Work Classification

Infrastructure (organism repos) — no library or workspace code is touched.

Affected Repositories

• PyAutoBrain (primary) — the alias table, the derivation, the guards
• PyAutoMind — repos.yaml gains the package: identity key

Branch Survey

Branch: claude/organ-repo-spelling-splits-hbbms0 (session-designated; this
is a web-github session with no task worktrees, so the usual
feature/<task-name> name is overridden). worktree_check_conflict organ-repo-spelling-splits PyAutoBrain PyAutoMind returns 0 — no claim conflict.

Worktree root: n/a (web-github — the session's clones are the checkout).

Implementation Steps

1. PyAutoMind/repos.yaml — add package: to the seven repos that ship
   one: PyAutoFit: autofit, PyAutoArray: autoarray, PyAutoGalaxy: autogalaxy,
   PyAutoLens: autolens, PyAutoCTI: autocti, PyAutoReduce: autoreduce,
   PyAutoNerves: autonerves. Document the key in the file header next to
   public_role / front_door: identity only — the import/PyPI name the repo
   ships, absent where it ships none. Every value is cross-checked in step 5
   against the test_witness rows fix: witness map missed 5 tested repos and split 2 more across keys #271 verified by reading each repo's tree.

2. PyAutoBrain/agents/faculties/sizing/_sizing.py — derive the join:

   • _body_map_specs() alongside the existing _body_map_categories(), so the
     package: field is readable without a second YAML load.
   • _canonical_key(name, spec) = spec["package"].lower() when present, else
     name.lower(). This is the fix: witness map missed 5 tested repos and split 2 more across keys #271 rule, now executable rather than prose.
   • _derived_aliases() — for every repo in a sizing category, map each
     registered spelling (name.lower(), the pyauto-stripped bare form, the
     package name, and the dot-truncated head for a dotted repo name) onto its
     canonical key.
   • REPO_ALIASES = _derived_aliases() | policy()["repo_aliases"], with a
     conflict check: a hand entry that disagrees with the derivation is drift
     and must raise, not silently win. Move the REPO_ALIASES assignment above
     _target_sets() so normalise_repo is usable during set construction.
   • Leave LIBRARY_REPOS / WORKSPACE_REPOS / ORGANISM_REPOS registering
     every spelling. Membership is always tested post-normalisation, and
     _intake._repos_in scans the set as free-text tokens — narrowing the sets
     to canonical keys alone would stop bare mentions being recognised in
     unheadered prose.

3. PyAutoBrain/config/policy.yaml — trim repo_aliases to the
   non-derivable rows (aa/af/ag/al; autoconf/pyautoconf and
   pyautobuild/autobuild renames) and head the block with the canonical-key
   rule plus a pointer saying the bare/prefixed join is derived, not typed here.
   The library rows and autohands: pyautohands become redundant and go.

4. PyAutoBrain/tests/test_policy_seams.py — three guards, every name
   derived from the body map (the tenant firewall allows only
   {PyAutoFit, PyAutoLens, autolens_workspace} as literals in this file):

   • no split spellings — for every body-map repo in a sizing category,
     all registered spellings normalise to one key, and that key is itself a
     known target. Fails today for five organs and PyAutoScientist.
   • no alias points at a dead key — every value in the effective alias
     table is the canonical key of some body-map repo. This is the
     pyautoconf -> autoconf defect (fix: retire PyAutoConf rename leftovers in Brain functional surfaces #267) killed at source.
   • package field agrees with the witness map — for every test_witness
     row of the form <Repo>/test_<pkg>, repos.yaml[<Repo>].package == pkg,
     and vice versa. Makes Mind identity and Brain policy check each other.

5. Mutation-test the guards (the fix: witness map missed 5 tested repos and split 2 more across keys #271 practice): drop one derived organ
   alias, mis-key one package: value, and restore the pre-fix policy.yaml
   — each must fail a named guard. Record which guard caught what.

6. Validate: pytest PyAutoBrain/tests (468 baseline + new),
   python3 PyAutoMind/scripts/repos_sync.py --check (tenant firewall + the 11
   identity checks), and re-run the reproduction — @autobrain, @autoheart,
   @automind, @automemory, @autogut, @autoscientist must each reach the
   prefixed key and its witness row, and the six library spellings must be
   byte-identical to their pre-change resolution.

Key Files

• PyAutoMind/repos.yaml — the body map; gains package: on seven repos
• PyAutoBrain/agents/faculties/sizing/_sizing.py — _target_sets, normalise_repo, REPO_ALIASES
• PyAutoBrain/config/policy.yaml — repo_aliases, test_witness, sizing_categories
• PyAutoBrain/tests/test_policy_seams.py — the coverage guards
• PyAutoBrain/agents/conductors/refactor/_refactor.py — behaviour_preservation, the surface where the defect shows
• PyAutoBrain/agents/conductors/intake/_intake.py — _repos_in / infer_target, the second surface

Trade-offs

• package: in the body map vs a Brain-only declaration. A
  bare_key_categories: [library] + PyAutoNerves-exception block in
  policy.yaml would keep this a one-repo change, but the exception list is
  hand-maintained again — the exact cause this task exists to remove — and the
  body map still could not answer what package a repo ships. The package name
  is repo identity, so it belongs in repos.yaml; the two-repo cost is
  accepted deliberately.
• Sets keep every spelling. Narrowing them to canonical keys would be
  tidier but silently breaks _intake._repos_in's free-text scan. Not worth it.
• Heart is not evaluable here. pyauto-heart is unreachable from a
  web-github session, so leg 4 of the ship gate will be recorded as NOT
  EVALUATED rather than claimed clean.

Testing approach

Guard-first: each new guard is written against the current main state, watched
to fail for the right reason, and only then made to pass by the derivation. The
library spellings serve as the behaviour-preservation witness — their
resolution must not move.

Organ repo spellings split across two normalised keys

Type: bug
Target: pyautobrain
Repos:

• PyAutoBrain
  Difficulty: large
  Autonomy: supervised
  Priority: normal
  Status: formalised
  Filed: 2026-08-24

Found while shipping the refactor witness-map audit (@PyAutoBrain#269 / PR #271).

The defect

The bare spellings of the organ repos do not normalise to the same key as their
repo spellings, so a policy map keyed on one is invisible to the other.

@PyAutoBrain/agents/faculties/sizing/_sizing.py (_target_sets.names_for)
registers BOTH spellings of every PyAuto* repo as known targets
(name.lower() and name.lower()[2:]), but the repo_aliases table in
@PyAutoBrain/config/policy.yaml only joins them for the libraries:

Same for the bare Heart, Memory, Mind and Gut spellings. @autohands was joined
in PR #271 because extra_organism_targets declares it explicitly and it
reaches real code today; the rest are latent — no prompt in the Mind history
uses them yet.

Decide before fixing

This is NOT a blind alias sweep. Should organs key on the bare form
(autobrain) or the prefixed form (pyautobrain)? target_signals keys the
prefixed form (pyautobrain, pyautomind, pyautoheart, pyautobuild), and
after #271 test_witness keys organs prefixed but libraries bare. Whichever way
it goes has to hold across both maps at once, so answer the question first and
let the edit follow.

The wider question this raises

Three separate repos have now hit this same defect class — PyAutoNerves
(#267), PyAutoCTI and PyAutoReduce (#269). The shared cause is that
repo_aliases is HAND-MAINTAINED while the known-target set is DERIVED from the
body map, so the two drift silently and the gap only ever surfaces as a
wrong-but-plausible conductor message ("strengthen tests first" for a
well-tested repo).

Guard 2 added in #271 — a witness-map key must be what its repo normalises to —
closes the witness-map instance only. An alias gap in a map with no coverage
guard is still invisible. Consider deriving repo_aliases from the body map so
the class closes rather than its fourth instance.

Notes on this prompt's header

pyauto-brain intake classify scored this bug at high confidence (kept), but
proposed Target: autocti — it read the prior-instance history above as the
subject — and Difficulty: too-large (score 10), inflated by the repo names and
the design-decision keywords. Corrected to pyautobrain / large on review.
An alternative classification is refactor: the fix is behaviour-preserving for
every spelling that resolves today. bug was kept because the Bug Agent's
investigate-first strategy matches the "decide before fixing" requirement above.

Acceptance: every spelling of every body-map repo that _target_sets registers
as a known target resolves to a key the policy maps are actually filed under, or
is deliberately not registered; a guard makes a future alias gap fail loudly
rather than degrade a conductor's advice.