A `git add -A` in a checkout with a worktree under .claude/worktrees/ records
it as a gitlink — a submodule pointer no clone can resolve. That happened in
gadfly today; this is the same one-line prevention, applied before it happens
here.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The progress entry was written before four review rounds reshaped the change:
it credited openaiCompatScheme alone, listed the tests as six per-provider
cases, and mentioned a captureRT detail that has since moved. Rewritten to
match what merges — registerOpenAICompatBuiltin owning both halves,
envKeyForProvider as the single LLM_<NAME> definition, and the shared table
that every OpenAI-compat built-in is now checked against.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
DSN.Scheme's list named kimi but not qwen — the same sibling drift this PR
keeps finding, in a doc comment this time (I updated the README's scheme list
and not this one). Added qwen, and llama-swap/llama-swaps while there, since
they were missing too.
envKeyForProvider's example ran backwards: "LLM_M1 → m1" describes registry
naming, not what the function does. Now reads name → variable in one
direction.
Not taking the third: the const block's alignment is gofmt's own output
(gofmt -l is empty), and the uneven padding is forced by the doc comments that
split the block into alignment groups. glm-5.2 reached that same conclusion in
round 2 before flagging it here.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Three findings, and the first two are the same recurring shape.
envKeyForProvider (env.go) is now the single definition of the LLM_<NAME>
form. It lived in two places — lazy resolution in registry.go and the
missing-key hint in openaiCompatScheme — with a comment on the second asserting
it matched the first. A comment is not enforcement: if either had drifted, a
keyless DSN target would have named a variable that does nothing, and nothing
would have failed.
The kimi and qwen test files had become near-identical, which is round 1's
finding at the level above it: I deduped the fixtures, then left two parallel
suites asserting the same four things. They are now ONE table in
builtin_openaicompat_test.go — endpoint + credential, missing key fails closed
naming its own variable and never reaching the network, the name:// DSN
reaching another host, and a keyless DSN naming LLM_<NAME> instead of the
built-in's key. Adding an OpenAI-compat built-in is a table row that
immediately owes all four; builtin_kimi_test.go is deleted because the table
covers it. Only genuinely qwen-specific tests remain in the qwen file: the
reverse credential leak and the reasoning_effort wire claim ADR-0027 rests on.
Also trimmed ProviderQwen's doc comment, which restated the ADR-0027 rationale
already given at the registration site.
The break-check suite caught its own rot again — two mutations went stale when
these tests were renamed, and the landed-check reported them loudly instead of
passing them off as green. Now 9 cases, including one that drifts
envKeyForProvider to prove the shared helper is load-bearing. 9/9 apply and are
caught.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Same class of finding as round 1, one level in: I factored the DSN-scheme half
of the kimi/qwen duplication into openaiCompatScheme and left the eager
provider half copy-pasted, so a third built-in still had six lines to clone —
including both credential rules, which is exactly the pair you do not want
re-typed.
registerOpenAICompatBuiltin now installs both halves from one call. The rules
that matter hold by construction for every future caller: WithAPIKey passed
unconditionally (an unset key must not fall through to OPENAI_API_KEY), and
WithAPIKeyName naming that same variable in the 401 hint. Registering kimi and
qwen is now one line each.
Also fixed a cross-reference the ADR got wrong: Qwen's image-input caveat is
README matrix footnote ⁴, not ³ — ³ is kimi's. I wrote "³, shared with kimi"
in the ADR and then gave Qwen its own footnote in the README.
The break-check harness needed fixing before any of this could be trusted:
three of its mutations targeted lines this refactor moved, so they matched
nothing, the code was never broken, and the suite reported "test still passed"
— identical output to a test that genuinely misses the bug. Mutations are now
verified to have landed (sha before/after) and the suite fails loudly if one
doesn't. Two new cases cover the helper: dropping the unconditional WithAPIKey,
and dropping the scheme-half registration. 8/8 apply and are caught.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Both findings were the same one, and both were fair: the PR that retires two
byte-identical DSN factories into openaiCompatScheme then copy-pasted the test
fixtures. qwenResponse was byte-identical to kimiResponse, and the single-key
env-lookup closure appeared three times in the new file (plus a fourth in the
kimi file, which neither reviewer was looking at).
Fixed for the class rather than for qwen: captureRT, the canned Chat
Completions body (now chatCompletionOK), and a new singleKeyEnv helper move to
builtin_openaicompat_test.go, owned by no single provider. The kimi tests adopt
them too, so the next OpenAI-compat built-in has nothing left to copy — the
same argument the production helper makes.
Also aligned the test model ids to the current Model Studio names
(qwen3.8-max / qwen3.7-plus), which the docs already cited. One reviewer called
those ids fictional and named the 2025 ones instead; they shipped 2026-08-03
and 2026-05-21 respectively, so that finding is stale model knowledge, not a
defect — but having tests and prose name the same models removes the smell that
prompted it. A dotted id also now proves it passes through verbatim.
Break-checked again after the refactor: all six mutations still fail their test.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Adds the `qwen` built-in provider and the `qwen://` DSN scheme, keyed by
QWEN_API_KEY and defaulting to Model Studio's international host. Like kimi
(ADR-0026) it is `provider/openai` pointed elsewhere — no new client.
Model Studio serves the same models over two protocols, so the real decision
was which wire format to speak. ADR-0027 records why it is the OpenAI one:
down the anthropic client `ReasoningEffort` is ignored by design, structured
output rides the first-party `output_config.format` mechanism the shim does
not implement, and cached-token accounting reads Anthropic-only usage fields.
Each of those fails silently rather than loudly, which is what makes the
choice worth writing down. The shim stays reachable ad hoc via an
`anthropic://` DSN.
The kimi and qwen DSN factories were byte-identical, so they now share one
`openaiCompatScheme` helper: the "credential comes from the DSN token, and
the missing-key hint names LLM_<NAME>" rules hold by construction instead of
by copy.
Tests are hermetic and break-checked (all six fail on a deliberate mutation),
including the reverse credential leak — a visible QWEN_API_KEY must not
authenticate the openai built-in — and reasoning_effort asserted on the wire
body, which is the ADR's load-bearing claim.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The 2/4 finding is right: after writeImagePart started passing an explicit
filename stem, initImageFilename had no caller in video.go, and its
"conditioning frame" doc no longer described its one remaining user
(lipsync.go's avatar image). A one-line wrapper that survives only to be
misdescribed is not indirection worth keeping.
lipsync now calls imageFilename(mime, "frame") directly, and imageFilename's
doc lists the real bases — including WHY the video keyframes need distinct
ones: a backend that stages uploads by filename would otherwise have the
second overwrite the first.
Not taken: consolidating the first/last-frame rationale to a single canonical
site. The copies address different readers — the wire encoding (provider), the
contract's undetectable-support caveat (videogen), and the mode table (README)
— and last round's finding was a doc pointing at a note that did not exist.
Trading duplication for cross-references is what produced that.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01PLjgrxvHjm1sJgUu9zBPH9
- README documented only t2v/i2v. Now a table of the four keyframe
combinations, plus the undetectable-support caveat, which is the one thing a
caller cannot work out for itself.
- The LastImage doc comment said "see the note on LastImage support in
provider/llamaswap" — there was no such note. A pointer to something that
does not exist is worse than no pointer; the comment is now self-contained.
- Generate's doc described only input_reference; it now names
input_reference_last and explains why an unsupporting backend returns a clip
rather than an error.
The 2/4 finding (writeImagePart reusing the "frame" base for both parts) was
already fixed in dbc9689 — from the receiving end, where the consequence is
concrete rather than stylistic: ComfyUI stages uploads by FILENAME with
overwrite=true, so a shared name means the second clobbers the first and both
keyframes resolve to one image.
Not taken: initImageFilename's name is no longer misleading (writeImagePart
stopped calling it), and it is still used by lipsync.go so it is not dead.
The empty-LastImage test stays standalone — it mirrors the existing standalone
empty-InitImage coverage rather than a table this file does not have.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01PLjgrxvHjm1sJgUu9zBPH9
Caught while writing the receiving end. Distinct multipart field names are not
sufficient: backends stage an uploaded frame under a name derived from the
FILENAME, and our own ComfyUI shim posts to /upload/image with overwrite=true.
Both parts were sending initImageFilename(mime) — literally "frame.png" for
each — so the second upload would have clobbered the first and BOTH keyframe
inputs would have resolved to the same stored image.
The failure mode is the worst kind: a clip pinned at both ends to the same
frame renders cleanly, returns 200, and looks like the feature not working
rather than like a bug. Nothing upstream or downstream would report a fault.
writeImagePart now takes the filename stem (frame / frame_last), and the test
asserts the two arrive under different filenames.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01PLjgrxvHjm1sJgUu9zBPH9
videogen.Request gains LastImage alongside InitImage, so one Request covers
t2v, i2v and FL2V without a mode flag. With InitImage it pins both ends of the
clip; alone it pins the destination and lets the backend invent the approach.
The llamaswap provider sends it as a SEPARATE `input_reference_last` part
rather than a second `input_reference`. Multipart permits repeated names, but
then which frame is first and which is last depends on part ORDER — an
ordering contract invisible in the payload, that nothing notices breaking. A
backend that does not know the new name ignores the part, the same degradation
as any other unknown field.
Both parts go through one writeImagePart helper so their encoding cannot
drift, and an empty LastImage is rejected up front exactly as InitImage
already is.
Support is per-model and deliberately NOT advertised in this contract: a
backend that ignores a trailing keyframe returns an ordinary clip, which is
indistinguishable from success. The doc comment says so, because a caller that
needs to know whether the pin took effect has to establish that out of band —
and the mort side gates on a convar for exactly this reason.
Motivated by mort's #1567 (long-form video): with both ends pinned, drift
becomes structurally bounded inside each shot instead of compounding across an
autoregressive chain.
Tests break-checked: sending the last frame under the shared name fails both
the distinct-name assertion and the last-alone case.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01PLjgrxvHjm1sJgUu9zBPH9
All tidiness, no behavior change: the leading-marker class is one shared
constant for citationLabelRe and summaryCloserRe (hand-copying it is how
'+' went missing the first time); the deliberate 'all' duplication across
summaryCopulas/summaryArticle is now stated at both sites; the weak-final
switch case assigns modeBackRef explicitly; test comments state the
constraint they guard instead of which reviewer asked for them.
Co-Authored-By: Claude Fable 5 <[email protected]>
Two behavioral fixes from the review:
- modeSummary's backward scan now stops at the most recent user message.
With the dwarf ratio rejecting the current turn's 1x-3x answer, the old
unbounded scan could walk into WithHistory content and resurrect a stale
answer to a DIFFERENT question — strictly worse than keeping the closer
(opus, correctness). Other modes keep their historical unbounded scan.
- A terminal matching BOTH the ack shape and a back-reference is now
classified back-ref: it carries no answer content, so the looser bar is
the right one (opus, error-handling).
Plus the nits: summaryCloserRe assembled from named fragments, the leading
marker class gains '+' (parity with citationLabelRe), verb-first form takes
'all the', dwarf ratio hoisted into one named local, and the 151-vs-153
char/byte comment inaccuracy corrected.
Co-Authored-By: Claude Fable 5 <[email protected]>
A third degenerate terminal shape from the glm-5.2 cite pattern: the model
front-loads its full answer into the cite-call turn, then closes with a
bookkeeping ack plus a one-line compression ("Citations are logged. Short
version: ..."). mort run b3cb9ee9 delivered 151 chars of a 2,089-char
answer this way — the closer was neither a back-reference (over the 120
cap, no back-ref phrase) nor a citations addendum (no label-colon, no
links), so finalOutput let it stand.
isSummaryCloser keys on the ack sentence alone (the verb must end the
sentence, so prose about citations never matches; a compression marker
without the ack is deliberately out of scope), and the new modeSummary
recovery bar makes the 3x dwarf ratio mandatory at every length: unlike a
back-reference this closer carries real answer content, so it is only
displaced by the clearly-fuller original it compressed.
The citations/back-ref bool becomes a three-way recoveryMode; existing
behavior for both old modes is unchanged.
Co-Authored-By: Claude Fable 5 <[email protected]>
ListFaces now carries head yaw, so a caller choosing WHICH face to swap can
see the thing that decides whether the swap will read — not only learn it
afterwards from the swap report. In the run that prompted this the target's
three faces sat at -82, -8 and -11 degrees; only the first was hopeless, and
nothing in a bounding box said so.
A face swap always returns an image and always looks like success. Whether the
likeness actually transferred is a different question, and until now nothing in
the response answered it — so a caller wanting to know went and asked a vision
model instead. That is wrong in precisely the cases that matter: shown a jogger
in a Georgetown cap holding McDonald's cups, a VLM answers "Bill Clinton"
whoever's face is on him. In the run that prompted this it reported failure on
six consecutive CORRECT swaps (measured afterwards at 0.79-0.84 cosine), and
the caller burned 21 minutes chasing a problem that did not exist.
Result.SwappedFaces now carries, per replaced face: pixel size, the target
image's dimensions, head yaw, and cosine similarity between the source face and
the face actually present in the output.
Yaw and FractionOfImage are the two that explain the complaint. The swap in
question replaced a 138px face in a 1010px-wide photo — 14% of the width,
correct and invisible at a glance — and elsewhere a face turned -82 degrees,
where the features carrying identity are edge-on and any swap reads as a
generic person. Same code on a 168px face in a 385px picture (44%, yaw 2) is
unmistakable. None of that was inferable from a bounding box.
Typed on Result rather than stuffed into Raw: a caller has to act on this, and
a value reachable only by type-asserting an `any` is one nobody finds in time.
doRawHeaders is doRaw with the whole header instead of only Content-Type; doRaw
delegates to it, so the other 25 call sites are untouched and there is still
one place where the status check and the size cap live.
A missing or malformed header yields nil, not an error — an older shim sends no
header, and a swap that produced a good image must not fail because the
diagnostics beside it were unreadable. Covered for absent/garbage/wrong-type,
and the parse is break-checked.
Gadfly on #23, blocking, 2/2 agreement — and it is the exact defect this
whole line of work has been about: a call that succeeds while handing back
the wrong bytes.
sniffImageMIME falls back to image/png when detection is inconclusive, and
the guard only consulted Content-Type. A response with NO Content-Type
therefore skipped the check entirely and was labelled a PNG. The shim answers
JSON on a semantic miss (no face found in the source or target), which is
precisely the body that would have sailed through as a successful image.
The check now validates the BYTES — http.DetectContentType must say image/ —
and the reported MIME prefers the server's own label only when that label is
itself an image type. Break-checked by restoring the header-only condition,
which fails the new test.
Also from that review:
- index is documented as ignored under all=true, so a negative one is no
longer rejected there; it is still rejected when it would actually be
sent, and both halves are tested.
- initImageFilename (video.go) was imageFilename with the base fixed to
"frame" and now delegates to it — two copies of one extension table is
how they drift.
- DetectedFace carried Width/Height alongside Box, two sources of truth for
one fact that can disagree after any transform. Now a Size() method
derived from Box.
- a dead `apiErr` in the test (declared, then `_ = apiErr`) was an
abandoned errors.As check; it is wired up and now asserts callers can
classify the error.
- swapImg duplicated editInit verbatim; removed.
Not taken: adding a FaceSwapProvider/ModelOption surface to match the other
optional imagegen capabilities (single-model finding). There are no options
to carry yet, and inventing an empty option type to look symmetrical would be
API surface with nothing behind it. Worth revisiting when a real knob exists.
Measured against the instruction-edit models on 2026-07-31: asking a diffusion
model to put a SPECIFIC person's face into a photo does not work by any route.
qwen-image-edit returns the picture essentially unchanged whether asked by
name, by attribute, or by supplying the portrait as a second reference image;
flux-kontext replaces the face with a different generic person. Identity
transfer is a detect/align/blend pipeline, not a better prompt, so it gets its
own interface rather than more Edit options.
imagegen.FaceSwapper is optional and type-asserted, like Editor — a provider
that cannot do this must not have Edit quietly stand in for it.
ListFaces is part of the interface, not a convenience: a caller asked to
change "the man on the right" needs a stable way to NAME one face, and pixel
boxes let it check its own choice. The llamaswap shim orders faces left to
right for exactly that reason (insightface's own order is score-ranked and
unstable between near-identical images), and a malformed box is a protocol
error rather than a zero-filled struct, because a wrong box aims the swap at
the wrong person.
The provider is the first here to POST more than one file, so buildMultipart
gained buildMultipartFiles and now delegates to it — one writer loop, so the
two cannot drift in how they escape names or terminate the body.
index and all are mutually exclusive ON THE WIRE: the shim ignores index under
all=true, and sending both would imply a precedence the caller cannot see.
A JSON body is refused rather than returned as image bytes — the shim answers
JSON on a semantic miss (no face in the source), and handing that back as a
picture would report success while delivering a file that is not one.
FLUX.1 Kontext and Qwen-Image-Edit are a different kind of edit from img2img
and reach sd-server by a different path, and nothing in imagegen could
express it: EditRequest only had Init, which is noised and denoised back
under the prompt.
Measured against FLUX.1-Kontext on the netherstorm host 2026-07-30, on a
synthetic scene with a red rectangle, a blue rectangle and a flat background,
prompted "change the blue rectangle on the right to bright green, keep
everything else exactly the same":
via init_images (the only path that existed)
right rect (60,60,200) -> (47,82,228) still blue, instruction ignored
left rect (200,60,60) -> (229,43,50) drifted
background (150,200,240) -> (154,211,229) drifted
via extra_images (this change)
right rect (60,60,200) -> (70,254,4) green, as asked
left rect (200,60,60) -> (204,57,57) intact
background (150,200,240) -> (151,202,247) intact
No mask, no strength, no compositing — the model is handed the picture as
conditioning and the prompt as an instruction about it.
EditRequest.RefImages selects the path; when set, Init/Mask/Strength are
ignored rather than rejected, so a caller handing the same request to
whichever model is configured gets the better result on a Kontext-class model
instead of an error. The provider posts /sdapi/v1/txt2img with extra_images
(sd-server reads that field on both routes into gen_params.ref_images, where
the CLI's -r/--ref-image also lands); there is no init latent to denoise, so
sending one would only add noise to a pipeline that does not want any.
An all-empty reference set is refused: it would otherwise degrade into a
plain txt2img and render the prompt from scratch, which is not the request.
Merge the parallel callCounts/lastResults maps into one repeatState struct
map (removes the "two maps in sync" smell + double lookup), drop the dead
i<len(results) bounds branch that contradicted the documented index invariant,
and note in-code that exact-string result equality is a deliberate err-toward-
not-tripping choice (a hung job whose poll reports a ticking field is left to
MaxRuntime / the job ceiling rather than risking a false kill of real progress).
No behavior change; guard tests still green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01HgEuVfZJN9mhRhzEsMEVog
The maxSameCallRepeats guard counted identical (name+arguments) tool calls
across a run and tripped ErrToolLoop past the ceiling — regardless of whether
each call made progress. This killed legitimate polling of long-running
background jobs: code_exec_poll must be called with identical args (same
job_id), so a render/encode that needs more than N polls was guillotined
mid-flight even as each poll returned an advancing result (elapsed/status
moving forward).
Only count an identical call toward the trip when its RESULT is unchanged
from the previous identical call. A call whose result keeps changing is
progress and resets its count; a genuinely stuck call returning the same
output still trips. This can never trip more than before, only less, and
covers every idempotent poller with no per-tool configuration — matching the
progress-over-usage thesis behind the stall-detection work.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01HgEuVfZJN9mhRhzEsMEVog
- kimi:// DSN scheme: missing-credential hint now names the LLM_<NAME> env
var that defines the provider (its token comes from the DSN, not
KIMI_API_KEY), matching providerFor's lazy-resolution key form. Fixes the
correctness/error-handling findings that the old hint misdirected users to
set KIMI_API_KEY when the fix is adding a token to the DSN.
- parse_test.go: add kimi to TestBuiltinsResolve. (llama-swap stays excluded
and is now documented — its no-URL built-in errors at Model() construction,
not just on use, so it can't resolve there; the finding's llama-swap half
was a false lead the test surfaced.)
- Add TestKimiSchemeMissingToken covering the corrected hint.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add a first-class `kimi` provider and `kimi://` DSN scheme for Moonshot AI's
OpenAI-compatible Chat Completions API. Both reuse provider/openai (no new
client, mirroring llama-swap's chat path). Default endpoint is the
international host; the China endpoint is reachable via a kimi:// LLM_* DSN.
- Credential is KIMI_API_KEY, read through the registry's injected envLookup
so it stays hermetically testable. WithAPIKey is passed unconditionally so
an unset KIMI_API_KEY can never fall through to the openai client's
OPENAI_API_KEY default.
- New openai.WithAPIKeyName option customizes the missing-key error hint
(default OPENAI_API_KEY); kimi names KIMI_API_KEY.
- Hermetic tests: built-in base URL + bearer, missing-key hint names
KIMI_API_KEY with no OPENAI fallthrough and no network hit, kimi:// scheme
round-trips against the China host.
- Docs in sync: README built-in table + DSN scheme list + support matrix,
.env.example, env.go DSN doc, ADR-0026 (+ index, backfilling 0024/0025),
progress.md.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Picks up gadfly's opencode/<model> engine — the reusable at this ref pins the
sha-bb98fae reviewer image that bundles the OpenCode CLI. Also correct the pin
comment: this is an immutable sha pin, not the v1 tag it claimed to track.
- SubmitChain rejects NaN/±Inf segment seconds with ErrUnsupported
(previously an obscure json.Marshal error; NaN fails every comparison
and +Inf passed the >= 0 check).
- ChainStatus skips segment entries with no usable id — JSON null
(which no-op-unmarshals into a string, previously appending ""),
empty strings, and id-less objects; the unfiltered list survives in
Raw. ChainJob.SegmentIDs doc now also says ChainSegmentResult takes
the segment index, not an id string.
- jobPath rejects '%' in job ids — %2F/%2E%2E percent-escapes decode
back into path structure server-side, bypassing the literal check on
this upstream-echoed value.
- singleVideoResult moves to video.go next to videoMIME, and the two
remaining hand-rolled copies of the video-result validation
(videoModel.Generate, Interpolate) now use it — one validation, one
message shape.
- videogen.LipSyncer renamed to videogen.Lipsyncer for consistency with
the rest of the surface's Lipsync* naming (LipsyncProvider,
LipsyncModel, LipsyncRequest); not yet consumed downstream, so the
rename is free now and never again.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01WWCQcYStWXBYUy5sZWnbLT
- speakWithReference now validates the response IS audio via the same
audioResultMIME sniff the sfx/enhance surfaces use — a 2xx JSON soft
error or HTML proxy page was previously wrapped up as audio/wav bytes.
- audioResultMIME moves to audio.go (next to speechMIME; it was defined
in sfx.go but shared by enhance/clone) and learns the Ogg container
normalization (application/ogg -> audio/ogg).
- Stems zip unpack gains entry-count (16) and total-decompressed (1GB)
caps on top of the existing per-entry cap — the per-entry bound alone
still let a many-entry bomb multiply up.
- sanitizeFilename drops NUL and both path separators too, so upload
metadata can never smuggle directory structure to a file-writing shim.
- New maxAudioResponseBytes (256MB) for bodies that ARE one audio clip
(clone, enhance, sfx): a long WAV legitimately passes the 64MB JSON
cap.
- speakWithReference local renamed path -> upPath (naming parity with
stems/enhance).
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01WWCQcYStWXBYUy5sZWnbLT
- Segment: reject NaN thresholds (NaN fails every comparison, so it
passed the [0,1] range check and reached the shim as the literal
string "NaN"); ±Inf were already caught by the range comparisons,
now covered by tests too.
- upstreamPath: reject '%' in model ids — %2F/%2E%2E percent-escapes
decode back into path structure server-side, bypassing the literal
/?#/.. rejection. Ids never legitimately contain '%'.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01WWCQcYStWXBYUy5sZWnbLT
- audio.StemSeparator/StemSeparationProvider: Demucs zip transport via
POST /upstream/<id>/v1/stems (Mode two -> two_stems=vocals; model +
format fields); bounded zip unpack, entry name -> stem, ext -> MIME.
- SFXModel reuses musicgen against the sync /upstream/<id>/v1/sfx route
(JSON prompt/seconds/steps/cfg_scale/seed -> WAV); musicgen.Request
gains CFGScale.
- audio.SpeechEnhancer/SpeechEnhancementProvider:
POST /upstream/<id>/v1/enhance -> WAV (result reuses SpeechResult).
- SpeechRequest.ReferenceAudio/ReferenceMIME (+WithReferenceAudio):
llamaswap switches to the chatterbox clone route
POST /upstream/<id>/v1/audio/speech/upload (input + voice_file),
wav MIME fallback.
- TranscriptionRequest.Translate (+WithTranslate): translate=true form
field, language=auto forced when no explicit hint (whisper.cpp default
en would skip translation).
- httptest contract tests (zip unpack, clone-route switch, translate +
auto-language injection); ADR-0024 (index row deferred — MJ-A backfills
the ADR index table and parallel edits would conflict).
Co-Authored-By: Claude Fable 5 <[email protected]>
Gitea >= 1.27 does not propagate workflow_dispatch inputs into a called
workflow's github.event; the stub must pass pr_number as an explicit
workflow_call input or manual dispatches die at 'PR required'. Mirrors
gadfly#24.
Co-Authored-By: Claude Fable 5 <[email protected]>
The netherstorm reviewer failed with 'unknown provider': the correctly
formatted GADFLY_ENDPOINT_NETHERSTORM user var was never forwarded by
the reusable workflow (hardcoded env list). Mirrors gadfly#22.
Co-Authored-By: Claude Fable 5 <[email protected]>
Gitea 1.27 hands called workflows event_name=workflow_call, so every
review since 2026-07-14 self-skipped in 1s while reporting success. The
hotfix lineage (gadfly 5007597 + entrypoint reclassification, image
sha-ed9e946) restores the exact pre-upgrade reviewer; gadfly main
carries the same fix for the executus re-platform rollout.
Co-Authored-By: Claude Fable 5 <[email protected]>
Round 2 from live smokes on netherstorm (2026-07-14):
- ACE-Step result blob is an ARRAY of objects and carries RAW control
characters inside string values (literal newlines) — strict JSON
rejected it. parseMusicResult sanitizes control chars (only legal
inside string values in the double-encoded blob) and accepts array or
object shapes. Regression test uses the live payload shape.
- Hunyuan3D GenerationRequest has NO output-format field (the documented
type param is fiction) — it always returns GLB. Results are now
labelled by sniffed magic bytes, never by the requested format.
- NEW meshgen.Converter/ConverterProvider optional surface + llamaswap
impl over the mediautils shim POST /v1/convert_mesh — the STL hop for
the printer pipeline.
Co-Authored-By: Claude Fable 5 <[email protected]>
- pollResult tolerates up to 5 CONSECUTIVE bad polls (transport blip,
unparseable payload, task momentarily absent) instead of killing a
multi-minute exclusive-GPU job on the first hiccup; only status=2, a
failure run, or ctx deadline aborts
- server-supplied result.File must be server-relative; combined with the
upstreamPath dot-dot/scheme rejection this stops a hostile upstream
from steering the follow-up GET at other proxy endpoints (test:
../../api/models/unload refused)
- WithSteps(<=0) rejected; embed responses repeating an index rejected;
musicFormatMIME now wraps speechMIME (one format table, wav32
normalized); poll interval is a test-shrinkable var (CI no longer
burns 2s+ per music test); parens + comments per review
Co-Authored-By: Claude Fable 5 <[email protected]>
- NEW musicgen leaf package: blocking Generate over ACE-Step's async job
queue (release_task -> poll query_result -> fetch file, all via
/upstream); tolerant envelope parsing, double-encoded result handled
- NEW embeddings leaf package: EmbedModel + RerankModel as separate mints
(two server instances on the host, llama.cpp #20085); InstructedQuery
helper for Qwen3-style query/document asymmetry
- provider/llamaswap: /v1/embeddings + /v1/rerank clients with strict
validation (index-ordered vectors, count mismatch and out-of-range
index are hard errors; rerank sorted descending, minimal parser)
Co-Authored-By: Claude Fable 5 <[email protected]>
- upstreamPath rejects '..' in model ids AND in the rest path — the rest
can embed SERVER-SUPPLIED components (ACE-Step result file URLs), so
dot-dot/scheme smuggling toward other proxy endpoints is refused
- singleImageResult requires positive image evidence (sniffed magic OR
declared image/*): an empty-Content-Type error page can no longer pass
as 'the image' via sniffImageMIME's PNG-default labelling
- upscale/background responses get a dedicated 256MB cap (the 64MB cap
is JSON-sized; a 4x PNG legitimately exceeds it)
- mesh JSON-detection widened (512-byte whitespace-tolerant peek + reject
declared application/json)
- Transcribe now reuses buildMultipart; transcriptionFilename takes
(filename, mime) so diarize shares it without a fake request struct;
truncateForError stops shadowing builtin cap; OnlyMask doc de-ambiguated
Co-Authored-By: Claude Fable 5 <[email protected]>
maxVideoResponseBytes (512MB) replaces the shared 64MB JSON cap on the
/v1/videos/sync read path (doRaw now takes the cap per call) — 3/6
models flagged that a legitimate long/high-bitrate clip would be
discarded after minutes of GPU work. Plus a stale stable-diffusion
comment in initImageFilename and a test-handler early return.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AXQxVhXBw8PwFAtsVrXSmj
- videoMIME no longer hard-falls-back to video/mp4: a 2xx body that is
neither declared nor sniffable as video (JSON job envelope, HTML error
page) is now an APIError instead of a 'successful' corrupt clip.
- Resolution rides the wire as width/height AND the OpenAI-style size
string, so either upstream convention honors an explicit request.
- writeFormFields + mimeFromContentType shared helpers replace the
copied multipart loop (audio.go/video.go) and Content-Type branch.
- ADR-0019 indexed in docs/adr/README.md; README gains the videogen
section + support-matrix mention (docs-parity rule).
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AXQxVhXBw8PwFAtsVrXSmj
New videogen/ contract package (ADR-0019): Request/Result/Model/Provider
with the imagegen conventions. Text-to-video and image-to-video are one
surface (Request.InitImage, nil = t2v) since hybrid checkpoints like
Wan 2.2 TI2V serve both from one model; Result carries a single clip.
provider/llamaswap gains VideoModel(id) targeting the blocking
POST {base}/v1/videos/sync (multipart, model-routed by the fork's new
video routes): vLLM-Omni parameter names, OpenAI-style input_reference
file part, optional fields stay off the wire so per-model launch-flag
defaults apply. CLAUDE.md package map picks up audio/ (missed in #12)
and videogen/.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AXQxVhXBw8PwFAtsVrXSmj
- Transcribe: sanitize the caller-supplied multipart filename (CR/LF
would inject Content-Disposition headers; upload metadata is
untrusted), always send the required model/response_format fields,
parse MIME parameters before extension matching, and give audio/opus
its own .opus extension.
- doRaw: a response larger than maxResponseBytes is now an error, not a
silent truncation.
- Shared plumbing: requireBaseURL() + newRequest() helpers replace the
7x-duplicated guard/error string and the triplicated request
building across doJSON/doRaw/Health.
- Health: non-2xx now returns *llm.APIError (package convention,
programmatically distinguishable from transport failure) instead of
a one-off unexported error type.
- Speak: reject negative Speed; speechMIME no longer accepts video/*
Content-Types.
- image.go: Generate/Edit share one sdWire validate+map helper.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AXQxVhXBw8PwFAtsVrXSmj
- New leaf package `audio` (ADR-0017): SpeechModel/SpeechProvider and
TranscriptionModel/TranscriptionProvider with imagegen conventions
(zero value = backend default, functional options + Apply, bytes
in/out, never URLs). Root re-exports added.
- imagegen.Editor (ADR-0018): optional image-to-image interface —
EditRequest carries the generation knobs plus Init image and
denoising Strength; separate interface so existing Models keep
compiling.
- provider/llamaswap implements all of it: POST /v1/audio/speech (JSON,
raw-audio response, MIME from Content-Type with format fallback),
POST /v1/audio/transcriptions (multipart, response_format=json),
ListVoices (GET /v1/audio/voices?model=, tolerant of string-list and
object-list shapes), POST /sdapi/v1/img2img (txt2img wire +
init_images/denoising_strength, shared image decode), and Health(ctx)
(GET /health) — a cheap liveness probe for often-offline hosts.
- Hermetic httptest coverage for every new wire shape and validation
path; README sections + support-matrix footnote updated in the same
commit (also corrects the stale /v1/images/generations claim — the
image path has been SDAPI since the seed fix).
First consumer: mort's llamaswap media tool cluster (status / image /
TTS / STT agent tools against the netherstorm host).
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AXQxVhXBw8PwFAtsVrXSmj
finalOutput now recovers the front-loaded answer when the terminal turn is a
sources/citations-only addendum ("Sources: [x](url), ..."), not just when it is
empty or a back-reference. It recovers the prior substantive answer and appends
the (real) citations below it. Guards: citation-DOMINANCE (a prose answer that
merely opens with "Source: ... http://..." is left as the answer), ^-anchored
heading, citations recovery decoupled from the terminal-length ratio (concise
answers recover too), preamble filter applied only in the borderline band
(long answers opening with "Sure,"/"Let me" are not vetoed), and a dedup that
ignores <url> angle-bracket wrappers. Healthy terminal answers unchanged; zero
extra model calls.
Fixes mort #1418. Gadfly-reviewed (6 reviewers) + adversarially pre-verified;
all findings graded, real ones addressed.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
The OpenAI /v1/images/generations endpoint ignores `seed` on our
stable-diffusion.cpp build — every render of a given prompt comes back
byte-identical, so a drawbot batch of N collapsed to one image. Switch the
image provider to sd-server's A1111 /sdapi/v1/txt2img endpoint, which honors
`seed` (verified live: distinct seeds -> distinct images on SDXL and
Qwen-Image). Size is split into width/height; llama-swap still routes by the
`model` field. Tests + ADR-0016 updated.
Add Steps, CFGScale, NegativePrompt, Sampler, Seed to imagegen.Request
(pointer/empty = leave the backend's per-model default), with mirror
options, and forward them in the llamaswap wire payload as the
stable-diffusion.cpp fields (steps/cfg_scale/negative_prompt/
sample_method/seed). Unset fields are omitted so sd-server keeps its
baked defaults.
Lets callers (e.g. mort drawbots) override only what they explicitly set.
ragnaros/qwen3.6-27b noted TestNormalizeOverCount matched 'omitted' by substring;
the test is in-package, so assert == imageOverflowPlaceholder instead — robust to
wording changes. No behavior change.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Review fixes (no behavior change):
- Fold the over-cap elide INTO the existing copy-on-write normalize pass: one
loop now replaces the first toElide (oldest) images with the placeholder and
size-normalizes the rest, so the Messages slice is copied at most once (the
prior dropOldestImages + the normalize loop double-copied when overflow and a
transform both applied — the dominant review finding, 5 models).
- Remove dropOldestImages (the name implied removal; it substituted) and the
one-shot hasImagePart helper — both subsumed by the single pass.
- Trim the 9-line inline comment that restated the package doc.
- Test: rename TestNormalizeTooManyImages_DropsOldest → TestNormalizeOverCount
(file convention) and assert the EXACT survivors ([b, c], in order) + a
content-based non-mutation check (first input part is still image a, which a
len check wouldn't catch).
Build + media + majordomo suites green (-race).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
media.Normalize refused (ErrUnsupported) when a request carried more images than
the target's MaxImagesPerReq, on the theory that a failover chain would try a
roomier target. In practice the chain's targets share the same cap — an agent loop
that accumulates a preview image per iteration (e.g. scaddy's write_scad) blows
past the cap, EVERY target rejects ("9 images, target allows at most 8"), and the
run dies. Observed live on ollama-cloud (cap 8).
Now: over-count keeps the most-recent MaxImagesPerReq images and replaces each
older one with a short text placeholder ("[earlier image omitted to fit this
model's per-request image limit]"), preserving each message's turn structure and
telling the model an image was elided. The most-recent images are the relevant
ones in an iterative run. Copy-on-write; the input request is never mutated. The
per-model threshold stays configurable via Capabilities.MaxImagesPerReq (0 still
means no image support); SupportsImages / MIME / byte-budget / dimension behavior
is unchanged, and the provider-side count backstop remains.
Test: TestNormalizeTooManyImages_DropsOldest — 3 images, cap 2 → 2 kept (the most
recent), 1 placeholder, no error, oldest dropped, input unmutated.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Drop the four lowest-graded reviewers — m5/qwen3.6:35b-mlx, gemma4:cloud,
gpt-oss:120b-cloud, kimi-k2.7-code:cloud. Removing m5/qwen3.6 takes the last
local Mac out, so this is now a cloud-only fleet of 6 ollama-cloud models;
GADFLY_ENDPOINT_M5 and the m5 concurrency entry are gone and the per-job timeout
drops to 45m. README/CLAUDE.md kept in sync.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
llama-swap was http-only by DSN, pushing TLS-fronted instances onto the openai://
scheme (which loses the management/image methods). Add a "llama-swaps" scheme
that builds an https base URL, alongside "llama-swap" (http, local-first) —
mirroring redis/rediss. Both share one factory; llama-swaps is scheme-only (no
default built-in). The choice stays explicit because a DSN has no reliable
http-vs-https signal.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
- Unload: reject model ids containing path separators (/?#) so a model name
can't redirect the request to another endpoint; ":" (common in ids) stays
verbatim.
- doJSON: take a model arg so image/management HTTP errors carry the target id
(was always ""); add a base-URL guard so management methods fail clearly
instead of building a bare-path request; cap the success-path JSON decode with
io.LimitReader (64 MiB) and drain the body when out is nil for conn reuse.
- image: reject negative Request.N before sending.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add provider/llamaswap, a tailored provider for llama-swap (the model-swapping
proxy over llama.cpp / stable-diffusion.cpp). Its chat path delegates to
provider/openai at {base}/v1 — no duplicated wire client (ADR-0007) — with
legacy max_tokens, a Bearer no-key placeholder for keyless local instances, and
a timeout-free client so cold model swaps rely on context deadlines. The
"tailored" surface is concrete management methods (ListModels / Running /
Unload) that don't belong on the canonical llm.Provider interface. The
llama-swap:// DSN scheme builds an http base URL (local-first); a no-URL
built-in errors clearly on use, mirroring foreman.
Add imagegen, a new canonical text-to-image interface separate from llm
(Request/Result/Model/Provider; Image = llm.ImagePart so generated images feed
straight back into chat). First backend is llama-swap via OpenAI
/v1/images/generations (b64_json, bytes-only). Re-exported from the root. v1 is
txt2img only.
Hermetic httptest coverage for chat delegation, management endpoints, image
decode, and scheme wiring. ADR-0015 + ADR-0016, README support matrix +
image-gen section, CLAUDE.md package map, and progress.md updated in the same
commit.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
M1 was consistently slow (26-29 min) for zero real findings, so pull it before
this workflow ever fires. Leaves the 9 ollama-cloud models + the M5 Mac;
removes GADFLY_ENDPOINT_M1 and the m1 concurrency entry. Mirrors the same change
on executus.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Records the PR workflow: push work to a PR (never straight to main), wait for
Gadfly to finish and weigh its findings, then grade each finding back to the
gadfly-reports MCP (record_finding_grade / list_findings / scoreboard) so the
telemetry can measure whether each model earns its keep.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Installs the standalone Gadfly agentic adversarial reviewer (advisory, never
blocks merge), mirroring executus's setup on the latest pinned image
(sha-d7f364d). Reviews majordomo PRs with the full fleet: 9 ollama-cloud models
plus the M1/M5 Macs via foreman, each running the 3-lens suite (security,
correctness, error-handling). Posts one consolidated comment per model.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
The agent loop took the final answer only from the terminal (no-tool-call)
turn. Models that "front-load" their answer into an earlier turn that also
calls a tool — then close with a trivial pointer like "(Already answered
above.)" — had their real answer discarded and the pointer delivered. This
recurs across several open-weight models (glm-5.2, etc.); well-behaved models
(Claude/GPT) defer their answer to the terminal turn and are unaffected.
finalOutput() now falls back to the last substantive assistant content in the
transcript when the terminal text is weak (empty, or a short back-reference).
The predicate is narrow and back-reference-gated so short-but-correct answers
("42", "It's down, restarting now.") are never overridden; recovery only picks
a prior turn that reads like a real answer, not a preamble. Zero extra model
calls. Terminal-answer behavior for normal runs is unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
A failover chain previously treated a successful-but-empty completion (no
content parts and no tool calls — a "stop with nothing") as a valid result
and returned it. The agent loop then ended the run with empty output, and
the configured backup models were never tried because no error was raised.
This let a single flaky model silently terminate an agent/skill run with
no answer (observed in the wild with ollama-cloud/glm-5.2 returning empty
completions right after a large tool/think turn).
- Add llm.ErrEmptyResponse (classified transient) and Response.IsEmpty():
true only when there are no tool calls and no meaningful content (no
parts, or whitespace-only text). A media/image part counts as content,
so image-only responses are NOT empty.
- chain.Generate converts an empty completion into ErrEmptyResponse so the
chain fails over to the next target. Unlike an ordinary transient it is
NOT retried on the same target (the model just produced it; these calls
are expensive) — the chain penalizes health (so a persistently-empty
target benches) and advances immediately.
- When every target returns empty the call fails with ErrChainExhausted
joined to ErrEmptyResponse — a visible error instead of a hollow success.
Single-element chains therefore also surface empties as errors.
Stream path is unchanged (can't inspect content before the consumer reads
it). Tests: Response.IsEmpty table; chain fails over past an empty head;
all-empty chain returns ErrChainExhausted/ErrEmptyResponse; repeated
empties bench the target across requests. Full suite green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Phase 8: all six live checks pass (tier aliases, thinking-tier chat, real
tool invocation, structured Generate[T], forced failover with bench+skip,
skill agent). Discovery: ollama.com ignores the format field — the
provider now also states the schema as a system instruction (constrained
decoding locally, instruction-guided JSON on cloud), with hermetic test.
Co-Authored-By: Claude Fable 5 <[email protected]>
Groundwork for the provider phase: reasoning levels map to native knobs
(OpenAI reasoning_effort, Ollama think); ErrUnsupported marks declared
capability mismatches that chains advance past without health penalty.
Co-Authored-By: Claude Fable 5 <[email protected]>
steveandClaude Opus 5