Skip to content

feat(senpi-codemode): teach eval batching via instruction and reuse-chain examples#205

Merged
code-yeongyu merged 1 commit into
mainfrom
feat/eval-prompt-batch-guidance
Jul 14, 2026
Merged

feat(senpi-codemode): teach eval batching via instruction and reuse-chain examples#205
code-yeongyu merged 1 commit into
mainfrom
feat/eval-prompt-batch-guidance

Conversation

@code-yeongyu

@code-yeongyu code-yeongyu commented Jul 14, 2026

Copy link
Copy Markdown
Owner

What

Rework the eval tool prompt so batch-through-eval usage surfaces on its own instead of relying on operator steering.

  • Instruction: new **Batch through eval.** paragraph — one cell replaces many one-shot tool calls: loop/comprehend over file sets with read()/stdlib, post-process tool.<name>() results programmatically, fan independent calls through parallel(thunks). Absorbs the old "Parallelize work within a cell" sentence (replacement, not addition).
  • Examples: the three py reuse-chain slots now demonstrate comprehension+rglob collection → batch read() loop with Counterparallel() fan-out of tool.grep calls, while keeping the cell-to-cell state-reuse narrative intact. The two rb slots are dropped — rb defaults off, so the default profile never rendered them.
  • Dedup: the <critical> sentence duplicated verbatim from the instruction ("Re-run setup only after reset/crash/NameError") is deleted.

Net: prompt teaches harness-specific, non-derivable patterns (thunk shape with d=d binding, tool.<name> args dict, batch-read idiom) in the same slot count; file-mutation and bash payloads intentionally excluded to avoid steering around the edit tool or double-bridging shell.

QA

  • npx tsx vitest --run test/ in packages/senpi-codemode: 350 passed / 6 skipped (snapshots updated deliberately, reviewed line-by-line).
  • npm run check at repo root: passed (biome, docs, types, neo build+vet+test).
  • Prompt-only change to the eval tool description string; no runtime code path touched.

Summary by cubic

Teaches eval batching in senpi-codemode with a new instruction and Python reuse-chain examples showing batch reads and parallel tool calls. Removes Ruby examples and duplicate guidance; prompt-only change with updated tests.

  • New Features
    • Added "Batch through eval." instruction, replacing the old "parallelize within a cell" line.
    • Reworked three Python reuse-chain examples to show: file collection with rglob, batch read() + Counter, and parallel tool.grep fan-out using thunks.
    • Removed Ruby example slots and a duplicated critical sentence about re-running setup.

Written for commit 31936d8. Summary will update on new commits.

Review in cubic

…hain examples

Replace the json-config example payloads with a three-cell chain that
demonstrates comprehension filtering, batch read() loops, and parallel()
fan-out of tool.<name>() calls. Add a 'Batch through eval' instruction
absorbing the old parallelize sentence, drop the rb example slots
(rb defaults off), and delete the critical-block sentence duplicated
verbatim in the instruction.
@code-yeongyu code-yeongyu merged commit 7ab215f into main Jul 14, 2026
6 checks passed
@code-yeongyu code-yeongyu deleted the feat/eval-prompt-batch-guidance branch July 14, 2026 03:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant