|
| 1 | +- completed: 2026-08-24 |
| 2 | +- issue: https://github.com/PyAutoLabs/autocti_workspace/issues/27 |
| 3 | +- prs: |
| 4 | + - https://github.com/PyAutoLabs/autocti_workspace/pull/28 (merged) |
| 5 | +- summary: | |
| 6 | + CTI resurrection epic **Phase 5**. Smoke coverage for the ordered-trap |
| 7 | + `modeling/start_here.py`-class scripts, unblocked by PyAutoFit#1520 |
| 8 | + (`438f56fac`). All nine verified running bypassed; a curated three promoted. |
| 9 | + This is autocti_workspace's FIRST CI. |
| 10 | + |
| 11 | +## The prompt's premise did not hold |
| 12 | + |
| 13 | +The prompt said to "add them back to the workspace `smoke_tests.txt`". **There is |
| 14 | +no `smoke_tests.txt`, and never was.** autocti_workspace has no `.github/` |
| 15 | +directory and no `config/build/`; `git ls-files` finds neither, and |
| 16 | +`git log --diff-filter=D` shows nothing was ever deleted. So the coverage had to |
| 17 | +be *created*, not re-enabled — which is why the task was larger than its |
| 18 | +"add lines to a file" sizing suggested. |
| 19 | + |
| 20 | +That also surfaced pre-existing drift worth knowing about: |
| 21 | +`PyAutoHeart/config/repos.yaml` lists autocti_workspace under `workspaces`, |
| 22 | +whose required checks are `["Smoke Tests", "Navigator Check"]`. Neither workflow |
| 23 | +existed, so that gate had been vacuous. This PR closes the smoke half. |
| 24 | +(Heart's *local* smoke runner table — the `smoke:` block in the same file — has |
| 25 | +no autocti entry either, and its `import_names` map has no PyAutoCTI. Untouched |
| 26 | +here; a genuine follow-up.) |
| 27 | + |
| 28 | +## Verification — all nine run bypassed |
| 29 | + |
| 30 | +Stack built from source `main` with PyAutoFit at `438f56fac` (which is `main` |
| 31 | +HEAD, so "a PyAutoFit main containing that SHA" is satisfied by definition), |
| 32 | +autonerves/autoarray 2026.8.17.1, autocti 2024.11.13.2, arcticpy 2.6. |
| 33 | + |
| 34 | +`output/` cleared between EVERY run. This matters: a bypassed fit calls |
| 35 | +`paths.completed()`, so re-running with the same `unique_tag` replays the old |
| 36 | +result via `result_via_completed_fit` and looks exactly like "the fix didn't |
| 37 | +work". |
| 38 | + |
| 39 | +| script | time | promoted | |
| 40 | +|---|---|---| |
| 41 | +| `dataset_1d/modeling/start_here.py` | 13 s | yes | |
| 42 | +| `dataset_1d/modeling/customize/priors.py` | 12 s | no | |
| 43 | +| `dataset_1d/modeling/features/species_x3.py` | 18 s | yes | |
| 44 | +| `dataset_1d/modeling/features/visualize_full.py` | 20 s | no | |
| 45 | +| `imaging_ci/modeling/start_here.py` | 96 s | yes | |
| 46 | +| `imaging_ci/modeling/features/cosmic_rays.py` | 120 s | no | |
| 47 | +| `imaging_ci/modeling/features/non_uniform.py` | 53 s | no | |
| 48 | +| `imaging_ci/modeling/features/serial_cti.py` | 94 s | no | |
| 49 | +| `imaging_ci/modeling/features/visualize_full.py` | 96 s | no | |
| 50 | + |
| 51 | +All nine **pass**. The fix is directly visible in the output: |
| 52 | +`dataset_1d/modeling/start_here.py` builds two `TrapInstantCapture` models with |
| 53 | +identical priors under `trap_0.release_timescale < trap_1.release_timescale`, |
| 54 | +and resolves them to `0.14` and `1.68` — distinct and correctly ordered — rather |
| 55 | +than tying at the prior medians and hard-failing the bypass. |
| 56 | + |
| 57 | +## Why three, not nine |
| 58 | + |
| 59 | +Every CTI repo's AGENTS.md says to keep the smoke list a small curated subset |
| 60 | +and not to mass-promote. Measured the existing CTI smoke suite |
| 61 | +(autocti_workspace_test) for a like-for-like baseline: **20 s** total |
| 62 | +(8 + 5 + 7). All nine here cost **~522 s** — a ~26x jump on the CTI smoke bill. |
| 63 | + |
| 64 | +Promoted three, covering both dataset geometries and the multi-species |
| 65 | +ordered-trap case — the surface the bypass fix actually unblocked. **132 s |
| 66 | +measured cold**, with `dataset/` wiped to its committed state so every dataset |
| 67 | +is simulated first (datasets are gitignored and auto-simulated by the scripts' |
| 68 | +`should_simulate` guards, so a fresh CI checkout always pays that). |
| 69 | + |
| 70 | +CI confirmed the estimate: the runner step took **129 s** on Python 3.12 and |
| 71 | +~123 s on 3.13. Both legs green. |
| 72 | + |
| 73 | +## What was added |
| 74 | + |
| 75 | +- `.github/workflows/smoke_tests.yml` — thin caller for Heart's reusable smoke |
| 76 | + workflow, chain `PyAutoNerves PyAutoFit PyAutoArray PyAutoCTI`, `arcticpy: true`. |
| 77 | +- `.github/scripts/smoke_install.sh` — install epilogue carrying **no** arcticpy |
| 78 | + recipe; the same-day arcticpy-install-standardisation task made Heart's |
| 79 | + `install-arcticpy` action its single owner, and Heart runs it before this |
| 80 | + epilogue. This repo never had a copy to delete. |
| 81 | +- `.github/scripts/run_smoke.py` — thin shim over PyAutoHands' `run_python.py`, |
| 82 | + mirroring autocti_workspace_test's. Holds no logic, so it needs none of the |
| 83 | + sweeps the old 198-line copies did. |
| 84 | +- `config/build/profile_smoke.yaml` — `PYAUTO_TEST_MODE=2` and cache dirs. |
| 85 | + Passed Heart's strict env-profile validator first time. |
| 86 | +- `smoke_tests.txt` — the curated three. |
| 87 | +- `AGENTS.md` — a "Smoke tests (CI)" section carrying the full nine-script |
| 88 | + timing table, so the next person promoting a script starts from numbers. |
| 89 | + |
| 90 | +## Left undone, deliberately |
| 91 | + |
| 92 | +**Navigator Check.** Heart lists it as a required check for this repo alongside |
| 93 | +Smoke Tests and it is still absent. Out of scope for an ordered-trap smoke task; |
| 94 | +adding a second unrelated workflow would have widened the PR. Worth a follow-up |
| 95 | +prompt. |
| 96 | + |
| 97 | +## Still open for a human |
| 98 | + |
| 99 | +autocti_workspace_test's `## Conventions` says integration scripts are |
| 100 | +"single-trap" with no stated reason. The reason was almost certainly the bypass |
| 101 | +crash, which is now fixed — so either drop it and exercise multi-trap ordered |
| 102 | +models (better coverage, since ordered traps are the realistic CTI case), or |
| 103 | +write the real reason in. Deliberately not decided here. |
| 104 | + |
| 105 | +## Original prompt |
| 106 | + |
| 107 | +# Re-enable autocti_workspace smoke coverage of the ordered-trap modeling scripts |
| 108 | + |
| 109 | +Type: test |
| 110 | +Target: autocti_workspace |
| 111 | +Repos: |
| 112 | +- @autocti_workspace |
| 113 | +Difficulty: medium |
| 114 | +Autonomy: supervised |
| 115 | +Priority: normal |
| 116 | +Status: formalised |
| 117 | +Filed: 2026-08-24 |
| 118 | +Issued: 2026-08-24 |
| 119 | +Epic: CTI resurrection — Phase 5 |
| 120 | + |
| 121 | +Re-homed from `draft/test/autofit/` by the filing session. Intake set |
| 122 | +`Target: PyAutoFit` and listed PyAutoFit as an affected repo — wrong on both: |
| 123 | +the PyAutoFit fix has already merged (438f56fac), so this task changes no |
| 124 | +library source and only needs a PyAutoFit main that contains it. Difficulty |
| 125 | +lowered from `large`: the work is verify-scripts-run, add lines to |
| 126 | +`smoke_tests.txt`, and time them against the gate cap — sized by script count, |
| 127 | +which is not yet known, hence `medium` rather than `small`. |
| 128 | + |
| 129 | +Re-enable @autocti_workspace smoke coverage of the modeling/start_here.py-class scripts, unblocked by PyAutoFit#1520 (merged 438f56fac). Those scripts were un-smokeable at PYAUTO_TEST_MODE=2 because ordered trap models tie at the prior medians and the bypass hard-failed; the bypass now selects a deterministic assertion-valid point, at TEST_MODE 2 and 3. This is CTI resurrection epic Phase 5. Work: confirm the scripts now run bypassed against a PyAutoFit main that includes 438f56fac, add them back to the workspace smoke_tests.txt, and check timings against the smoke gate cap before adding them. |
| 130 | + |
| 131 | +<!-- formalised by the Intake (Conception) Agent on 2026-08-24 from user-intake --> |
0 commit comments