`deepseek-v4-flash-free` (zen free tier): stream drops the final text delta before `tool_calls`; non-streaming is complete

Description

deepseek-v4-flash-free via https://opencode.ai/zen/v1/chat/completions (free tier) drops the final text delta before tool_calls from streamed responses: the sentence before the tool call ends mid-word, no error, finish_reason: "tool_calls". The same request with stream: false returns the complete text. Non-deterministic — identical requests sometimes complete.

Plugins

None — raw HTTP, no client SDK involved.

OpenCode version

N/A (API-level; also reproduced via Vercel AI SDK streamText, pi's SDK streamSimple, and headless opencode run)

Steps to reproduce

repro.sh — curl + python3, no other deps:

#!/usr/bin/env bash
# Usage: ./repro.sh [runs]
set -u
KEY="${OPENCODE_API_KEY:-public}"          # free tier accepts "public"
N="${1:-8}"
PHRASE="The quick brown fox jumps over the lazy dog"

req() { # $1 = stream(true/false)
  curl -sS --max-time 90 \
    -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
    -d "{\"model\":\"deepseek-v4-flash-free\",\"messages\":[{\"role\":\"user\",\"content\":\"Say exactly '$PHRASE' then call the tool get_time with an empty object.\"}],\"stream\":$1,\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_time\",\"parameters\":{\"type\":\"object\",\"properties\":{}}}}]}" \
    "https://opencode.ai/zen/v1/chat/completions"
}

stream_text() {
  python3 -c '
import sys, json
text = ""
for line in sys.stdin:
    line = line.strip()
    if not line.startswith("data:"): continue
    p = line[5:].strip()
    if p == "[DONE]": continue
    try: o = json.loads(p)
    except Exception: continue
    for c in o.get("choices", []):
        d = c.get("delta") or {}
        if isinstance(d.get("content"), str): text += d["content"]
print(text)'
}

trunc=0
for i in $(seq 1 "$N"); do
  body=$(req true)
  id=$(printf '%s' "$body" | grep -o '"id":"[^"]*"' | head -1 | cut -d'"' -f4)
  text=$(printf '%s' "$body" | stream_text)
  if [[ "$text" == *"lazy dog"* ]]; then s="complete"
  elif [[ "$text" == *"fox"* ]]; then s="TRUNCATED"; trunc=$((trunc+1))
  else s="no-text"; fi
  printf 'run %2d  id=%-22s  %-9s  ...%s\n' "$i" "${id:0:22}" "$s" "${text: -32}"
  sleep 0.3
done
echo
echo "streaming: $trunc/$N truncated (expected tail: '$PHRASE')"
echo "--- non-streaming control (stream:false) ---"
req false | python3 -c 'import sys,json; o=json.load(sys.stdin); print("content:", o["choices"][0]["message"]["content"])'

Screenshot and/or share link

Output of one run:

run  1  id=router-c651e84ac345a25  TRUNCATED  ...e quick brown fox jumps over the
run  2  id=router-4b3abcc145d862d  TRUNCATED  ...e quick brown fox jumps over the
run  3  id=router-59f37695f6c301f  TRUNCATED  ...
run  4  id=router-906ca39fd5009d0  TRUNCATED  ...e quick brown fox jumps over the
run  5  id=router-4f990f930cbd9cd  TRUNCATED  ...e quick brown fox jumps over the
run  6  id=b7ea5f8c-1b71-4d5a-9d3  complete   ...rown fox jumps over the lazy dog
run  7  id=router-87d80b13af7041b  TRUNCATED  ...e quick brown fox jumps over the
run  8  id=router-b0831d8d0f852b9  TRUNCATED  ...e quick brown fox jumps over the

streaming: 7/8 truncated (expected tail: 'The quick brown fox jumps over the lazy dog')
--- non-streaming control (stream:false) ---
content: The quick brown fox jumps over the lazy dog

Second run: streaming: 6/6 truncated, non-streaming control complete.

Operating System

Linux (container); client-independent

Terminal

N/A

Additional evidence

• Raw SSE of a truncated run (verbatim capture, no parser): content='The' → content=' quick brown fox jumps over the' (last content delta) → tool_calls → finish_reason: tool_calls → [DONE]. The tail ' lazy dog' never appears as a content delta. The full sentence IS present in the same stream's reasoning_content deltas and in the stream: false response — the model generated it; the SSE dropped it.
• Truncation correlates with the response chunk id prefix. Across ~100 streamed runs with tools: every truncated response had a router-* id; uuid-prefixed responses were always complete. In the systematic matrix: 28/33 router-* runs truncated (~85%), 0/14 uuid runs.
• Request parameters don't change it: reproduced with/without temperature, x-opencode-session, stream_options, max_tokens, and different tool-schema shapes.
• Not a parsing artifact: with the Vercel AI SDK and pi's SDK, the SDK's emitted text matched the captured raw bytes byte-for-byte (both truncated).

Related: #40888 (streaming + tools truncated on the Go tier), #40176 (deepseek-v4-flash on Go drops the tool call after text — same text↔tool transition defect, opposite side).

Start with the issue's repro.sh and compare the raw SSE for stream:true against the stream:false control, focusing on the content delta immediately before tool_calls. The issue is complete when the streamed response preserves the full text before the tool call across the affected router-* responses without regressing tool-call delivery.