feat(platform): voice brain-dump onboarding step

autogpt_platform/backend/.env.default

@@ -294,3 +294,30 @@ TALLY_API_KEY=
 
 # Other Services
 AUTOMOD_API_KEY=
+
+## ===== ONBOARDING BRAIN DUMP ===== ##
+# The voice brain-dump onboarding step is gated on the `onboarding-brain-dump`
+# LaunchDarkly flag and is fail-closed: the endpoints 404 unless it is on.
+#
+# Left commented deliberately — .env.default ships to every install, and an
+# active override here would force the feature on for everyone and bypass
+# LaunchDarkly. Uncomment in your own .env to test locally, and set the
+# matching NEXT_PUBLIC_FORCE_FLAG_ONBOARDING_BRAIN_DUMP in frontend/.env.
+# FORCE_FLAG_ONBOARDING_BRAIN_DUMP=true
+
+# Speech-to-text model for the brain dump. Requires a direct OpenAI key
+# (OPENAI_INTERNAL_API_KEY, else OPENAI_API_KEY) — OpenRouter has no
+# /audio/transcriptions endpoint. Falls back automatically on failure.
+# BRAIN_DUMP_TRANSCRIPTION_MODEL=gpt-4o-transcribe
+# BRAIN_DUMP_TRANSCRIPTION_FALLBACK_MODEL=whisper-1
+
+# Greeting generation for the copilot home after onboarding. The prompt is
+# managed in Langfuse under BRAIN_DUMP_GREETING_PROMPT_NAME when Langfuse
+# credentials are configured; otherwise a built-in prompt is used.
+# BRAIN_DUMP_GREETING_MODEL=anthropic/claude-sonnet-5
+# BRAIN_DUMP_GREETING_PROMPT_NAME=Brain Dump Greeting
+
+# Model that picks which integrations to suggest in the welcome dialog's
+# "Connect your tools" panel. Recommendations are best-effort — a failure
+# here never blocks onboarding.
+# BRAIN_DUMP_RECOMMEND_MODEL=anthropic/claude-sonnet-5

autogpt_platform/backend/backend/api/features/onboarding_dump/EVAL.md

@@ -0,0 +1,137 @@
+# Brain-dump transcription eval
+
+WER (word error rate) harness for the onboarding brain-dump transcription
+pipeline (`transcription.py`). It runs the **real** pipeline — same models,
+same retry/fallback, same ffmpeg splitting and stitching — against a corpus of
+recordings with human-written reference transcripts, and reports how far the
+machine transcript drifts from the truth.
+
+**Release gate: aggregate (pooled) WER must stay under 5%.** The harness exits
+`1` when it is at or above 5%, `0` otherwise, so it can be wired into a release
+check as-is.
+
+## Running it
+
+```bash
+# from autogpt_platform/backend
+poetry run brain-dump-eval --dir /path/to/dumps
+
+# override the primary model (fallback stays whisper-1)
+poetry run brain-dump-eval --dir /path/to/dumps --model gpt-4o-mini-transcribe
+
+# also dump machine-readable results
+poetry run brain-dump-eval --dir /path/to/dumps --json /tmp/wer.json
+```
+
+Run `poetry install` once after pulling this in — `brain-dump-eval` is a new
+console script and won't exist in an older virtualenv. Equivalent without it:
+
+```bash
+poetry run python -m backend.api.features.onboarding_dump.brain_dump_eval --dir /path/to/dumps
+```
+
+Requires a direct OpenAI key (`OPENAI_INTERNAL_API_KEY` or `OPENAI_API_KEY`) —
+OpenRouter does not implement `/audio/transcriptions`. Recordings above the
+single-request byte cap also need `ffmpeg` on `PATH`.
+
+**This costs real money and real minutes.** A 25-file, 2–15 minute corpus is
+roughly 3 hours of audio per run. Run it before a transcription change ships,
+not on every commit.
+
+## The corpus
+
+Put the recordings and their reference transcripts in one flat directory,
+matched **by basename**:
+
+```
+dumps/
+  dump01.webm      dump01.txt
+  dump02.m4a       dump02.txt
+  dump03-es.mp3    dump03-es.txt
+```
+
+- Audio extensions: `.webm` `.mp4` `.m4a` `.mp3` `.wav` `.ogg`
+- Reference: same basename, `.txt`, UTF-8, plain text (no timestamps, no
+  speaker labels)
+- Unmatched files are skipped and listed at the bottom of the report, so a
+  typo'd basename never silently shrinks the corpus.
+
+### What to record — please help fill this out
+
+We need **25+ real rambling dumps**. Not scripts read aloud: the pipeline's
+failure mode is messy natural speech, so read-aloud audio will pass at ~1% WER
+and tell us nothing. Target mix:
+
+- **25+ recordings**, each **2–15 minutes** long
+- **Mixed accents** — at minimum US, UK, Indian, and non-native-English
+  speakers
+- **Background noise** — a few from a café, a street, a room with a fan, a
+  phone on speaker
+- **At least 3–4 non-English dumps** (e.g. Spanish, French, Hindi). The
+  pipeline must transcribe in the language spoken, never translate — a
+  non-English dump that comes back in English is a bug even at low WER
+- **At least 2 long ones (>20 MB)** so the chunked/stitched path is exercised
+- Real rambling: false starts, "um", topic switches, self-corrections, long
+  pauses
+
+### Writing the reference transcript
+
+Type what was actually said, verbatim: keep filler words and false starts, drop
+nothing. Don't clean up grammar. Punctuation, casing and hyphenation don't
+matter — the harness lowercases, strips punctuation and collapses whitespace
+before comparing — but *words* do.
+
+## Reading the report
+
+```
+file                      WER   sub   ins   del     ref    secs  segments
+dump01                  3.20%     8     2     1     344    12.4  1
+dump02                  7.10%    15     4     3     312    41.9  3
+
+Stitch boundaries (chunked files):
+  dump02 — 3 segments
+    seam 0/1: dedup dropped 7 word(s)
+      ...so then I went back to the drawing
+      board and I thought maybe the whole thing
+    seam 1/2: dedup dropped 0 word(s)
+      ...
+
+pooled WER      4.30%  (656 reference words)
+mean WER        5.15%
+median WER      5.15%
+wall clock       54.3s
+
+Release gate: aggregate WER must stay under 5% — PASS (exit 0)
+```
+
+- **pooled WER** — errors over all reference words in the corpus. This is the
+  gated number; long files count more, which is what we want.
+- **mean / median WER** — per-file averages. A mean far above the pooled number
+  means a small file is failing badly; look at the per-file rows.
+- **sub / ins / del** — substitutions dominate for accents and jargon,
+  insertions for hallucination (the failure mode that matters most: a
+  hallucinated sentence becomes a "fact" about the user), deletions for
+  dropped audio and bad seams.
+- **segments** — `1` means single-request; more means the chunked path ran.
+- **Stitch boundaries** — for chunked files only, the last ~10 words of each
+  segment and the first ~10 of the next, plus how many words the overlap-dedup
+  removed. `dropped 0` at a seam usually means the dedup failed to find the
+  overlap and the transcript now repeats a phrase; a large drop means it ate
+  real speech. Both show up as insertions/deletions in the WER but are only
+  diagnosable here.
+
+A file whose transcription raised is shown as `ERROR` and forces a failing exit
+regardless of WER.
+
+## Notes / limits
+
+- The harness has no duration metadata for the files, so it takes the
+  single-request vs chunked decision on the byte cap
+  (`SINGLE_REQUEST_MAX_BYTES`) alone — exactly what
+  `transcribe(audio, filename, duration_secs=None)` does. In production the
+  frontend supplies a duration, so a long-but-small recording can chunk there
+  and not here.
+- WER is computed with an inline Levenshtein over normalised words
+  (`brain_dump_wer.py`); no `jiwer` dependency.
+- The corpus is deliberately **not** committed — the recordings are personal
+  speech. Keep them in shared storage and point `--dir` at a local copy.