fix(grillme): require a recommended answer on inline questions

[johnmcollier]

Summary

Grillme's inline threads on #7 asked good questions but never took a stance. The original one-question model always included Recommended answer; that was lost in the thread rewrite.

This restores it:

• Each new inline comment is question + **Recommended:** <answer and why>
• Follow-up thread replies that still ask a decision question do the same
• Omit the recommendation only when there is genuinely no technical default (a values/priority call)

Prompt + grilling skill both updated. Example comment bodies in the agent output contract now include the recommendation line.

After merge, a new /fs-grillme session on a PR will pick this up (workflow reads main). Existing threads on #7 will not be rewritten; start a fresh session or continue and look for recommendations on new/follow-up comments.

Test plan

• Merge this PR
• /fs-grillme on docs: OpenSpec for task status workflow #7 (or a new OpenSpec PR)
• Confirm each inline question has a Recommended: line

Made with Cursor

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 7:22 PM UTC · Completed 7:35 PM UTC

Commit: 88b3ce5 · View workflow run →

[fullsend-ai-review]

Review

Findings

Low

• [internal consistency] .fullsend/custom/skills/grilling/SKILL.md:30 — SKILL.md step 2 ("Reply in-thread when an answer needs a follow-up") is not updated to mention including **Recommended:** on follow-up replies, while grillme.md explicitly adds this rule at lines 225–226: "When the follow-up is itself a decision question, include **Recommended:** the same way as new comments." The two documents now diverge on follow-up behavior. In practice the agent definition (grillme.md) is the primary runtime document and does contain the instruction, so the risk is low.
  Remediation: Add a clause to SKILL.md step 2 such as: "When a follow-up is itself a decision question, include **Recommended:** <answer and why> the same way as new comments."

[fullsend-ai-review]

Note: The following inline comments could not be posted on the diff (GitHub returned 422) and are included here instead:

• .fullsend/custom/skills/grilling/SKILL.md (file-level): Line 30 · [low] internal consistency

SKILL.md step 2 ('Reply in-thread when an answer needs a follow-up') is not updated to mention including Recommended: on follow-up replies, while grillme.md explicitly adds this rule at lines 225-226. The two documents now diverge on follow-up behavior. In practice the agent definition (grillme.md) is the primary runtime document and does contain the instruction, so the risk is low.

Suggested fix: Add a clause to SKILL.md step 2 such as: 'When a follow-up is itself a decision question, include Recommended: the same way as new comments.'

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 7:41 PM UTC · Completed 7:48 PM UTC

Commit: 88b3ce5 · View workflow run →

[fullsend-ai-retro]

Retro: PR #9 — fix(grillme): require a recommended answer on inline questions

What happened

PR #9 was a human-authored fix (johnmcollier, with Cursor) to the grillme agent configuration. It restored the **Recommended:** answer requirement on inline questions after real-world testing on PR #7 showed the grillme agent asked good probing questions but never took a stance — the author even replied "I'm not sure, what do you think?" expecting the agent to have an opinion.

The review agent (run 32518019927, fullsend v0.36.0, agents from fullsend-ai/agents@6bdcab69) ran four sub-agents (correctness, style-conventions, intent-coherence, docs-currency) plus an adversarial challenger. It approved the PR with one low-severity finding: SKILL.md step 2 was not updated to mention **Recommended:** on follow-up replies, creating a consistency gap with grillme.md. The PR was merged ~19 minutes after creation without addressing this finding.

Workflow quality

• Review quality: Good. The review agent caught a real consistency gap, rated it appropriately as low-severity, and produced zero false positives.
• Rework rate: Zero. Single commit, single review iteration, merged. The unaddressed finding creates a small latent consistency issue.
• Token cost: Slightly elevated. The sonnet model was unavailable on GCP Vertex, forcing 3 of 4 sub-agents to re-dispatch on opus. This added ~3 minutes and higher per-token cost. The agent handled the fallback gracefully.
• Time to resolution: 19 minutes from PR creation to merge, with ~14 minutes spent on the review agent.

Observations on existing tracked issues

• Sonnet model fallback: The model unavailability is already tracked in fullsend-ai/agents#379 and fullsend-ai/fullsend#1771. This run provides additional evidence — 3 sub-agents were re-dispatched, adding ~3 minutes.
• 422 inline comment posting: The review agent's inline comment failed with a 422 error and was included in the review body instead. This is extensively tracked across fullsend-ai/agents#760, #699, #430, and fullsend-ai/fullsend#6039. No new proposal needed.

What went well

The overall workflow was healthy: real-world testing (grillme on PR #7) surfaced a UX gap, a human authored a targeted fix, and the review agent provided useful feedback. The review agent correctly identified the one remaining consistency issue in a 2-file, 24-line change to agent configuration files — solid performance on a meta-review task (reviewing agent definitions).

Proposals filed

• Update SKILL.md step 2 to require Recommended on follow-up replies (in rhdh-parasol/node-api)