Closeout Remote

Session Knowledge Closeout (Remote / No-Store Variant)

This session cannot read or write the zmem store — no store.py, no ZMEM_DATA, no local box to connect to. You cannot recall what is already known and you cannot supersede anything. Do not attempt either; do not fabricate a namespace derivation you cannot verify. Instead, reflect on the session and hand off a harvest that a session WITH store access will dedup, trim, and ingest later via /ingest-harvest.

The capture bar (same bar as local closeout — restated, not skipped)

The store compounds only if what goes in is true, general, and non-redundant. A store that only grows becomes a store that lies. Capturing nothing is a perfectly good outcome; capturing five mediocre items is a bad one, because retrieved-wrong costs more than retrieved-nothing. Because this harvest will be ingested without your having seen what the store already contains, err toward a smaller, higher-bar list than you would in a local closeout — the ingesting agent is your only backstop against redundancy, not a rubber stamp.

An item earns a row only if all of these hold:

1. A future session facing a different but similar task would act differently because of it.
2. It is not already discoverable in the repo (README, CLAUDE.md/AGENTS.md, docstrings). Don't mirror documentation into memory.
3. It is not a one-off — not a typo, a transient network failure, or a now-fixed bug in code you already corrected.
4. Getting it wrong again would cost real time.

Good: "vec0 KNN is namespace-blind — filtering after a fixed k silently starves same-namespace neighbours; escalate k until a below-threshold row appears." Bad: "Fixed the consolidate bug." (narrative, not reusable) Bad: "Use pytest for tests." (already in the repo docs)

Write each item's content as an actionable claim, not a story. Include the trigger condition ("when X, do Y, because Z") so recall can match a future situation.

Aim for 0–5 items. If you have more than five, you are probably capturing narrative or duplicating docs — re-apply the bar.

Signal honesty is load-bearing — signal sets confidence, confidence gates recall. Never inflate:

Namespace rule

You cannot run host.py:resolve_namespace here (no store checkout to import from, and even if the code is visible, git remote resolution needs to match what the ingesting box would derive). Default to user:global for anything portable. Only propose a project:github.com/<org>/<repo> namespace when you can verify it from a real github.com remote URL:

git remote get-url origin

• If that prints https://github.com/<org>/<repo>.git, git@github.com:<org>/<repo>.git, or an equivalent GitHub URL, use project:github.com/<org>/<repo> (lowercase).
• Cloud / CCR sessions see a loopback proxy remote instead, of the form http://local_proxy@127.0.0.1:<port>/git/<org>/<repo> — this is NOT a real GitHub host, it is a local port-forwarded proxy standing in for one. If you see this shape, you may derive <org>/<repo> from the path segments after git/ and propose project:github.com/<org>/<repo> — but you MUST say so explicitly in that item's why field (e.g. "derived org/repo from the CCR loopback proxy remote path, not a verified github.com URL"), so the ingesting agent can decide whether to trust it.
• If the remote is anything else, unreadable, or you are not confident in the derivation, fall back to user:global and say why in why.

Never hand-write a project: namespace from a guessed folder name — an unverified namespace writes somewhere nothing ever queries.

Output format

Emit results as a single fenced JSON array, one object per captured item, with exactly these keys (no extras, no omissions):

[
  {
    "namespace": "user:global",
    "type": "lesson",
    "content": "Specific, actionable claim including the trigger condition.",
    "tags": "comma,separated,tags",
    "signal": "test",
    "why": "One sentence: which signal, and what concretely verified it."
  }
]

Field rules:

• namespace — user:global or a verified project:github.com/<org>/<repo> (see above).
• type — exactly one of fact, lesson, convention, preference.
• content — the actionable claim itself, not narrative.
• tags — a comma-separated string (not a JSON array).
• signal — exactly one of test, compile, lint, reviewer, user, none.
• why — one sentence of grounding evidence: why this signal was chosen and what concretely verified it (e.g. "pytest run passed after the fix", "user explicitly stated this preference in chat", "no tool ran; this is my own inference").

If nothing clears the bar, emit an empty array [] — do not omit the fenced block, and do not pad it with narrative to look productive.

Never

• Never include secrets, credentials, tokens, API keys, or PII in content — the store is local plaintext and the write-time scanner is advisory only. If a lesson is only meaningful with a secret embedded (e.g. "the deploy key starts with X"), drop the item entirely rather than redact-and-keep.
• Never inflate a signal to make an item look more grounded than it is.
• Never capture in-trajectory narrative ("first tried X, then Y") — capture only the conclusion that would help next time.
• Never claim a namespace derivation you did not actually check.

Delivery

1. Paste the fenced JSON array directly in your reply, along with a short plain-language report: what you captured (or didn't, and why not).
2. If this session has a working branch, also write the same JSON to zmem-harvest/<short-session-id>.json in the repo (create the zmem-harvest/ directory if absent) and include that file in the branch (stage/commit per the session's normal workflow) so the harvest survives after the conversation ends. Use a short, stable session identifier for <short-session-id> (e.g. the first 8 characters of the session/task id you have available) — this filename is only a delivery handle, not a namespace or an identity the store reasons about.
3. Do not attempt to write to any zmem store, and do not invoke store.py from this skill — that is the ingesting agent's job, via /ingest-harvest, applying its own recall-before-write dedup pass against the real store.

Report

State plainly, same as local closeout:

• Items captured, with signal and namespace for each (or "none").
• What you deliberately did not capture, and why — this is the evidence the bar was actually applied, not skipped.
• Where the harvest was delivered (pasted only, or also written to zmem-harvest/<short-session-id>.json).