magpie-setup-privacy-llm

setup-privacy-llm

Decide which LLMs this project's skills may send private foundation content to, record that decision where the skills look for it, and prove it holds.

Two mechanisms, and this skill configures one of them:

1. The approved-LLM gate — applies to <private-list> mail and any other private foundation list. A skill refuses to fetch unless every LLM in the active stack is approved. This is what the skill configures, by picking a variant.
2. PII redaction — applies to <security-list> mail. Third parties the reporter names are replaced with hash-prefixed identifiers before any LLM sees them. It runs under every variant and needs no per-variant configuration; this skill only verifies it works.

The recipes are docs/setup/privacy-llm.md; the contract behind them is tools/privacy-llm/tool.md. This skill walks the recipe rather than restating it — read the variant you land on before writing anything.

[!IMPORTANT] This is a project decision, not a personal preference. What goes in the committed copy says what the project permits for everyone's sessions. Configure locally for yourself freely; see Step 5 before promoting anything to the committed half.

---

Step 1 — Does it already resolve?

Resolve privacy-llm.md through the lookup chain, per file, local first:

1. .apache-magpie-local/privacy-llm.md — gitignored, yours;
2. .apache-magpie-overrides/privacy-llm.md — committed, the project's.

Both present → say so and name which one wins here (the local one does), because a local file silently shadowing the project's answer is exactly the case an operator needs told. Offer to show the difference, then go to Step 4 and verify rather than rewrite.

Only one present → read it, report the stack it declares, and go to Step 4. Do not rewrite a working configuration because this skill happened to be invoked.

Neither → continue to Step 2. This is the first run.

Step 2 — Detect the stack, do not ask for it

Ask the user nothing you can determine. Establish what is actually in the stack:

• The agent itself — which harness is running this skill (Claude Code, Codex, Gemini CLI, …). Always in the stack.
• Local inference — is an Ollama, llama.cpp or vLLM endpoint reachable and configured for this project? Check the adapter configuration and the environment, not a guess.
• Third-party endpoints — anything in the project's existing configuration naming a model endpoint the framework does not ship.
• Whether this project has private lists at all. A project with no <private-list> still wants the file, because the redactor and the gate both read it — but the variant is the simple one and the interview is one question shorter.

Report what you found as a list before proposing anything. A detection the user can see is a detection they can correct.

Step 3 — Pick the variant and write it

Map what Step 2 found onto a variant in docs/setup/privacy-llm.md:

What the stack is	Variant
The agent and nothing else	1 — agent only (the default)
A local Ollama model beside it	2 — local inference (Ollama)
A local vLLM endpoint beside it	3 — local inference (vLLM)
An Apache-hosted endpoint	4 — Apache-hosted endpoint
AWS Bedrock	5 — Bedrock
The Anthropic API directly	6 — direct API (opt-in)

Propose the variant your detection implies, say what it means in one sentence — this permits the private list to be read by X and nothing else — and let the user correct it.

Then write .apache-magpie-local/privacy-llm.md from that variant's block, substituting the project's real list addresses. Local, always, on this step. Writing the gitignored copy needs nobody's permission, is invisible to every other clone, and is undone by deleting a directory.

Two things this step may not do: invent an endpoint — if the detection is ambiguous, ask rather than assume — and widen the stack to make the gate pass. The gate failing is information.

Step 4 — Prove it, do not declare it

Run all three. A privacy configuration nobody has exercised is a comment.

# 1. The gate: is every LLM in the active stack approved?
uv run --project <framework>/tools/privacy-llm/checker \
  privacy-llm-check --reads-private-list

# 2. The redactor, end to end. "Other Researcher" is a third party
#    the reporter named — never the reporter themselves.
echo "I worked with Other Researcher (other@example.com) on this" | \
  uv run --project <framework>/tools/privacy-llm/redactor \
  pii-redact --field name:"Other Researcher" \
             --field email:"other@example.com"

# 3. The resulting local map.
uv run --project <framework>/tools/privacy-llm/redactor pii-list

Expected: exit 0 from the gate; the two values in step 2 replaced by N-… and E-… identifiers; step 3 listing them.

A non-zero gate is the finding, not a failure of this skill. Report which member of the stack is unapproved and what the options are — drop it from the stack, or take the opt-in recipe for it — and stop. Do not edit the file until the user picks.

Step 5 — Say that adopting exists, then stop

Everything so far is local and yours. If the project should commit this answer so every contributor's session is bound by it, that is adopt, which promotes .apache-magpie-local/ into the committed .apache-magpie-overrides/ and stages it for review.

Say that in one line and name the command. Do not run it, do not offer to run it, and do not repeat it on later invocations. Deciding what the project permits its agents to do with private foundation mail is a maintainer decision taken with the other maintainers — and unlike the rest of this skill, it is the one part that is not undone by deleting a directory.

Step 6 — Re-run after an upgrade

What counts as default-approved can narrow between framework versions. After /magpie-setup upgrade, re-run Step 4: if an entry that used to pass is now opt-in, the gate surfaces it and the user takes the matching variant's opt-in recipe.

Say this once, at the end of a first run. It is the reason this skill is worth invoking a second time.

---

Status — provisional pending ASF Legal

The registry this skill checks against is provisional: it reflects the framework maintainers' working position in the absence of a ratified ASF Legal Affairs policy for AI-assisted handling of foundation private data. Say so when a user asks whether a variant is "allowed" — the honest answer is that the framework enforces a list nobody has yet ratified, and docs/setup/privacy-llm.md carries the full caveat.