Wiki: lock scope to user-centric only; CLAUDE.md cleanup-script note

[sysread]

=head1 SYNOPSIS

Two changes batched in one PR:

• Lock the wiki's scope to user-centric subjects only. The
  per-conversation agent and the librarian both now refuse to
  create / accept articles about generic world-knowledge topics
  that merely came up in conversation. The librarian deletes
  any that exist on its next sweep.
• CLAUDE.md doc tweak: drop the delete_ref 403 explanation in
  favour of just "leave the remote feature branch alone; the
  user runs a cleanup script."

=head1 PURPOSE

After the wiki shipped, the per-conversation agent started
writing standalone articles about generic topics that came up in
conversation. The concrete case from production: a brainstorm
about app naming mentioned that the app's name was inspired by
the 1980s "Kermit" file-transfer protocol; the agent then
created a standalone "Kermit protocol" article. The wiki was
meant to be ABOUT the user - their projects, people in their
life, things they're learning, their work - not a general
encyclopedia of topics that came up.

The earlier prompt language ("a project, a person in their life,
a place, an interest, a recurring situation") didn't draw the
line firmly enough; "an interest" was reading as "any topic the
user was interested in for the purposes of one conversation".

=head1 DESCRIPTION

=head2 Layer 1: how the existing wiki agents framed scope

The autonomous agent's prompt named some kinds of things
articles are about (project / person / place / interest /
situation) but did not specify what is NOT about the user, and
did not explicitly forbid creating standalone articles for
external topics that came up in conversation. The librarian's
prompt was scoped only to consolidate duplicates, fact-check,
and tighten boundaries - it had no notion of "this article
shouldn't exist at all".

=head2 Layer 2: what this PR changes, parallel to layer 1

WIKI_AUTONOMOUS_PROMPT now carries an explicit Scope block
between the wiki overview and the workflow steps:

• IN scope (article-worthy when discussed): projects, people in
  their life, places they care about, things they're learning or
  reading, habits and experiments they're tracking, their
  career, hobbies, and the user themselves.
• OUT of scope (do NOT create articles for these): generic
  technical concepts, world-knowledge topics, public people the
  user does not know personally, news, tutorials, debug
  sessions, generic Q&A.

When an OUT-of-scope reference shows up inside a user-centric
article (e.g. the conversation noted that the app being built is
named after a 1980s file-transfer protocol), the agent is told
to add a Markdown link inside the user-centric article rather
than creating a separate page. The Kermit case is baked in as
the concrete example. It also explicitly cautions against
fabricating URLs.

buildWikiLibrarianPrompt gains a matching scope block AND an
expanded workflow step 1: "Scan for out-of-scope articles
first." The librarian must wiki_search the full body before
deleting (some titles look generic but are about a specific
project), and if a related user-centric article exists,
optionally edit a Markdown link into it before deleting the
out-of-scope article. Critically, this scope-cleanup pass runs
BEFORE the duplicate-consolidation pass so the librarian doesn't
tidy two off-topic articles into one off-topic article.

wiki_delete's description in the librarian prompt now lists
two legitimate cases: (a) consolidation as before, plus (b)
out-of-scope cleanup with no merge required.

WIKI_MANUAL_PROMPT (the per-article "ask agent to update"
flow) gets a brief scope note - the user is in control here so
the rule is softer, but it still tells the agent to prefer a
noop with a one-sentence reason over silently drifting an
article away from being about the user (e.g. into a generic
explainer of an external topic).

WIKI_BLOCK in chat-prompt.ts updates the main LLM's mental
model of the wiki to match: it now describes the wiki as
articles ABOUT THE USER and explicitly "not a general
encyclopedia of topics that came up".

The docs/user/wiki.md user-facing manual gains a "Scope:
about you, not about the world" section that mirrors the
prompt's IN/OUT lists in plain language. docs/dev/wiki.md
adds a Gotcha that records the failure mode and the rationale
behind workflow ordering on the librarian.

The CLAUDE.md edit (already pushed in commit 1) just drops the
delete_ref 403 explanation - the user has a cleanup script.

=head2 Layer 3: how that resolves PURPOSE

The per-conversation agent now has an explicit IN/OUT list and
a concrete example (Kermit) baked into the prompt, making the
"is this about the user?" decision much harder to slip on. The
chat assistant's WIKI_BLOCK matches the same scope so it
doesn't over-eagerly suggest articles for things outside that
scope. And critically, the librarian's first workflow pass
deletes any out-of-scope articles that already exist, so the
existing wiki should self-heal over the next 12-24 hours
without manual intervention.

=head1 Notes for AI reviewers

• The librarian's "delete on out-of-scope" path is intentional
  and reviewed. wiki_delete's contract has been deliberately
  widened from "consolidation only" to also include "out-of-
  scope cleanup with no merge required". If you suggest
  reverting this to consolidation-only, you'd leave the
  existing off-topic articles in place.
• The librarian still has no wiki_create. Out-of-scope
  cleanup uses delete-only; it cannot mint replacement
  articles. That's by design.
• "External topics get linked, not given their own pages" is
  the load-bearing rule. A reviewer suggesting "but the
  conversation went deep on this" is exactly the failure mode
  the prompt now defends against.
• Workflow ordering in the librarian (scope cleanup first,
  duplicates second) is deliberate. Swapping them would
  consolidate two off-topic articles into one.

https://claude.ai/code/session_015XcR7xzLdij66ZbYERUdLH

---

Generated by Claude Code