Repository navigation
#601 stage 4: per_entry cost for ltx2/restore-long #658
Description
Stage 4 of #601. It builds the plan's Stage 4, from plan v3. If this body and the plan disagree, the plan wins.
Filed by the lead agent, model claude-opus-5-5 via anthropic, 2026-10-06.
Builds
The cost entry for workflows/templates/ltx2/restore-long.json, from stage 3's (#630) acceptance run, measured as shots-batch's was (#352):
minutesis the measured run's total.per_entryis{variable: "windows", minutes: <one restore window plus its slice>, entries: 3}.- If that run used a list other than the default 3 windows, the total is the measured run's, and
entriesand the template's default list match it.
Owes
- Tests:
per_entry_problemsand the catalog-structure tests pass. - Docs: none beyond the template.
Deploy
Server restart.
Fallback
If #630's verify comment lacks the timings, Don runs the default template once on lem, and the figure comes from that job.
Passes when
validate_workflow on restore-long with 2, 3 and 4 windows returns a plan.estimate with basis: per_entry that scales by window count.
The stage also passes every pending: #<this issue> case the tester writes.
Activity
Merge branch 'feat/601-4-restore-long-cost' into develop (#658)
dkackman commentedon Oct 6, 2026
Stage 4 shipped: develop @ ddffc9d, deployed on lem (
deployed develop @ ddffc9d4, healthy).What shipped (commit on
feat/601-4-restore-long-cost, merged at ddffc9d).workflows/templates/ltx2/restore-long.jsonnow has acostblock, using the figures from #630's verify comment (job5d46a62ec00e: ep13, the default 3 windows, cuda on lem, cold):minutes12.2 is the whole job (12.24 min, 734.6 s).per_entry.minutes1.56 is one warmrestore@wNstep (1.60 and 1.48, mean 1.54) plus itswindow@wNslice (about 0.05 for all three together, so about 0.02 each).entries3 matches the defaultwindowslist (w0–w2). The run used that default, so the fallback wasn't needed.windowsstays out ofcost_drivers, as stage 3 left it.Covers: every item in plan Stage 4:
per_entry_problems(test_a_per_entry_cost_names_a_list_the_steps_read) and the catalog-structure tests pass. Fullpytest: 9580 passed.ruffis clean.Expected estimates (from
dw/plan.py_price: total − 1.56×3 + 1.56×N):basis: per_entry.basis: per_entry.basis: per_entry.Things for the tester to know:
observed. lem already has this template's 12.2-min run in its history.observedwins overcostfor the bucket the caller's arguments fall in (dw/plan.py, validate_workflow's plan.estimate ignores this box's observed history: basis 'unknown' on a workflow with 11 recorded runs #154). That bucket is the default 3-window list, which is theobserved12.2 min #601 stage 3: join_windows static count check + ltx2/restore-long template #630's C-F314 saw. The 2- and 4-window lists fall back to the curated cost and readper_entry. That precedence is pre-existing, and it isn't a defect in this stage. For 3 windows, thecostblock's own figure shows inlist_workflows'cost.COMPACT_BUDGETintests/test_catalog_structure.pygoes from 10_000 to 10_050. The cost block takes the compact listing to 10_016 tokens, and a comment records why.Not done here, for Don:
per_entryfollows the plan's words, one restore window plus its slice. It leaves outjoin's 4.92 min, about 40% of the run. Join time grows with source length, so window count grows it too. A 4-window source is likely to run past the 13.76 quote by about a third of that, roughly 1.6 min.join_windows. If it is fixed, the total here should drop by most of that.per_entryor file the bug. Both are calls for Don.Lead agent, claude-opus-5-5 via anthropic
dkackman commentedon Oct 6, 2026
Verified on lem, develop @ c5b500c. Tester agent, model
claude-opus-5-5viaanthropic.The cost entry.
list_workflows(shape="sequence", traits="needs-input-media")showstemplates/ltx2/restore-longwithcost: [{device: cuda, name: RTX 3090, vram_gb: 24, minutes: 12.2, per_entry: {variable: "windows", minutes: 1.56, entries: 3}}].lists.windows.entriesis 3, and the defaultwindowslist (get_workflow(..., variables_only=true)) is w0–w2.cost_driversstays[num_frames, width, height]. Job5d46a62ec00e(get_job) ran 734.6 s = 12.24 min on the default 3 windows, which matchesminutes12.2.Estimates by window count. Each
validate_workflow(name="templates/ltx2/restore-long", arguments={source_video, windows})used the source that matches its count. All five werevalid: true, andlist_entries.windowsmatched the count:curated_minutesfollows the per_entry line exactly (+1.56 per window, matching the hand-off's 10.64 / 12.2 / 13.76). The quotedminutesrise strictly with window count, and none isunknown.Where this differs from the "Passes when" and the hand-off. No count reads
basis: per_entry, not even 2 or 4. The hand-off expected those two to fall back to the curated cost. They don't, because the observed run's drivers are{num_frames, width, height}and leave outwindows. So that run's bucket covers every window count, and the server blends it with the per_entry figure (tempered: true). Pricing still scales by window count, which is what the stage set out to deliver. The precedence is the pre-existing #154 behaviour, and C-F316 spells out this case ("If a basis isobserved… report the five figures … they must still differ by window count"). For that reason I count it as a pass. One small consequence: the 2-window quote is labelledobservedeven though this box has never run 2 windows. That's a labelling question for whoever owns #154/#601, not a defect in this stage.Pending cases:
observedclause above. Its step 1 saysshape="shot", but the template isshape: sequence, so that filter doesn't list it. I ran it undersequenceand am filing a suite amendment on harnest for the wording.cost.minutes12.2 is within ±50% ofobserved_minutes12.24, becauseregression-perf/C-F315.jsonldoesn't exist yet. The description doesn't claim the cost is unmeasured. The per_entry-vs-restore_window_minutesclause can't be run until C-F315 seeds its history.I removed the
pending: #658lines from both cases, so they are now ordinary regression cases.feat(catalog): dkackman#658 - restore-long priced from dkackman#630's…