) whosemem_mesh.pin_idis set (visible viaxm inbox list --json):- Call
mcp__mem-mesh__pin_get(pin_id)yourself. - If it comes back not-found, OR
status: "completed"(mem-mesh's 7-day auto-close — never a real user action on a delivery pin), the pin is dead: callmcp__mem-mesh__pin_add(content="<item id> — <item title>", project_id=<this project's mem-mesh id>, tags=["inbox"])to recreate it, then persist the new id:xm inbox record <id> --pin-id <new pin id> --scope inbox --json - If the pin is still
in_progress, or the item isresolved/dismissed, do nothing — this matchesledger.mjs'sreconcile()rules exactly (dead pin + unresolved item → renotify; anything else → no action). - If you have no MCP tools available at all (plain shell, no Claude Code MCP session), skip this step entirely and say so — don't guess at pin state or claim re-notification happened.
- Call
Common Rationalizations
| Excuse | Reality |
|---|---|
| "I'll remember item #2 from the last list and take() it later." | Lists are re-sorted (unresolved-first) and re-swept (archive) on every call — position 2 can point at a different item next time. Always address by id. |
| "Inbox is empty, something's broken." | An empty inbox is a normal, valid state — say so; don't assume the CLI or mem-mesh is malfunctioning. |
"search(query=\"inbox\")가 비었으니 새 수신함도 비었다." |
search는 tags 필터가 없고 넓은 질의는 랭킹에서 밀린다. 먼저 pin_list(tags=[\"inbox\"], limit=10)으로 알림을 찾고 toss id로 정확히 검색한다. |
"pin_list(tags=[\"inbox\"]) 한 번이면 모든 상태가 나온다." |
status 생략 시 서버 기본 상태만 반환될 수 있다. open, in_progress, completed를 각각 조회해 합쳐야 신규 delivery pin을 놓치지 않는다. |
"I compared the pin list against list's output by eye — same count, so nothing is missing." |
Counting agrees even when a pin points at an id the ledger never had. reconcile diffs both directions by id; eyeballing missed exactly this for five days (toss-20260721-666aa5a0, taken 07-21, local file gone by 07-26, pin still in_progress). Run the command. |
| "list didn't print anything about pins, so they must all still be fine." | list never checks pin state (t11 — no network in the CLI). Silence from list says nothing about pin health; you have to actually call pin_get yourself to know. |
"The memory body is short enough — I'll just paste it into --content '...'." |
Bodies embed a repro command and its captured output, both of which routinely contain quotes, backslashes and newlines. Re-quoting them by hand corrupts the repro with no error — nothing validates the repro against its source. Write the body to a file and use --content-file. |
"take() gave me the item, I'll now later promote it to keep the body." |
later promote has no field for why/repro/fix_direction — the body does not survive that trip. Use the returned item content directly instead. |
| "The fix is obvious, I'll put 'fixed' in --verification and move on." | --verification is the sender's only evidence that this project checked anything. A restatement of the summary is what let 5 x-kit items sit marked done while only 2 were actually fixed. Name the command you ran and its result, or the file:line you confirmed. |
| "I resolved it and delivered the receipt — the pin will sort itself out." | Nothing closes a delivery pin but you. Left open it reads as an unhandled report at every future session start. If mcp_calls.pin_complete is in the output, call mcp__mem-mesh__pin_complete with it. |
| "take means the issue is done." | take means work started. Only resolve is terminal, and only after the fix and checks complete. |
| "I'll drop an item I'm not sure about, just to clean up the list." | Drop is for items that genuinely don't need action. When unsure, leave it delivered (or take it) and ask the user rather than guessing it away. |
| "No MCP tools here, I'll just tell the user the inbox looks fine." | The listed items are accurate (local files), but you silently skipped pin re-notification. Say explicitly that you couldn't check/renew pin state — don't imply you did. |
| "I recreated the pin via pin_add but that's the whole job, no need to persist it." | Without xm inbox record --scope inbox, the new pin id only lives in your chat output — the next session's ledger still points at the dead pin id and would recreate ANOTHER one. Always record it. |
| "The item is old, I'll delete the file myself." | Only the CLI's own opportunistic archiveExpired() sweep (run before every read) may relocate terminal items, and only after 30 days. Manual deletion bypasses the recoverable archive path entirely. |
Red Flags
- You referenced an inbox item by list position instead of
id. - You decided what to materialize or renotify by reading the pin list yourself instead of running
xm inbox reconcile. - You used a generic memory search instead of an inbox pin's toss id to materialize a delivery.
- You dropped an item without the user's confirmation when its relevance was unclear.
- You called
resolvebefore implementation and verification completed. - You passed a
--verificationthat restates the summary instead of naming a check you ran. - You delivered a terminal receipt but skipped the
mcp_calls.pin_completecall it returned. - You treated a terminal local state as proof that the sender received its receipt.
- You claimed pin re-notification happened without actually calling
pin_get/pin_add. - You called
pin_addto renotify but never ranxm inbox record --scope inboxafterward. - You had no MCP tools available and didn't tell the user re-notification was skipped.
- You edited or deleted a
.xm/inbox/*.jsonfile directly instead of going throughtake/drop/record.
Verification
- Every item referenced by
id, matching whatlistprinted. take's full body (why/repro/fix) was relayed or acted on, not summarized away.- Completed work was closed with
resolve, while unfinished work remainedin_progress. - For resolve/drop,
mcp_calls.pin_completewas called when present, so no closed item leaves an open pin. - For resolve/drop, the receipt MCP call was delivered verbatim and its memory id recorded; otherwise the pending transport state was stated plainly and left retryable.
- Any pin re-notification you performed was via real
pin_get/pin_addMCP calls, followed byxm inbox record --scope inbox— never assumed or skipped silently. - If no MCP tools were available, that limitation was stated plainly, not glossed over.
- No direct file edits under
.xm/inbox/— only via the CLI subcommands.