Change the way of working with sepo

[lolipopshock]

1. before the review gate, can you generate a list of "decisions (implicit / explicit) made by the code implementation"?
2. in the review process, can you additionally judge the decisions --> are some of them hacky decisions?

@sepo-agent /answer

[sepo-agent-app]

Yes. I’d make this a first-class implementation-to-review handoff.

Today, the implementation prompt returns only summary and PR metadata. The review prompt already checks for hacky or brittle solutions, but it does not systematically evaluate implementation decisions.

Proposed behavior:

1. After coding, Sepo produces a structured list of material decisions:

   • Decision
   • Explicit or implicit
   • Rationale and trade-offs
   • Alternatives considered
   • Supporting files/symbols

2. Deterministic post-processing renders it as ## Implementation Decisions in the draft PR before review starts. Fix passes update the inventory when decisions change.

3. Reviewers independently validate the list against the current diff, add omitted decisions, and classify each as:

   • SOUND
   • PRAGMATIC_COMPROMISE
   • HACKY

   “Hacky” findings must include evidence, risk, and a more principled alternative; the label alone would not make a finding blocking.

4. The review synthesis includes a decision-review table and converts risky hacky decisions into normal findings and action items.

Implementation would touch the prompt contract, response parsing in response.ts, PR rendering, review prompts, tests, and matching documentation. Existing PRs without an inventory would remain reviewable—the reviewers would reconstruct it from the diff.

To proceed, send:

@sepo-agent /implement Add a structured implementation-decision inventory before review, keep it current across fix-pr runs, and require review synthesis to classify decisions as sound, pragmatic compromises, or hacky with evidence.

---

Sepo activity

Sepo finished — answer · 9m29s · 10 steps

Last message

Yes. I’d make this a first-class implementation-to-review handoff. Today, the implementation prompt returns only summary and PR metadata. The [review pro…

Activity

• 📖 Read "sed -n '1,260p' .github/prompts/agent-implement.md sed -n '1,240p' .github/prompts/review.md sed -n '1,220p' .github/p… (completed)
• 📖 Read rg -n "normalizeImplementationResponse|summaryFromAgentResponse|ImplementationResponse" .agent/src/response.ts .agent/s… (completed)
• 📖 Read sed -n '1,260p' .agent/src/cli/create-pr.ts rg -n "create-pr|PR_BODY|normalizeImplementationResponse|pr_body|agent-impl… (completed)
• 🔧 Used tool sed -n '120,165p' .agent/docs/usage/agent-actions.md sed -n '1,130p' .agent/docs/usage/agent-actions.md sed -n '350,405… (completed)
• 🔍 Searched "sed -n '1,220p' .github/prompts/agent-fix-pr.md rg -n \"Parse response|RESPONSE_FILE|pr_body|post-comment|fix\" .githu… (completed)
• 💬 Message "Yes. I’d make this a first-class implementation-to-review handoff. Today, the implementation prompt returns only summary and PR metadata. The [review pro…"

---

codex | gpt-5.6-sol[max] | Macmini-runner-2