Overview
A census of the llms.txt layer (2026-07-31) ahead of the assistant paper going live found the generated catalogues (llms-full.txt / workspace_index.json) byte-identical and all navigator CI checks green, but the hand-curated routing layer has drifted behind recent workspace work: multi_galaxy/ is unrouted, the weak/ entry makes a false claim, a cross-repo link is dead, and the PyAutoHands GROUP_ORDER still lists the pre-rename multi folder. One approved performance change: scope llms-full.txt (168KB, ~30–40k tokens, grows linearly with script count) to non-chat harnesses so connector chats don't ingest it.
Plan
- Add
multi_galaxy/ routing to autolens_workspace/llms.txt and fix its stale weak/ entry (a start_here.py exists now).
- Fix
GROUP_ORDER in PyAutoHands' navigator (multi → multi_dataset, add multi_galaxy, ellipse) and regenerate both workspace catalogues.
- Repoint autocti_assistant's dead link to the non-existent autocti_workspace navigator.
- Signpost the autolens_assistant from
PyAutoLens/llms.txt.
- Complete the
scripts/README.md folder lists in both workspaces (autolens missing weak; autogalaxy missing multi_galaxy + cluster) and fix the shared "inculding" typo.
- Scope the
llms-full.txt fetch instruction to non-chat harnesses in autolens_assistant/AGENTS.md and refresh the stale "(added in a later phase)" tail lines in both workspace llms.txt files.
Merge order: PyAutoHands first, then the workspaces — the workspaces' navigator_check CI regenerates against PyAutoHands main (same-named branches keep PR CI green in the meantime).
Detailed implementation plan
Work Classification
Workspace (docs/routing files; one tooling constant in PyAutoHands; no library API changes)
Affected Repositories
- autolens_workspace (primary)
- autogalaxy_workspace
- PyAutoHands
- autocti_assistant
- PyAutoLens
- autolens_assistant
Branch Survey
| Repository |
Current Branch |
Dirty? |
| ./autolens_workspace |
main |
clean (claimed by interferometer-start-here-integrate-oom, disjoint files) |
| ./autogalaxy_workspace |
main |
clean |
| ./PyAutoHands |
main |
clean |
| ./autocti_assistant |
main |
clean |
| ./PyAutoLens |
main |
clean (claimed by 2 active tasks, disjoint files — llms.txt only) |
| ./autolens_assistant |
main |
1 uncommitted user edit in README.md (untouched by this task) |
Suggested branch: feature/llms-txt-census-fixes
Worktree root: ~/Code/PyAutoLabs-wt/llms-txt-census-fixes/
Implementation Steps
- PyAutoHands —
autohands/navigator.py: in GROUP_ORDER, replace "multi" with "multi_dataset", add "multi_galaxy" and "ellipse" (after the dataset-type groups, before "guides"). PR merges FIRST.
- autolens_workspace —
llms.txt:
- "Start here": add
scripts/multi_galaxy/start_here.py entry (multi-galaxy lenses — 2+ co-dominant deflectors, no host halo).
- "I want to…": add
scripts/multi_galaxy/modeling.py entry.
- Fix weak entry: route to
scripts/weak/start_here.py (it exists now; drop the "no start_here.py" claim), keep fit.py/simulator.py as companions.
- Tail line: drop "(added in a later phase)"; add "large file — local agents only; in a browser chat use this file's routing instead" caveat.
- autolens_workspace —
scripts/README.md: add weak to the folder list; fix "inculding" → "including".
- autolens_workspace — regenerate catalogue with the fixed GROUP_ORDER (
regenerate_navigator.py autolens with sibling PyAutoHands on the same-named branch) so llms-full.txt/workspace_index.json pick up the new group order. No notebook regeneration (no script changes).
- autogalaxy_workspace —
llms.txt tail line (same refresh + caveat as step 2); scripts/README.md: add multi_galaxy + cluster, fix typo; regenerate catalogue (regenerate_navigator.py autogalaxy).
- autocti_assistant —
llms.txt: repoint the "autocti_workspace navigator" link to https://github.com/PyAutoLabs/autocti_workspace with wording that doesn't promise an llms.txt routing file.
- PyAutoLens —
llms.txt: add one autolens_assistant line under "Use it".
- autolens_assistant —
AGENTS.md "Working with skills" step 3: qualify the "check the autolens_workspace catalogue (llms-full.txt)" instruction — local/coding agents read llms-full.txt; connector chats use the workspace llms.txt routing + targeted single-file reads (the full catalogue is ~30–40k tokens and slows every subsequent turn).
Key Files
PyAutoHands/autohands/navigator.py — GROUP_ORDER constant
autolens_workspace/llms.txt, scripts/README.md, llms-full.txt (generated), workspace_index.json (generated)
autogalaxy_workspace/llms.txt, scripts/README.md, llms-full.txt (generated), workspace_index.json (generated)
autocti_assistant/llms.txt
PyAutoLens/llms.txt
autolens_assistant/AGENTS.md
Acceptance
check_navigator.py path + banner checks green in both workspaces after edits.
regenerate_navigator.py zero-diff against the committed catalogues (run with patched PyAutoHands).
- Generated files change only via the generator; notebooks untouched.
- Workspace
url_check CI green (all URLs in edited files resolve).
Out of scope (deliberately)
- autocti_workspace navigator layer (future task; today's fix is just the dead link).
- euclid_assistant llms.txt (possibly deliberate absence — collaboration-internal).
- autogalaxy_assistant birth checklist (workspace INTERIM science section, capability-boundary assistant link, PyAutoGalaxy/llms.txt signpost) — lands with that task.
Notes
- Brain sized this too-large (25) with a generic 4-phase split; overridden — the score is prose/repo-count driven, every edit is a few lines of docs with zero API risk. One task, per-repo PRs, PyAutoHands-first merge order.
Original Prompt
Click to expand starting prompt
See PyAutoMind/active/llms_txt_census_fixes.md — census findings (2026-07-31), verified green baseline, full scope. Original user request (verbatim) is embedded in the prompt file.
The llms.txt file is used by the autolens_assistant to support conversation AI like ChatGPT. Can you do a census of this file against all the workspace updates we have done recently, and make sure it is up to date and fit for purpose. In particular, also make sure its paired to the autolens_assistant correctly (e.g. they are in sync). Soon, I will make the autogalaxy_assistant, therefore also make sure the llms.txt file in autogalaxy_workspace and other repos is ok. These may include making sure README.md files with contents are up to date, or any other API / docs / structure checks. Can you also do an assessment of whether the llms.txt file structure and general design might be leading to some sort of slow down?
Overview
A census of the
llms.txtlayer (2026-07-31) ahead of the assistant paper going live found the generated catalogues (llms-full.txt/workspace_index.json) byte-identical and all navigator CI checks green, but the hand-curated routing layer has drifted behind recent workspace work:multi_galaxy/is unrouted, theweak/entry makes a false claim, a cross-repo link is dead, and the PyAutoHandsGROUP_ORDERstill lists the pre-renamemultifolder. One approved performance change: scopellms-full.txt(168KB, ~30–40k tokens, grows linearly with script count) to non-chat harnesses so connector chats don't ingest it.Plan
multi_galaxy/routing toautolens_workspace/llms.txtand fix its staleweak/entry (astart_here.pyexists now).GROUP_ORDERin PyAutoHands' navigator (multi→multi_dataset, addmulti_galaxy,ellipse) and regenerate both workspace catalogues.PyAutoLens/llms.txt.scripts/README.mdfolder lists in both workspaces (autolens missingweak; autogalaxy missingmulti_galaxy+cluster) and fix the shared "inculding" typo.llms-full.txtfetch instruction to non-chat harnesses inautolens_assistant/AGENTS.mdand refresh the stale "(added in a later phase)" tail lines in both workspace llms.txt files.Merge order: PyAutoHands first, then the workspaces — the workspaces'
navigator_checkCI regenerates against PyAutoHandsmain(same-named branches keep PR CI green in the meantime).Detailed implementation plan
Work Classification
Workspace (docs/routing files; one tooling constant in PyAutoHands; no library API changes)
Affected Repositories
Branch Survey
Suggested branch:
feature/llms-txt-census-fixesWorktree root:
~/Code/PyAutoLabs-wt/llms-txt-census-fixes/Implementation Steps
autohands/navigator.py: inGROUP_ORDER, replace"multi"with"multi_dataset", add"multi_galaxy"and"ellipse"(after the dataset-type groups, before"guides"). PR merges FIRST.llms.txt:scripts/multi_galaxy/start_here.pyentry (multi-galaxy lenses — 2+ co-dominant deflectors, no host halo).scripts/multi_galaxy/modeling.pyentry.scripts/weak/start_here.py(it exists now; drop the "no start_here.py" claim), keepfit.py/simulator.pyas companions.scripts/README.md: addweakto the folder list; fix "inculding" → "including".regenerate_navigator.py autolenswith sibling PyAutoHands on the same-named branch) sollms-full.txt/workspace_index.jsonpick up the new group order. No notebook regeneration (no script changes).llms.txttail line (same refresh + caveat as step 2);scripts/README.md: addmulti_galaxy+cluster, fix typo; regenerate catalogue (regenerate_navigator.py autogalaxy).llms.txt: repoint the "autocti_workspace navigator" link tohttps://github.com/PyAutoLabs/autocti_workspacewith wording that doesn't promise an llms.txt routing file.llms.txt: add one autolens_assistant line under "Use it".AGENTS.md"Working with skills" step 3: qualify the "check theautolens_workspacecatalogue (llms-full.txt)" instruction — local/coding agents readllms-full.txt; connector chats use the workspacellms.txtrouting + targeted single-file reads (the full catalogue is ~30–40k tokens and slows every subsequent turn).Key Files
PyAutoHands/autohands/navigator.py—GROUP_ORDERconstantautolens_workspace/llms.txt,scripts/README.md,llms-full.txt(generated),workspace_index.json(generated)autogalaxy_workspace/llms.txt,scripts/README.md,llms-full.txt(generated),workspace_index.json(generated)autocti_assistant/llms.txtPyAutoLens/llms.txtautolens_assistant/AGENTS.mdAcceptance
check_navigator.pypath + banner checks green in both workspaces after edits.regenerate_navigator.pyzero-diff against the committed catalogues (run with patched PyAutoHands).url_checkCI green (all URLs in edited files resolve).Out of scope (deliberately)
Notes
Original Prompt
Click to expand starting prompt
See
PyAutoMind/active/llms_txt_census_fixes.md— census findings (2026-07-31), verified green baseline, full scope. Original user request (verbatim) is embedded in the prompt file.