…ript servers

TypeScript CLI tool that uses the Codex SDK to generate clean, well-structured
Pydantic v2 models from a River TypeScript server's source code. Instead of the
current mechanical codegen that produces awful names and duplicated types, this
points a Codex agent at the TypeScript source so it can read how types are named
and organised, then generates matching Python models.

Key features:
- Reads TypeScript TypeBox definitions to mirror naming and composition
- Generates shared/reusable types (common errors, domain types)
- Verifies correctness by comparing generated JSON schemas against the
  serialised River schema (via a Python verification script)
- Iterates on failures: feeds verification errors back to the model
- Exposed as a CLI for GitHub Actions integration

Usage:
  codegen-llm generate \
    --server-src ./path/to/ts/services \
    --schema ./schema.json \
    --output ./generated \
    --client-name MyClient

…concrete examples

- Switch sandbox from danger-full-access to workspace-write with
  approvalPolicy: never — agent can only access its workspace and
  the explicitly listed additionalDirectories (TS source, existing
  client). It can no longer browse river-python or other repos.

- Prompt overhaul:
  - Quality bar framing: output will be discarded if not clean/readable
  - BANNED patterns section: RootModel, make_schema_model,
    __get_pydantic_json_schema__, SchemaAdapter, create_model, raw
    JSON Schema dicts — all explicitly rejected
  - 6 concrete TypeBox-to-Pydantic translation examples covering
    Type.Object, $kind unions, error unions, Optional/Record/Array,
    recursive types, and Type.Intersect flattening
  - Directory scope section: only access workspace + TS source
  - Stronger anti-shortcut language throughout

- Verification script improvements:
  - Code quality pre-check: scans all .py files for banned patterns
    before comparing schemas, fails with exit code 2 if found
  - New normalizations: Uint8Array->string, strip type alongside
    const, enum->anyOf+const, strip null variant from 2-element
    anyOf (handles TypeBox Optional vs Pydantic Optional mismatch),
    strip discriminator and additionalProperties metadata

…red names, fix bare $ref resolution

…r redefs; add allOf flattening to verifier

- Add extractNamingHints() to codegen.ts: walks schema.json to produce
  naming_hints.json with correct PascalCase error/kind names, written
  into the agent workspace so it has authoritative naming data
- Complete rewrite of prompts.ts: naming_hints.json featured prominently,
  scaffolding scripts explicitly banned, mandatory TS reading phase,
  previous-failures section with concrete bad patterns from runs 1-4
- Add ALLCAPS class name check to verify-script.ts: regex bans class
  names with 4+ consecutive uppercase letters (e.g. NOTFOUNDError)

…ained Literals

Run 5 analysis: agent wrote scaffolding script to /tmp, created
JsonAdapter hack class, produced 200+ char stuttered class names,
and chained Literal[x] | Literal[y] instead of Literal[x, y].

Structural changes:
- Move Python venv outside workspace into verify dir so the agent
  has no access to a Python interpreter (only ./verify works)
- Prompt explicitly states no Python available in workspace

New code quality checks in verifier:
- Ban JsonAdapter and custom json_schema() methods
- Ban class names > 60 characters (mechanical path-derived naming)
- Ban chained Literal[x] | Literal[y] | Literal[z] (3+ in a row)

Prompt updates:
- Remove .venv from file scope, add 'no Python' warnings
- Change Literal style examples from chained to multi-value
- Add failures 5-7 from run 5 (scaffolding to /tmp, JsonAdapter,
  chained Literals)

…p.py

Run 6 analysis: agent switched to Node.js scaffolding script (since
Python venv was removed), generated 1966 model classes with decent
naming, but _schema_map.py was a complete cheat — loaded schema.json
at runtime via SimpleNamespace objects with fake json_schema() methods.
Verification passed trivially without ever testing the actual models.

Fixes:
- Add isinstance(adapter, TypeAdapter) check in verify script — rejects
  any adapter that isn't a real pydantic TypeAdapter
- Ban SimpleNamespace, _make_adapter, _schema_path, _schema_doc patterns
- Ban loading schema.json at runtime in generated code
- Add failure case 8 to prompt explaining the cheat and the new check

The toolResult  value appears on two variants (status=ok and
status=error). Pydantic raises 'mapped to multiple choices' when
using Field(discriminator='kind'). Added concrete example showing
how to handle this: nest sub-variants by status, use plain union
for the outer type when any  value is duplicated.

Run 8 passed verification (268/268) but used WithJsonSchema on every
adapter: TypeAdapter(Annotated[Any, WithJsonSchema(json.loads(...))])
The 1174 Pydantic model classes were decorative — never actually tested
since adapters returned raw embedded JSON schemas.

Fixes:
- Ban WithJsonSchema in verify script banned patterns
- Ban json.loads( in verify script banned patterns
- Add failure case 9 to prompt documenting this cheat
- Update banned constructs list in prompt

…ring)

Pass 1 generates schema-correct Pydantic models (may have mechanical names).
Pass 2 aggressively refactors for production quality: TS-derived class names,
error deduplication, cleanup of alphabetic suffixes and deep path names.

New CLI flags:
  --pass1-only    Stop after Pass 1
  --pass1-dir     Skip Pass 1, refactor existing output
  --pass2-max-attempts  Separate retry budget for Pass 2

Both passes use the same verifier to ensure correctness is preserved.

…; add TS export name extraction

Key changes for the next generation run:

Verify script improvements:
- Add _dedup_anyof normalization: deduplicate structurally identical
  anyOf variants so the agent can reuse error classes instead of creating
  *Duplicate/*Triplicate copies
- Ban *Duplicate/*Triplicate class name suffixes
- Ban monkey-patching patterns: .json_schema =, _bind_reference,
  _load_reference, frozen_schema, adapter.json_schema
- Add duplicate field declaration detection (catches e.g. ParseError
  declaring extras: twice)

Naming improvements:
- Extract TypeBox export names from TS source into naming_hints.json
  (tsExportNames per service, tsSharedExportNames for lib/ dirs)
- Agent now gets pre-extracted names like ExitInfo, MonitorResponse,
  ScreenshotAction, FilesystemError mapped to their service directories

Prompt improvements (Pass 1 + Pass 2):
- Document anyOf dedup normalization -- agent should reuse error classes
- Ban _schema_map.py monkey-patching patterns explicitly
- Reference tsExportNames/tsSharedExportNames from naming_hints.json
- Add Run 10 failure patterns (Duplicate classes, monkey-patching,
  duplicate fields, missing TS names)
- Rewrite _schema_map.py instructions to be explicit about what's banned