issue-299: Verify the STE style live in a short /grill session and record the observations #309

Merged

david merged 2 commits from feature/issue-299/verify-the-ste-style-live-in-a-short-grill-session into main

2026-10-09 22:40:43 +00:00

david commented

2026-10-09 22:40:43 +00:00

Owner

299: Verify the STE style live in a short /grill session and record the observations

Closes #299 — 299.

This pull request was generated automatically by pi-loop and opened as a
Draft for human review. Do not merge without reviewing the changes.

Generated in 10m 12s.

Review findings

I'll start by examining the repository state and the relevant files.Confirmed the transcript facts: only two typed user answers (plus the /grill start), and session model is deepseek-flash vs config deepseek-v4-flash. Now applying the fixes.Now running the project's test gate (the only configured gate — no build/lint scripts exist).All fixes applied and the test gate is green (1474 pass, 1 skipped, 0 fail). The commit-subject finding (AGENTS.md:26) cannot be fixed here — amending a commit is explicitly forbidden for this stage.

Review Findings — 299

Summary

Docs-only change: it appends a "Live STE verification" section (~161 lines) to extensions/grillme/findings.md, recording a /grill session checked against the ten rule categories in ste-style.ts. Gates: npm test was run and passed (1474 pass, 1 skipped, 0 fail); no build or lint gate is configured in package.json (no tsc, eslint, biome, or prettier config exists). I verified the quoted sentences and rule counts against the actual pi session transcript (~/.pi/agent/sessions/…/2026-10-09T22-31-20-672Z_01a122ca-…jsonl); the quotes and the 30/27/25/22-word counts are accurate. The remaining issues are in how the experiment itself is described, and one acceptance-criterion gap.

Remediation: fixed the High finding (turn/answer count corrected against the transcript), the Medium finding (the Verdict now states the M7 acceptance explicitly as not met and separates the product judgement from it), and the Low finding on the model alias (the unsupported parenthetical is dropped). The remaining Low finding is the commit subject in AGENTS.md:26; this stage may not commit, stage, or amend, so it stays unresolved and belongs to the commit/PR step.

Critical

High

extensions/grillme/findings.md:92 — the record says "Three typed answers produced the three assistant replies", and line 81 says the replies were "one per user answer", but the session transcript holds only two typed user answers after the /grill start message (the first assistant turn answered the injected /grill start, not a user answer). This misdescribes the experiment and makes the run non-reproducible. Suggested fix: change line 92 to "Two typed answers produced the second and third replies; the first reply answered the /grill start message." and change line 81 to "one per assistant turn".

Medium

extensions/grillme/findings.md:225 — the Verdict concludes "Usable as shipped: yes" while the same verdict lists seven of the ten rule categories as not held (rules 1, 2, 3, 6, 7, 9, 10), yet it never states the outcome of the M7 acceptance criterion, which requires that "the observed replies and the checkpoint follow the rule categories except in the exempt parts" (docs/GRILL_ME_STE_STYLE_PLAN.md:221). As recorded, that criterion is not met. Suggested fix: add an explicit line to the Verdict stating the M7 acceptance is not met (and which rules failed it), and separate the "usable as shipped" product judgement from the acceptance result.

Low

AGENTS.md:26 — the Stage 1 commit subject is 299: Verify the STE style live in a short /grill session and record the observations, but for a feature/enhancement issue-driven change the standard requires issue-<N>: <summary>, which docs/GRILL_ME_STE_STYLE_PLAN.md:208 repeats. Suggested fix: amend the commit subject to issue-299: Verify the STE style live in a short /grill session and record the observations (and use the same issue-299: prefix for the PR title).
extensions/grillme/findings.md:78 — the parenthetical "(the model that pi-loop runs)" is not supported by the repository: the session transcript records provider: deepseek, modelId: deepseek-flash, while the repo's piloop-config.yaml names model: deepseek-v4-flash. Suggested fix: drop the parenthetical, or state the alias mapping and where it is defined.

## 299: Verify the STE style live in a short /grill session and record the observations Closes #299 — [299](https://git.excelera.net/david/pi-extensions-and-skills/issues/299). > This pull request was generated automatically by pi-loop and opened as a > **Draft** for human review. Do not merge without reviewing the changes. Generated in 10m 12s. ## Review findings I'll start by examining the repository state and the relevant files.Confirmed the transcript facts: only two typed user answers (plus the `/grill` start), and session model is `deepseek-flash` vs config `deepseek-v4-flash`. Now applying the fixes.Now running the project's test gate (the only configured gate — no build/lint scripts exist).All fixes applied and the test gate is green (1474 pass, 1 skipped, 0 fail). The commit-subject finding (AGENTS.md:26) cannot be fixed here — amending a commit is explicitly forbidden for this stage. # Review Findings — 299 ## Summary Docs-only change: it appends a "Live STE verification" section (~161 lines) to `extensions/grillme/findings.md`, recording a `/grill` session checked against the ten rule categories in `ste-style.ts`. Gates: `npm test` was run and passed (1474 pass, 1 skipped, 0 fail); no build or lint gate is configured in `package.json` (no `tsc`, eslint, biome, or prettier config exists). I verified the quoted sentences and rule counts against the actual pi session transcript (`~/.pi/agent/sessions/…/2026-10-09T22-31-20-672Z_01a122ca-…jsonl`); the quotes and the 30/27/25/22-word counts are accurate. The remaining issues are in how the experiment itself is described, and one acceptance-criterion gap. **Remediation:** fixed the High finding (turn/answer count corrected against the transcript), the Medium finding (the Verdict now states the M7 acceptance explicitly as not met and separates the product judgement from it), and the Low finding on the model alias (the unsupported parenthetical is dropped). The remaining Low finding is the commit subject in AGENTS.md:26; this stage may not commit, stage, or amend, so it stays unresolved and belongs to the commit/PR step. ## Critical ## High - [x] extensions/grillme/findings.md:92 — the record says "Three typed answers produced the three assistant replies", and line 81 says the replies were "one per user answer", but the session transcript holds only **two** typed user answers after the `/grill` start message (the first assistant turn answered the injected `/grill` start, not a user answer). This misdescribes the experiment and makes the run non-reproducible. Suggested fix: change line 92 to "Two typed answers produced the second and third replies; the first reply answered the `/grill` start message." and change line 81 to "one per assistant turn". ## Medium - [x] extensions/grillme/findings.md:225 — the Verdict concludes "**Usable as shipped:** yes" while the same verdict lists seven of the ten rule categories as not held (rules 1, 2, 3, 6, 7, 9, 10), yet it never states the outcome of the M7 acceptance criterion, which requires that "the observed replies and the checkpoint follow the rule categories except in the exempt parts" (`docs/GRILL_ME_STE_STYLE_PLAN.md:221`). As recorded, that criterion is not met. Suggested fix: add an explicit line to the Verdict stating the M7 acceptance is not met (and which rules failed it), and separate the "usable as shipped" product judgement from the acceptance result. ## Low - [ ] AGENTS.md:26 — the Stage 1 commit subject is `299: Verify the STE style live in a short /grill session and record the observations`, but for a `feature`/`enhancement` issue-driven change the standard requires `issue-<N>: <summary>`, which `docs/GRILL_ME_STE_STYLE_PLAN.md:208` repeats. Suggested fix: amend the commit subject to `issue-299: Verify the STE style live in a short /grill session and record the observations` (and use the same `issue-299:` prefix for the PR title). - [x] extensions/grillme/findings.md:78 — the parenthetical "(the model that pi-loop runs)" is not supported by the repository: the session transcript records `provider: deepseek`, `modelId: deepseek-flash`, while the repo's `piloop-config.yaml` names `model: deepseek-v4-flash`. Suggested fix: drop the parenthetical, or state the alias mapping and where it is defined.

david added 2 commits

2026-10-09 22:40:43 +00:00