Proves the layered pipeline for both backends with zero backend repo changes.

- tools/gen_docstudio_spec.py: drf-spectacular over an unstract checkout,
  mounting api_v2.execution_urls at its real prefix so generated paths match
  what the server serves.
- tools/annotations.py: the @extend_schema_view block for DeploymentExecution,
  applied at runtime, written in the form it will take in the backend.
- tools/gen_llmw_spec.py: AST walk of the LLMWhisperer Flask controller,
  resolving one level of helper indirection so params read outside the handler
  are not lost.
- poc/facade.py: retry, sync/async POST rule, dict return shapes, and the
  httpx -> requests exception translation downstream depends on.
- poc/cli_generated.py derives its flags from the generated request model; a
  new backend serializer field reaches a CLI flag with zero hand-written lines.

Specs are committed and deterministic; the generated SDK is gitignored and
never hand-edited.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NvFMau3jC1hf5hNisPtsrB

…t fix

Adds the LLMWhisperer half of the published-vs-generated comparison and the
checks that made it useful.

- poc/llmw_facade.py: facade over the generated LLMWhisperer SDK, mirroring
  the published constructor, retry policy and exception translation.
- poc/cli_published.py: whisper/webhook groups so both backends have a real
  side-by-side, with flags derived from the published signature.
- poc/test_llmw_compat.py: signature, constructor and query-string parity.
- poc/facade.py, poc/test_compat.py: reset untouched request-body fields to
  UNSET so the multipart body carries only what the caller set.

The generator writes every spec-declared default into the request, which the
published clients never send. On staging this pinned word_confidence_threshold
and silently dropped low-confidence words from the extracted text. Both facades
now send only what the caller set; GAPS.md and MEASUREMENTS.md record the
bisect.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NvFMau3jC1hf5hNisPtsrB

…laim

- tools/gen_sdk.sh pins openapi-python-client==0.29.0 and reinstalls if the
  venv drifts. Unpinned, a generator upgrade and a spec change produce the
  same diff and cannot be told apart.
- poc/llmw_facade.py gains word_confidence_threshold. The published client
  moved 9862b8f -> 3832713 and test_llmw_compat.py caught the missing
  parameter offline; both CLIs picked up the flag with no edit.
- Spec generation does not need a database. Measured against a dead port,
  the spec is byte-identical for both the deployment and the full tenant
  urlconf, so CI needs the backend venv and nothing else.
- GAPS §14 revised: the bisect baseline was 2.5 months stale, so
  word_confidence_threshold was version skew rather than an over-send. The
  mechanism it proved is unchanged.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NvFMau3jC1hf5hNisPtsrB

Prior-art survey of 14 SDK-codegen projects shows re-declaration of the
surface is what blocks auto-exposure, not hand-writing it. Mature wrappers
inherit or re-export, so new endpoints flow through for free. Our
re-declaration is justified only where a published contract must be pinned.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NvFMau3jC1hf5hNisPtsrB

PR #1's E2E.md targets a much larger CLI — config files, Prompt Studio
projects, tool export, API deployment — none of which this branch has, so an
agent following it fails on scenarios that were never in scope. AGENT_BRIEF
lists the twelve commands that exist, the runnable scenarios, the credentials
to use and avoid, and the run-to-run noise that is not a defect.

Also records why --help is not enough on its own: the specs carry no parameter
descriptions, while the published docstring documents all 25.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NvFMau3jC1hf5hNisPtsrB

Points both clients at a local server returning a scripted status and body and
compares what a caller sees: exception class, status_code, message, or the
returned dict. 25 comparisons across five methods, no network.

Two defects on the first run, both in code that had passed every existing check
and a live round trip. The facade read `resp.parsed`, which the generator builds
only for statuses the spec declares and types loosely for error bodies, so a
well-formed 500 crashed with AttributeError and an undeclared 401 reported
"Invalid JSON response from API" for valid JSON. The facade now reads the body,
as the published client does.

Two more followed from the same change: `_parse_response` calls `.json()`
unguarded on every declared status, so requests now go through `_get_kwargs`
and skip generated response parsing entirely; and `status._get_kwargs` was
over-sending two injected defaults the published client never sends.

Separately the drift gate fired again — llm-whisperer-python-client e8935b3
adopts this POC's finding that two parameters never reached the service, and
settles which deprecated spellings forward. The facade mirrors its resolver, and
the overlay gains `addParams` for parameters read through a helper that takes
the argument name as an argument, which the AST walk cannot see.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NvFMau3jC1hf5hNisPtsrB