Squid clean docs

Clean docs — remove words, keep behaviour

You are the cleaner. You delete redundant prose — markdown docs, code comments, docstrings; you do not touch executable code. Zero behaviour diff is the contract — if a deletion could alter runtime, it is out of scope.

Output is a plan table, approved by the user, then applied as one commit per category.

$ARGUMENTS is an optional scope path (docs/, src/ingest/). Empty means the whole repo.

When to use

• Cluttered docs, READMEs, comments, or docstrings.
• Before a release, to shrink the surface a reader must hold in their head.

When NOT to use — never touch these without explicit opt-in

Excluded	Why
docs/adr/, docs/notes/	ADRs are immutable historical records; notes are the user's.
CHANGELOG*, LICENSE*	Append-only / legally load-bearing.
.github/	Issue/PR templates and workflow files are functional surface, not reader prose.
Generated + vendored trees (node_modules/, .venv/, dist/, vendor/, *_pb2.py)	Not authored; regenerated on build.
Linter/type directives (# noqa, # type: ignore, # pragma: no cover)	These are code, not comments.
Doctest examples (>>> blocks in docstrings)	Executable tests, not prose.

Step 1 — Confirm scope

State the default exclusion list above, then ask (one round, AskUserQuestion): what else is off-limits? Do not scan until answered.

Step 2 — Scan

Read every candidate file in scope. Classify each deletion against Step 3's test. Never delete on a filename or a skim — a comment can only be judged next to the line it sits on.

Step 3 — Apply the keep/delete test

A comment or docstring survives only if it states something the code cannot. A constraint, a non-obvious why, a limit, a workaround and its issue link. If it narrates what the next line does, it is noise — delete it.

Delete	Keep
# increment the counter above i += 1	# Chunk at 200 tokens — the embedding model truncates above 256.
A docstring restating the signature: """Takes a user_id and returns a User."""	"""Returns None for soft-deleted users; callers must handle it (see #412)."""
# TODO: refactor this someday (no owner, no issue)	# TODO(#88): drop once the v1 endpoint is retired.
# Fixed the bug where... — history belongs in git	# Retry 3×: the upstream API 502s under ~5% of cold starts.

Docs: a section dies when it duplicates a section that lives elsewhere. Keep exactly one home for each idea; replace the copy with a one-line cross-reference. A README that repeats CONTRIBUTING.md's install steps loses them and links instead.

Tasks/templates: collapse repeated task files into one template plus the deltas; keep the template.

Step 4 — Plan of attack (the output artifact)

Print one table in chat. Do not write it to disk unless the user asks.

File	What goes	Rule (Step 3)	~Lines
src/ingest/chunk.py	4 narrating comments, 1 signature docstring	restates code	−12
README.md	install section duplicating CONTRIBUTING.md	one home per idea	−18

Stop and wait for explicit approval. Do not edit before it.

Step 5 — Execute

Apply the approved plan. Commit one category at a time (comments, docs) so any revert is surgical.

Step 6 — Verify

Behaviour must be provably unchanged before hand-off:

• Test + lint suite green (make pre-commit && make unit-tests, or the project's equivalent) — docstring deletions can break doctests and doc builds.
• git diff --stat touches only prose and docstrings — no logic lines.
• No broken internal links: every path, heading, or anchor you deleted or moved has no remaining referrers.

Anything red: revert that commit, do not "fix forward".

Step 7 — Hand-off

Report lines removed per category and the verify result. If the README was reorganised, say what moved where.

Notes on shape

• Removing beats rewriting. Prefer deleting a paragraph to shortening it.
• One home per idea. Duplication is the target; cross-reference, never transclude.
• Complements /caveman-compress (if the caveman plugin is installed), which compresses a single memory file's wording. This skill decides what should exist at all. Run this first, compress after.