Solid domain-specific tool set with clear naming (verb_noun pattern) and comprehensive descriptions. All 20 tools have descriptions (194-char baseline met). Input schemas present with types and descriptions for most parameters. However, several tools lack output schema documentation, and some parameter descriptions could be more prescriptive about constraints and formats. Error handling is present but inconsistent, some tools guide recovery, others return generic summaries. Two-step safety pattern (plan/apply) is well-implemented across reversible operations, reducing accidental data loss risk. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications.
Batch-set a reference on objects: Attenuation / Conversion ShareSet, or OutputBus (a Bus), by name or id. plan/apply, verified by read-back, undoable.
Audit audio sources: flag sources with no original WAV path and original files used by multiple sources. Read-only. (WAAPI does not expose sample rate/channels.)
Read-only project health audit. Checks: 'no_output_bus', 'empty_container', 'empty_sound', 'event_not_in_soundbank', 'duplicate_name' (default: all). Returns counts + a capped findings list.
Count how many objects route to each Output Bus (mix-routing overview). Read-only.
Report the connected Wwise version and which tools are expected to work on it. Run this first on an untested Wwise version to self-diagnose compatibility.
Wrap a set of objects into a new container (default RandomSequenceContainer), created under their parent, then move them into it. plan/apply, undoable. The create + moves run in one undo group; if creation fails, the group is cancelled.
Output schemas not documented. Tools return dict/to_dict() but LLMs cannot see field structure, forcing them to infer response shape and risking downstream tool chaining failures.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications (READ_ONLY, REVERSIBLE, WRITE). LLMs cannot distinguish safe reads from destructive writes without explicit hints.
Parameter descriptions lack prescriptive constraints. E.g., 'transform' in wwise_rename accepts 'spaces_to_underscores' but no enum or validation guidance. 'pattern' in wwise_naming_check is 'Optional regex pattern' but no format/example provided.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | <=2025-11-25 | v2 |
Create a Play event '<prefix><SoundName>' for target Sounds that lack one (default prefix 'Play_'). Two-step safety: mode='plan' previews + returns a token; mode='apply' with that token creates them (idempotent, undoable). If object_ids is omitted, all Sounds are considered.
Generate a Markdown audio-design doc (project summary, event list, Mermaid bus hierarchy). Returns the markdown; if `output_path` is given, also writes it there.
Audit the project, then propose/apply the auto-fixable issues: whitespace names (→ rename), events not in a SoundBank (→ add to default bank), sounds with no Output Bus (→ default bus). Issues that need a human (empty sound/container, duplicate name, non-ASCII names) are reported, never auto-changed. `dry_run=True` previews.
Deep-dive one object: all property/reference values, parent, children, and (for Events) action targets. `object` may be a GUID '{...}', a path '\...', or a name.
Move objects to a new parent. plan/apply, verified by read-back, undoable.
Flag objects with naming-convention issues (whitespace, non-ASCII, and optionally names not matching `pattern`, a regex). Read-only.
Batch-import audio files into Wwise: create AudioFileSource objects and optionally wrap them in Sounds. plan/apply, undoable.
Run a read-only WAQL query against the open Wwise project. `waql` example: '$ from type Event where name : "Footstep"'. Results are capped.
Batch-rename objects by a transform (default 'spaces_to_underscores', which fixes whitespace in names). Two-step: mode='plan' previews + returns a token; mode='apply' with that token renames (verified by read-back, undoable).
Batch-set a property on objects. Two-step safety: mode='plan' -> returns a ChangeSet + token (no writes); mode='apply' -> pass the token to execute the planned changes.
Add objects (e.g. Events) to a SoundBank's inclusion list (by bank name or id). Two-step: mode='plan' previews + returns a token; mode='apply' with that token writes (idempotent, undoable).
Generate SoundBanks to disk (all banks if `soundbanks` omitted). Returns a log summary with any error count. This is an action — it writes output, not project data.
Check the connection to Wwise and report the open project.
Cross-check Unity C# event usage against Wwise. Scans .cs files under `unity_path` for PostEvent("X") string literals and diffs against actual Wwise events: reports events referenced in code but missing in Wwise, and Wwise events never referenced in code. (String-literal calls only; enum IDs not detected.)
Error handling inconsistent. Some tools return generic 'Invalid mode' messages; others return structured dicts with 'summary', 'succeeded', 'failed'. No guidance on retryability or next steps for LLM recovery.
Plan/apply token storage via ChangeSetStore is server-side but not documented in tool descriptions. LLMs may not understand that tokens are ephemeral or session-scoped, risking stale token errors.