Update requirements.txt

requirements.txt

@@ -1,40 +1,41 @@
 # ── PlotWeaver Hausa Voice AI Agent ──────────────────────────────────────────
-# Versions verified by actually launching the app (see test_app.py), not by
-# assumption. Two things were caught this way:
-#   * gradio 4.44 is now BROKEN against current huggingface_hub (HfFolder was
-#     removed) — a fresh Space build on 4.44 fails at import.
-#   * `stream_every` on Audio.stream does not exist before gradio 5.
-# Keep sdk_version in README.md frontmatter in sync with the gradio pin below.
-
-gradio>=5.49.0,<6.0.0       # Audio.stream(stream_every=...) + gr.skip()
+#
+# CRITICAL: do NOT add gradio here.
+#
+# A Gradio Space installs its own gradio from `sdk_version` in README.md
+# frontmatter, as:
+#     pip install -r requirements.txt gradio[oauth,mcp]==<sdk_version> ...
+# Any gradio pin in this file is resolved against that in the SAME pip
+# command, so a mismatch is an unresolvable conflict and the build dies with
+# ResolutionImpossible. To change the gradio version, edit `sdk_version` in
+# README.md — never this file.
+#
+# For the same reason, avoid pinning anything the Space base image already
+# constrains. The base image installs "protobuf<4" and "click<8.1"; requesting
+# protobuf>=4 here is a guaranteed build failure.
+#
+# The base image is Python 3.13, which removed the `audioop` stdlib module —
+# integrations/sip.py implements G.711 mu-law decoding in numpy for that reason.
 
 # ── Core ML ──────────────────────────────────────────────────────────────────
-torch>=2.2.0                # Whisper, NLLB, MMS, Silero VAD
+torch>=2.2.0                # Whisper, NLLB, MMS
 transformers>=4.40.0        # >=4.40 required for VitsModel (MMS-TTS)
 sentencepiece>=0.2.0        # NLLB tokenizer — NOT optional
 accelerate>=0.28.0          # device placement
-protobuf>=4.25.0            # slow→fast tokenizer conversion for NLLB
 
-# ── Audio ────────────────────────────────────────────────────────────────────
-numpy>=1.26.0,<3.0.0
+# ── Audio / numeric ──────────────────────────────────────────────────────────
+numpy>=1.26.0
 scipy>=1.12.0               # scipy.signal.resample — the only resampler used
-soundfile>=0.12.1           # Gradio audio I/O backend
 
 # ── Misc ─────────────────────────────────────────────────────────────────────
-requests>=2.31.0            # WhatsApp Cloud API, Zendesk, HF Inference API
-
-# ── Optional ─────────────────────────────────────────────────────────────────
-# torchaudio is NOT required by this app, but Silero VAD imports it via
-# torch.hub. Without it the built-in energy VAD is used instead, which works
-# fine — uncomment only if you want Silero's noise robustness.
-# torchaudio>=2.2.0
-#
-# Only needed with SIP_PROVIDER=twilio and real credentials; sip.py imports it
-# lazily, so the demo never touches it.
-# twilio>=9.0.0
+requests>=2.31.0            # WhatsApp Cloud API, Zendesk, HF Inference router
 
 # ── Deliberately absent ──────────────────────────────────────────────────────
-# librosa    — resampling uses scipy.signal.resample.
-# silero-vad — loaded via torch.hub with an energy-VAD fallback.
-# audioop    — removed from the stdlib in Python 3.13; G.711 mu-law decoding
-#              is implemented in numpy in integrations/sip.py.
+# gradio      — injected by the Space from README sdk_version (see above).
+# protobuf    — base image pins <4; do not fight it.
+# soundfile   — comes with gradio.
+# torchaudio  — only needed for Silero VAD; the energy VAD fallback is used
+#               otherwise. Uncomment if you want Silero's noise robustness.
+# twilio      — only for SIP_PROVIDER=twilio; sip.py imports it lazily.
+# librosa     — resampling uses scipy.signal.resample.
+# audioop     — removed in Python 3.13; mu-law decode is numpy in sip.py.