|
| 1 | +Repaired the `aplt.Output` stale-API drift in the last two repos the |
| 2 | +2026-08-04 `plot-array-stale-kwargs` task deliberately stopped short of — and |
| 3 | +found that the drift's real shape was not what the prompt described. |
| 4 | + |
| 5 | +## Outcome |
| 6 | + |
| 7 | +| Repo | PR | Merged | |
| 8 | +|---|---|---| |
| 9 | +| PyAutoGalaxy | [#586](https://github.com/PyAutoLabs/PyAutoGalaxy/pull/586) | `d68a8f6` | |
| 10 | +| euclid_strong_lens_modeling_pipeline | [#39](https://github.com/PyAutoLabs/euclid_strong_lens_modeling_pipeline/pull/39) | `62b2fd4` | |
| 11 | + |
| 12 | +Classification changed during planning from workspace-only to **library + |
| 13 | +workspace**, which is why a PyAutoGalaxy PR exists at all. |
| 14 | + |
| 15 | +## The prompt was wrong in three ways, each found by verifying rather than trusting |
| 16 | + |
| 17 | +**1. `autocti_workspace_test` is out of scope — the hedge resolved, then the |
| 18 | +scope collapsed.** The prompt flagged 27 files as "UNVERIFIED — check first", |
| 19 | +suspecting PyAutoCTI might still export `Output`. It does not: |
| 20 | +`autocti/plot/__init__.py` is 49 lines of flat functions. So those files are |
| 21 | +genuinely broken — but an alias-aware scan found **31**, not 27, and: |
| 22 | + |
| 23 | +- **18 are under `legacy/`**, which `autocti_workspace_test/AGENTS.md:52` |
| 24 | + forbids editing ("preserved Euclid VIS history"). Its README says outright |
| 25 | + that they target the removed Plotter object stack and are not runnable. |
| 26 | +- **13 are top-level `imaging_ci/`** — undocumented in AGENTS.md's structure |
| 27 | + block, absent from `smoke_tests.txt`, substantive history **2023-02**, the |
| 28 | + same era as `legacy/` (created 2026-07-17 by "CTI resurrection Phase 5"). |
| 29 | + Its only 2026 commits are repo-wide mechanical sweeps that also hit `legacy/`. |
| 30 | + |
| 31 | +Following the prompt literally would have modernised dead code *and* violated |
| 32 | +the repo's own hard rule. Filed as |
| 33 | +`draft/maintenance/autocti_workspace_test/imaging_ci_heritage_sweep.md` for a |
| 34 | +human decision against the CTI resurrection epic (PyAutoCTI#82). |
| 35 | + |
| 36 | +**2. The drift was never really about `Output`.** The whole plotter-object stack |
| 37 | +is gone from every public namespace — `MatPlot2D`, `MatPlot1D`, `Cmap`, |
| 38 | +`Array2DPlotter`, `Visuals2D`, `MassProfileCentresScatter`, the `*Plotter` |
| 39 | +family. `Output` was 6 of 13 missing symbols in autocti, 2 of 6 in euclid. |
| 40 | + |
| 41 | +**3. Two library defects the prompt never mentioned.** `Cmap` is exported by |
| 42 | +**no** public plot namespace, yet: |
| 43 | + |
| 44 | +- `Scribbler.__init__` required a `Cmap`-shaped object for `cmap=`, so callers |
| 45 | + had no public way to colour the GUI. |
| 46 | +- `Clicker.start()` built `aplt.Cmap(...)` on `autoarray.plot` |
| 47 | + *unconditionally* — reproduced on the installed stack as |
| 48 | + `AttributeError: module 'autoarray.plot' has no attribute 'Cmap'`. That GUI |
| 49 | + was dead for every caller. |
| 50 | + |
| 51 | +Fixed library-side because the only workspace alternative was a private-path |
| 52 | +import — the "autoimmune reaction" the Bug Agent's fix-locus rule warns against. |
| 53 | + |
| 54 | +## The plan's worst assumption, caught before it shipped |
| 55 | + |
| 56 | +The approved plan said euclid's explicit `mask=` could simply drop out, since |
| 57 | +`plot_array` auto-derives the outline. It cannot. `data` there comes from |
| 58 | +`Array2D.from_fits` and is **unmasked**, so `auto_mask_edge(data)` returns |
| 59 | +`None` — verified directly. Worse, the mask being drawn is a *separately |
| 60 | +constructed* circular one whose radius is grown at line 160 to enclose the |
| 61 | +clicked galaxies, so it is the entire subject of the figure. |
| 62 | + |
| 63 | +A literal reading would have silently deleted the mask radius from the PNG that |
| 64 | +exists to show it, and nothing would have failed. Fixing it needed a second |
| 65 | +library change: autogalaxy's `plot_array` wrapper delegates to autoarray's but |
| 66 | +was dropping its `mask=` parameter. |
| 67 | + |
| 68 | +## Verification |
| 69 | + |
| 70 | +- PyAutoGalaxy CI green on **every leg**: `unittest (3.12)`, `unittest (3.13)`, |
| 71 | + `unittest-nojax`, Docs. 1129 tests, 16 new across two files (`test_autogalaxy/gui/` |
| 72 | + did not previously exist). |
| 73 | +- Alias-aware AST re-scan — aliases resolved from each file's own imports, the |
| 74 | + correction carried forward from HowToGalaxy#56 — **0 residual stale symbols** |
| 75 | + in both repos. |
| 76 | +- All 10 changed `aplt.*`/`al.*` calls bound against real signatures: 10 ok, 0 failures. |
| 77 | +- Mask passthrough verified by **rendering**: 2436 pixels differ vs. the same |
| 78 | + call without `mask=`, confirmed by pixel diff rather than file size. |
| 79 | +- `compileall` clean; CRLF preserved uniformly (see below). |
| 80 | + |
| 81 | +## Not verified, stated plainly |
| 82 | + |
| 83 | +- **The two euclid GUIs were never run end-to-end.** They need TkAgg and FITS |
| 84 | + data absent from the repo, and euclid has **no CI workflows at all**. The |
| 85 | + merge was made with that stated and acknowledged. |
| 86 | +- **`Scribbler.__init__` has no direct test** — `matplotlib.use("TkAgg")` and |
| 87 | + `wm_geometry` cannot run headless. The extracted `norm_from` helper carries |
| 88 | + the logic and is fully tested; the constructor is covered only by binding. |
| 89 | + |
| 90 | +## Behaviour change accepted |
| 91 | + |
| 92 | +Extra-galaxy centre markers were cyan via `MassProfileCentresScatter(c="cy")` |
| 93 | +and are now the `plot_array` default. `autoarray/plot/array.py:258-262` |
| 94 | +hardcodes the overlay colour cycle, and `line_colors` (265-273) applies to |
| 95 | +`lines`, not `positions` — an early misreading of mine, corrected before it |
| 96 | +reached the code. Restoring the colour needs a PyAutoArray change; markers |
| 97 | +remain visible and distinguishable from the mask outline by size and z-order. |
| 98 | + |
| 99 | +## Process note |
| 100 | + |
| 101 | +An intermediate "suite green" reading was wrong twice and caught both times: |
| 102 | +pytest had exited 4 on an unrecognised `--timeout` flag without running, and a |
| 103 | +later run was corrupted when a `git stash` reverted the tree mid-run. Reported |
| 104 | +results come from clean runs on the final tree. Separately, a stop-hook flagged |
| 105 | +the PyAutoGalaxy branch as unpushed: the commit *was* on GitHub, but the |
| 106 | +`--depth 1` clone's single-branch refspec meant no remote-tracking ref existed, |
| 107 | +so `@{u}` could not resolve — repaired by widening the refspec. |
| 108 | + |
| 109 | +## Follow-ups filed, not fixed |
| 110 | + |
| 111 | +- `draft/maintenance/autocti_workspace_test/imaging_ci_heritage_sweep.md` — the |
| 112 | + 13 heritage files the `legacy/` sweep missed. |
| 113 | +- `autoarray` duplicates this normalisation inline in `plot/array.py` and |
| 114 | + `plot/inversion.py`; a shared helper there is the real fix. This task added |
| 115 | + one copy in autogalaxy serving both GUIs rather than a third inline copy. |
| 116 | +- euclid's three `tools/` files are **100% CRLF on `main`**, contradicting its |
| 117 | + own AGENTS.md ("CRLF will break shell scripts on the HPC"). Preserved rather |
| 118 | + than converted, to avoid whole-file diffs that would not fix the repo-wide |
| 119 | + problem. |
| 120 | + |
| 121 | +## Original prompt |
| 122 | + |
| 123 | +# `aplt.Output` stale-API drift in the remaining workspace repos |
| 124 | + |
| 125 | +Type: bug |
| 126 | +Target: workspaces |
| 127 | +Repos: |
| 128 | +- autocti_workspace_test |
| 129 | +- euclid_strong_lens_modeling_pipeline |
| 130 | +Difficulty: small |
| 131 | +Autonomy: supervised |
| 132 | +Priority: normal |
| 133 | +Status: formalised |
| 134 | +Filed: 2026-08-04 (backfilled from git) |
| 135 | +Issued: 2026-08-24 |
| 136 | + |
| 137 | +Split out of `plot-array-stale-kwargs` (HowToGalaxy#56, 2026-08-04), which |
| 138 | +repaired this same drift in `autolens_workspace_developer` but deliberately |
| 139 | +stopped at that repo's boundary. |
| 140 | + |
| 141 | +## The drift |
| 142 | + |
| 143 | +`aplt.Output` no longer exists on the **autolens / autogalaxy** plot namespace — |
| 144 | +verified: `hasattr(autolens.plot, "Output") == False`. It survives only as |
| 145 | +`autoarray.plot.Output`. The removal was deliberate and is already documented in |
| 146 | +`autolens_assistant/AGENTS.md:218` ("the `aplt.MatPlot2D` / `aplt.Output` objects |
| 147 | +have been removed — do not use them"). |
| 148 | + |
| 149 | +Callers must move to the flat convention. Note the accepted kwargs differ per |
| 150 | +callee — check each signature rather than blanket-renaming: |
| 151 | + |
| 152 | +```python |
| 153 | +# plot_array takes all three |
| 154 | +aplt.plot_array(array=..., output_path=P, output_filename=F, output_format="png") |
| 155 | +# subplot_* take only path + format (no output_filename) |
| 156 | +aplt.subplot_tracer(tracer=..., grid=..., output_path=P, output_format="png") |
| 157 | +``` |
| 158 | + |
| 159 | +## Sites |
| 160 | + |
| 161 | +| Repo | Files | Status | |
| 162 | +|------|-------|--------| |
| 163 | +| `autocti_workspace_test` | 27 | **UNVERIFIED — check first** | |
| 164 | +| `euclid_strong_lens_modeling_pipeline/tools/` | 2 (`psf_size.py`, `extra_galaxies_centres_gui.py`) | confirmed broken | |
| 165 | + |
| 166 | +**Do not assume the autocti files are broken.** Those import |
| 167 | +`import autocti.plot as aplt` — a *different* library's plot namespace. |
| 168 | +`autocti` was not installed in the 2026-08-04 session so it could not be |
| 169 | +checked. PyAutoCTI may still export `Output`, in which case those 27 files are |
| 170 | +correct as written and must be left alone. Verify with |
| 171 | +`hasattr(autocti.plot, "Output")` before touching anything. |
| 172 | + |
| 173 | +Confirmed **not** bugs, do not "fix" them: |
| 174 | +- `PyAutoArray/test_autoarray/plot/test_output.py` — there `aplt` *is* |
| 175 | + `autoarray.plot`, which does export `Output`. |
| 176 | +- `autolens_assistant` markdown — documents the removal. |
| 177 | + |
| 178 | +## Verification |
| 179 | + |
| 180 | +Re-run an alias-aware AST scan after the fix (the 2026-08-04 session's first |
| 181 | +sweep hardcoded the alias `aplt` and **missed** a call site written as `aaplt`; |
| 182 | +resolve aliases from each file's own imports). Then bind each changed call's |
| 183 | +kwargs against the real callee signature via `inspect.signature`, since these |
| 184 | +repos have little or no CI to catch a wrong kwarg name. |
0 commit comments