Parallel Mock Implementation

Parallel mock implementation

This is a building block, not a standalone bring-up workflow. Invoke it as part of /implement-mock or /compute-llk-bringup when the work-list is large enough to amortize the dispatch overhead.

When to use

• ≥4 similar files to author (e.g. a batch of SFPU shims, a sweep of missing NOC overloads, a fanout of fabric primitives).
• Each output is independent — one file per worker, no shared edits, no read dependencies between workers.

Don't use it for:

• Single-file work — call the worker prompt directly inline.
• Cross-cutting wiring that touches a shared file (e.g. one sfpu_split_includes.h, one .claude/references/structure.yaml). The orchestrator handles those centrally after workers return.

Worker prompt skeleton

Each worker writes exactly one file. Customize the constraints for your scope; the structure below stays the same.

You are implementing an emule mock for tt-emule. Write exactly ONE file.

TARGET FILE: <relative path under include/jit_hw/ or src/>
UPSTREAM REFERENCE (read-only): <path to the silicon API definition>
ADDITIONAL REFERENCE (read-only, optional): <e.g. LLK impl, RTL spec>
REFERENCE EMULE FILES (read for boilerplate):
  <path to an existing similar mock — keep this list short>

REQUIRED SIGNATURES (mirror upstream — if upstream differs, follow upstream):
<list>

SEMANTICS:
<formula / contract>

HARD CONSTRAINTS:
- Do NOT modify any file other than the TARGET FILE.
- Do NOT touch shared wiring files (the orchestrator handles those).
- Do NOT run pytest, build, or commit.
- Read at most 3 files: upstream, optional second reference, one
  existing mock as boilerplate.
- (Add scope-specific exclusions — e.g. for compute shims:
  "Do NOT read tt_metal/llrt/, tt_metal/impl/dispatch/,
  tt_metal/soc_descriptors/".)

OUTPUT: write the file, return status=DONE with file=<relative path>
and lines=<count>.
If signature is ambiguous or upstream missing, return status=STUCK with
a one-sentence reason.
Do NOT invent signatures.

Result schema

const RESULT_SCHEMA = {
  type: 'object',
  properties: {
    status: { type: 'string', enum: ['DONE', 'STUCK'] },
    file:   { type: 'string' },
    lines:  { type: 'integer' },
    reason: { type: 'string' },
  },
  required: ['status', 'file'],
  additionalProperties: false,
}

Orchestration

phase('Write mocks')
const results = await parallel(ITEMS.map(it => () =>
  agent(makePrompt(it), {
    label: `mock:${it.name}`,
    phase: 'Write mocks',
    schema: RESULT_SCHEMA,
  }).then(r => ({ item: it.name, ...(r || {status: 'NULL'}) }))
))
const done  = results.filter(r => r.status === 'DONE')
const stuck = results.filter(r => r.status !== 'DONE')
log(`${done.length}/${ITEMS.length} DONE; ${stuck.length} STUCK`)
return { done, stuck }

Orchestrator finalization

After the workers return, the orchestrator (you) handles centrally:

1. Visual diff each new file.
2. Wire any shared aggregator/registration file (e.g. for compute LLK shims: SFPU_OP_*_INCLUDE branches in sfpu_split_includes.h).
3. Update .claude/references/structure.yaml with one line per new file (per CLAUDE.md project rule).
4. Build.
5. Run targeted tests for each new mock — this loop can itself parallelize via one pytest call per item.
6. Promote 100%-pass entries into the regression script (e.g. scripts/run_ttnn_pytests.sh for compute LLK shims).
7. For partial passes, decide per-item: pytest -k filter, or defer with a written note.

Anti-patterns

1. Don't have parallel workers touch shared files. Each worker owns a unique target path; the orchestrator centralizes wiring after the batch returns. Without this, you get racing writes that silently lose changes.
2. Don't skip the STUCK branch. When a worker hits ambiguity it must report STUCK with a reason; the orchestrator decides how to resolve (e.g. surface to the user, drop the item, hand-write it). Workers that "guess" produce subtle bugs that surface much later.
3. Don't bypass the schema. Free-text worker output is unparseable at scale. Even a STUCK item needs structured {status: STUCK, file: <intended path>, reason: <one sentence>}.
4. Concurrency cap is automatic. The Workflow runtime limits concurrent agent() calls (typically ≤16); excess queue up. Don't manually batch in chunks unless you have a reason — parallel() over the full work-list is fine.