Repository navigation
Conversation
…ery surface) Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…t, segment pattern, step_output_dir, step_end, sub-workflows, steering rule) Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…aps it Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…e_name after expansion Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… name not a path Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… feeds Result.save and the pipeline wrapper Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every manifest step entry and step_end event now carries "subfolder" (empty string when the step is unfoldered, the definition's subfolder on a cache hit). Adjusted the exact-key-set assertion in tests/test_workflow_step_cache.py::test_cache_hit_marks_its_manifest_entry_and_event_reused to include the new field. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…id wins; CLAUDE.md states the output: ceiling Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Replaces mcp-feedback.md in the iterate harness repo: tickets now live as GitHub Issues on this repo, with owner:*/status:* labels standing in for the old owner/status fields and comments standing in for notes/verify-notes. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… CLAUDE.md dedupe, fence, wording) Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… subfolders list Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…orkflow state the subfolder convention Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…termediate convention in the authoring guide and its CLAUDE.md mirror Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…n spill lands in the step's subfolder Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ads straight Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…te is convention Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Design for a since_seq/last_seq cursor on wait_for_job/get_job_events (mirroring the SSE endpoint's existing resume semantics) plus an explicit failure_kind field, to close the gap where an active polling loop can silently miss a job's terminal transition (completion or OOM-kill). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ts bullet, CLAUDE.md period Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every saving step of a template with two or more saving steps carries result.subfolder - final for the output the user is shown, intermediate for the scratch that went into it. tests/test_template_subfolders.py asserts the rule so it cannot drift and pins the packaged builtins as unmarked: a role is the parent's to assign. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Where the skill reads the manifest it now says what the family's templates put in final and intermediate, that list_gallery(subfolder=) filters on it, and that a composed workflow keeps the convention. Pinned against the templates' actual roles. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ms in CLAUDE.md Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… is the builtins', not their parents' Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…first Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…headings Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…stages done Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ds live from step_end; doc and test tidy-ups Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The newest-N slice started at `len(jobs) - limit`, which goes negative once the limit exceeds what matched and is then read from the end: `limit=12` against 9 matching jobs answered with the last 3 and reported 6 "not listed". Only bites when `len < limit < 2 * len`, which is why the same data at `limit=20` looked correct. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The result is already in host memory and saving never touches the pipeline, so the old ordering held ~10 GB on the device through the longest phase of a video step. Releasing now also clears the loop's own reference to the step action and reclaims on the spot - a popped pipeline the frame still holds is not freed. A sub-workflow's manifest is read before the release so the rollup is unaffected. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…uilt Two leaks that compounded across retries: - Pipeline.load kept whatever it had built when it raised partway - the pipeline, its quantized weights, a half-applied placement - and the caller never received it, so nothing dropped it. It is now torn down and reclaimed on the way out, and a component shared before the failure is unpublished so no later step reuses a component of a pipeline that does not exist. - The worker reclaimed on success and on cancellation but not on failure: no collection, no device-cache empty, and no eviction of the variants the attempt superseded. It now does all three. The exception's traceback is dropped first - it reaches every frame between the handler and the failure, so a collection with it still live frees none of what those frames hold. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
They are the engine's answer to decode memory pressure and carried no description, so the schema did not say that decoding a batch one sample at a time is already available. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ng, memory - ltx-2.5: the write-out warning states ~4 s/frame at 1536x896 so a reader can scale it to any frame count, rather than "minutes" at 121 frames (#76); intermediate steps take `"result": {"save": false}` (#77). - minimax-music3: the generated arc stretches to fill audio_duration, so the ceiling is set at 1.5x the intended length and trimmed (#78). - both: check idle gpu_memory_allocated_mb with get_memory before a near-ceiling render, and ask for a worker restart rather than retrying into what a failed attempt left behind (#79). Trimmed prose elsewhere in the ltx skill to stay under the 12 KiB cap. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
/api/memory answered live/info only, so a cached reading taken while a job was loading a model was indistinguishable from a live idle one - and reads low, which is wrong in the dangerous direction. The response now carries stale, reason (job_running, worker_stopped, worker_busy, worker_unreachable) and age_seconds beside the existing keys, and info: null keeps its meaning: nothing has been measured because nothing is resident. The get_memory tool description, docs/MCP.md and docs/SERVER.md state what live means and that only live readings compare with each other; the ltx-2.5 and minimax-music3 skills' memory step gives the rule for each state. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The tool description promised "VRAM and RAM" and returned only the card. On this box that is the less informative half: the templates offload weights to host memory by design, so a generation can run with ~8 MB allocated on the GPU while the weights are very much resident - which is exactly the leak path #72 describes and the one a consumer could not see. host_memory_stats() reads the worker process's RSS and high-water mark and the machine's total/available, via psutil when it is installed (a transitive dependency here, never required), /proc otherwise, and resource(2) for the peak. Nothing raises; a reading the platform cannot take is left out of the payload rather than sent as null, so an absent key means "not measurable here" rather than "measured, and nothing". Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The release moved to before the write in 2e2ac05, and the tester found the ordering unverifiable from outside: it happens inside the window between a step's generation and its files appearing, which is 0.4 s for an image step, and nothing in the event stream marked it. Polling get_memory cannot land a sample in that window. A step with release_pipeline now emits `pipeline_released` (step, index, allocated MB before and after, null where the backend cannot say), which sits before that step's `step_end` - so the ordering reads directly off get_job_events on any workflow, at no GPU cost. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Shots generated independently land at whatever level the model chose -
two shots of one scene measured -2.65 and -12.56 dBFS peak - and each
reads as fine alone, because a shot is only wrong relative to what it is
cut against. Butt-joined that is an audible drop, and it is the one seam
artifact none of the fade controls can touch: it is either side of the
seam rather than at it. The workaround was normalize_audio + pair_audio
per shot, two extra steps each, ten for a five-shot cut.
`match_levels` ("rms" for perceived level - the measurement
get_gallery_metadata reports as mean_dbfs - or "peak") scales every track
before the join, to `match_levels_dbfs` or the measure's default; a gain
that would clip is held just below full scale and logged. Off by default,
so nothing existing changes, and an unmatched join whose tracks span 6 dB
or more says so in the log rather than passing in silence.
Both cut templates and assemble-and-score expose it as a variable,
defaulting to null, so a run can even its shots out with an argument.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The metadata route decodes an audio or video file whole to probe it, and the job page showed nothing a probe supplies - so a multi-shot video job opened with one full decode per clip. Under the flat output layout two runs share file names, so the map now clears on a job change. Both seed draws use OS entropy; model ids read in mono, as the engine reads them. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The page reads each image's embedded metadata and links a download, so the mock has to answer both; one test pins that a video is never probed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.