fix(providers): cloudflare-ai flattens message content unconditionally, but the #2539 constraint is model-scoped — this blocks image input to Cloudflare vision models

[davidlinfr]

Summary

open-sse/executors/cloudflare-ai.ts flattens every message.content array into a plain
string, and throws when any part is non-text. The comment explains why:

the Workers AI /ai/v1/chat/completions endpoint requires each message content to be a
plain string; it rejects the OpenAI content-part array shape ([{ type:"text", text }])
with HTTP 400 (#2539)

The constraint is real, but it belongs to the model, not to the endpoint. Measured today
against the exact URL buildUrl() produces, same payload, only the model changing:

Model	content = string	content = all-text part array
@cf/mistralai/mistral-small-3.1-24b-instruct	200	200
@cf/meta/llama-4-scout-17b-16e-instruct	200	200
@cf/meta/llama-3.3-70b-instruct-fp8-fast	200	200
@cf/qwen/qwen2.5-coder-32b-instruct (text-only)	200	400

The 400 is AiError: Bad input: … oneOf at '/' not met … — the same error family as #2539,
alive today. Multimodal models declare content: string | array; text-only models declare
content: string. So the guard is unconditional while the constraint is not, and the
#6390 throw it entails refuses image input to vision models that would accept it.

To be explicit about what this PR does not propose: deleting flattenContent outright
would regress every text-only model. I measured that before writing this. The #6390 throw
is also correct given the flattening — refusing beats silently dropping an image. The
defect is one level up: the flattening is applied where it is not needed.

Reproduction

All calls against https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1/chat/completions,
on 2026-08-29. Model is @cf/mistralai/mistral-small-3.1-24b-instruct unless stated:

#	payload	HTTP	What it establishes
T0	content = "plain string"	200	control — credentials and endpoint work
T1	content = [{"type":"text","text":"…"}]	200	array accepted on a multimodal model
T2	content = [{"type":"text",…},{"type":"image_url",…}]	200	image text read back correctly
T3	content = [{"type":"not_a_real_type","blob":"x"}]	400	negative control — still validates
T4	3 messages (user/assistant/user), all content arrays	200	multi-turn, as in #2539
T5	same 3 messages, messages[2] missing role	400	messages[2].role Required
T6	T1 on @cf/qwen/qwen2.5-coder-32b-instruct	400	the constraint, alive, model-scoped

T2 was only run on multimodal models (mistral-small-3.1, llama-4-scout); sending an image
to a text-only model would not establish anything.

Two notes on #2539 itself, which I read in full:

• Its error is an aggregated oneOf, and one line reads
  required properties at '/messages/2' are 'role,content'. T5 reconstructs that shape — a
  message missing role — and gets a 400 naming exactly that field. A malformed message is
  sufficient on its own to fail the request, whatever the other messages' content looks
  like. (T5 is a reconstruction from the reported error, not the original payload.)
• The surrounding lines contradict each other — messages/0 and /1 faulted for being
  arrays "not in string", messages/2 for being a string "not in array". T3 shows the same
  aggregation noise from a single bad part. Those lines are branch noise, not a diagnosis.

I am not claiming #2539 was misdiagnosed with certainty. I am claiming the fix it produced
is broader than the constraint it was aimed at.

Minimal T6, the one that matters for scope:

for M in "@cf/mistralai/mistral-small-3.1-24b-instruct" "@cf/qwen/qwen2.5-coder-32b-instruct"; do
  printf '%s ' "$M"
  curl -s -o /dev/null -w '%{http_code}\n' \
    "https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1/chat/completions" \
    -H "Authorization: Bearer $CF_TOKEN" -H "Content-Type: application/json" \
    -d "{\"model\":\"$M\",\"stream\":false,\"messages\":[{\"role\":\"user\",
         \"content\":[{\"type\":\"text\",\"text\":\"Reply exactly: OK\"}]}]}"
done
# @cf/mistralai/mistral-small-3.1-24b-instruct 200
# @cf/qwen/qwen2.5-coder-32b-instruct 400

### Scope

Two files. `open-sse/executors/cloudflare-ai.ts` — `transformRequest()` only; the flatten
helper is split into `isTextPart` / `flattenTextParts` so the condition reads as the rule it
encodes. `tests/unit/cloudflare-ai-image-parts-6390.test.ts` — the `#6390` expectation is
restated (the attachment must survive, which is what #6390 asked for) and a `#2539`
no-regression witness is added on a text-only model. No refactor in passing, no other
provider touched, no behaviour change for plain-string or all-text content.

### TDD proof (Hard Rule #18)

Tests written first, against unchanged source — **red**, and for the right reason:

not ok 1 - transformRequest passes image_url content parts through untouched (#6390)
error: 'Cloudflare Workers AI chat endpoint does not accept image/non-text content parts …'
not ok 2 - transformRequest never silently drops a non-text part (#6390)
error: 'Cloudflare Workers AI chat endpoint does not accept image/non-text content parts …'

tests 11 · # pass 9 · # fail 2

Same command after the `transformRequest()` change — **green**:

tests 11 · # pass 11 · # fail 0

### Gates

node --import tsx/esm … --test tests/unit/cloudflare-ai-image-parts-6390.test.ts
tests/unit/executor-cloudflare-ai.test.ts 11/11 pass
npm run typecheck:core clean
npx eslint --suppressions-location config/quality/eslint-suppressions.json
open-sse/executors/cloudflare-ai.ts tests/unit/cloudflare-ai-image-parts-6390.test.ts
ESLint: No issues found
node scripts/check/check-file-size.mjs OK
node scripts/check/check-complexity.mjs OK — 2672 (baseline 2774)

`pre-commit` also ran prettier, `eslint --fix`, `check:docs-sync`,
`check:any-budget:t11` and `check:tracked-artifacts` — all clean.

`tests/unit/executor-cloudflare-ai.test.ts` and
`tests/unit/cloudflare-ai-image-parts-6390.test.ts` are the only files importing this
executor, and the thrown message is referenced nowhere else in the codebase (only in
`CHANGELOG.md`), so nothing downstream keys on it.

### Changelog

Fragment added under `changelog.d/fixes/` once this PR has a number.