Repository navigation
Inline workflows get cost basis unknown; price by pipeline identity like the VRAM projection #797
Description
Field report
A per-shot-reference copy of the music-video template, run inline, reported cost basis "unknown". The server had observed runs of the same pipeline.
Assessment (verified)
- Observed cost history is keyed by catalog name.
dw/server/routes/jobs.py:750-762passesobserved=Nonewith nocatalog_name, andObservedCosts.rows_for(name)(dw/server/observed_cost.py:390) looks up by name. - The VRAM projection already matches by pipeline identity:
component_type+model_name+workflow(dw/vram_inheritance.py). - This is documented behaviour: the MCP instructions tell agents to quote the
models/entry by hand. But an agent that inlines a catalog template to change one knob loses its price.
Proposed
Fall back to pricing by pipeline identity for an inline workflow, the same way vram_inheritance matches. Report it under a distinct basis (e.g. "inherited") so it reads as an approximation.
Activity
dkackman commentedon Oct 10, 2026
triage: escalate — adds a new cost basis consumers must learn ('inherited') and a pipeline-identity price-matching path in validate_workflow with an open edge-case matrix (offload/quant/frames differing from the observed runs); needs sign-off on the concept before a fix session builds it. (claude-opus-5-5 via anthropic)
dkackman commentedon Oct 10, 2026
Approved (Don). Build the pipeline-identity fallback for inline workflows. It should match on
component_type+model_name+workflow, the same keydw/vram_inheritance.pyuses, and report under a distinct"inherited"basis so callers read it as an approximation. When the inline workflow's offload, quantization or frame count differ from the observed runs, the result should say so the way validate's VRAM-ceiling warning does: warn, never refuse. Document the new basis wherever the existing bases are listed.Ruling recorded by claude-opus-5-5 (anthropic) for Don.
feat(plan): #797 - price an inline workflow by pipeline identity; bas…
test: #797 - patch.object instead of a string target
10 remaining items
dkackman commentedon Oct 10, 2026
Architecture review: bounced (map finding). I read f90f9df and 4e56a9e under
dw/anddw_mcp/againstdocs/ARCHITECTURE.mdand item 3 ofdocs/stabilization/harness/stage-c-guardrails.md. 1eee13a (#799) is in the same range but I didn't review it.Finding: the map has no row for this. The change adds a new rule with new owners: an inline workflow (one with no catalog name) is priced from a catalog template's observed runs, matched by pipeline identity, under basis
inherited, and is warned, never refused. Those owners aredw/server/deps.py:inherited_observed,dw/vram_inheritance.py:inherited_differences, anddw/plan.py:_observed(theINHERITEDbasis), with the warning indw/server/routes/jobs.py:_inherited_cost_warnings.docs/ARCHITECTURE.mdat a2db75f doesn't name any of them. Under "Validation, plan and cost", the "Inherited VRAM ceiling" row covers only the VRAM ceiling (ceiling_index), and "The estimate" row covers only observed ahead of curated. The checklist says a change that adds an owner updates its row in the same change, andtests/test_architecture_map.pycan't catch a row that's missing.What would resolve it: add an "Inherited cost" row (or extend "The estimate"). Name the owners above and state the rule: same
component_type+model_name+workflowkey asvram_inheritance; most cold runs wins, then name;differscovers offload, quantization and frame count; warn, never refuse. List the enforcing tests fromtests/test_inherited_cost.py. Because the catalog parse now lives in_catalog_listing, which the ceiling index and the cost lookup share, it would help to name it in the "Inherited VRAM ceiling" row too.Clean, for the record:
template_identityandvram_estimate._resolvedfor identity andvariable:resolution,observed_cost.declared_driversplusobserved_for_namefor buckets and history, andplan._temperedfor tempering when there are few runs.CLAUDE.mdchanges; the rationale is in the module docstrings.Reviewed by claude-opus-5-5 (anthropic).
dkackman commentedon Oct 10, 2026
triage: batch with #800 — both are shipped on develop and bounced by architecture review only for a missing docs/ARCHITECTURE.md row; one session adds both rows (per each review comment) and hands both off. (claude-opus-5-5 via anthropic)
docs(arch): #797 #800 - map rows for inherited cost and the level-spr…
dkackman commentedon Oct 10, 2026
Architecture-review finding addressed in 8b52148 on
develop(docs-only:docs/ARCHITECTURE.mdrows added — "Inherited cost", plus `_catalog_listing` named in the ceiling row;tests/test_architecture_map.pypasses). No code changed since the earlier hand-off, so the MCP verify described in the fix comment stands; the driver deploysdevelop. I left offdocs-reviewbecause the tester still has the code fix to verify.claude-sonnet-5-5 via anthropic
dkackman commentedon Oct 10, 2026
Verified on the mini-ai server (mps on mini-ai.lan, develop @ 8b52148) by claude-opus-5-5 (anthropic).
Baseline:
get_workflow("templates/text-to-image")hasobservedwith mps runs (cold 0.18 min, 13 cold runs, drivernum_images_per_prompt: 1).All calls below are
validate_workflow(workflow=<inline copy of templates/text-to-image, id changed>):qa-inline-t2i):basis: "inherited",inherited_from: "templates/text-to-image",differs: [],minutes: 0.2,runs: 13. Warning: "cost estimate is inherited from this box's runs of templates/text-to-image (same pipeline), not of this workflow". Previously this would have beenunknown.configuration.offload: "model":basis: "inherited",differs: ["offload"], and the warning adds "its offload differ from that template, so the real cost may too".valid: true, so it warns and doesn't refuse. ✔StableDiffusionPipeline+stabilityai/stable-diffusion-2-1):basis: "unknown", no inherited warning. ✔arguments: {num_images_per_prompt: 4}):basis: "unknown". That matchesvalidate_workflow(name="templates/text-to-image", arguments={num_images_per_prompt: 4}), which also givesunknown. So the inline copy uses its own driver value, as the hand-off says. ✔I didn't test the
quantizationandframe countentries indiffers;offloadcovers the same mechanism. One cosmetic nit, not a blocker: with a single entry the warning reads "its offload differ" (a plural verb for a single noun).test(e2e): the fixture server keeps its own settings root