Create agent for user-driven, complete reviews. #10829

New issue

Closed

opened 2026-04-22 21:05:17 +00:00 by brent.edwards · 1 comment

brent.edwards commented

2026-04-22 21:05:17 +00:00

Member

Metadata

Commit Message: feat(agents): add user-driven review agent
Branch: feature/m7-user-driven-review-agent

Background and Context

The project has an existing pr-reviewer.md agent that performs fully automated,
bot-driven code reviews without human involvement. There is currently no agent that
assists a human reviewer — one who wants to conduct a structured review themselves
but needs the LLM to read the code, surface findings, ask clarifying questions, and
ultimately submit a formal Forgejo decision (approve or request changes) on their
behalf.

This issue creates human-reviewer.md in .opencode/agents/: a human-facing,
interactive review agent that embeds the project's review playbook criteria and
guides the reviewer through a structured evaluation.

Expected Behavior

A human invokes the agent via OpenCode. The agent:

Reads the PR diff, changed files, linked issue, and all prior review comments.
Evaluates the code against the embedded review criteria (focus areas, checklists,
and priority matrix from docs/development/review_playbook.md).
Tags each finding with a severity label in the conversation output: P0:blocker,
P1:must-fix, P2:should-fix, or P3:nit.
Asks the human clarifying questions about the code where context is ambiguous.
Posts a formal Forgejo review decision — APPROVED or REQUEST_CHANGES —
based on the findings and the human's input.

Acceptance Criteria

The file /app/.opencode/agents/human-reviewer.md exists.
The agent is mode: agent (human-callable directly from OpenCode).
The agent is permitted to ask the human questions (question: allow).
The agent reads the PR diff, changed files, linked issue, and prior comments
before beginning its review.
The agent tags every finding with a severity level (P0–P3) in its output.
The agent can post a formal Forgejo review (APPROVED or REQUEST_CHANGES) via
curl using credentials provided in the conversation.
The agent does not need Behave/Robot tests or a separate YAML config file —
the agent file itself is the deliverable.

Supporting Information

Existing automated reviewer for reference:
.opencode/agents/pr-reviewer.md
Review criteria source:
docs/development/review_playbook.md
⚠️ No parent Epic linked yet. A parent Epic must be added once identified.

Subtasks

Create /app/.opencode/agents/human-reviewer.md
Write the frontmatter: mode: agent, permissions (allow question,
sequential-thinking*, read-only Forgejo tools, curl * for posting
the formal review; deny everything else)
Embed key review criteria from review_playbook.md directly in the agent
instructions: focus areas, skip rules, P0–P3 priority matrix,
per-type checklists (architecture, CLI, DB migration, security),
escalation rules, and routing table
Write instructions for reading prior comments and PR diff before reviewing
Write instructions for tagging each finding P0–P3 in the conversation output
Write instructions for asking the human clarifying questions when context
is ambiguous
Write instructions for posting a formal Forgejo APPROVED or REQUEST_CHANGES
review via curl

Definition of Done

This issue is complete when:

The file /app/.opencode/agents/human-reviewer.md is created and the agent
is functional when invoked from OpenCode.
A Git commit is created where the first line matches the Commit Message in
Metadata exactly: feat(agents): add user-driven review agent
The commit is pushed to feature/m7-user-driven-review-agent.
The commit is submitted as a PR to master, reviewed, and merged.

Automated by CleverAgents Bot
Agent: new-issue-creator

## Metadata **Commit Message:** `feat(agents): add user-driven review agent` **Branch:** `feature/m7-user-driven-review-agent` ## Background and Context The project has an existing `pr-reviewer.md` agent that performs fully automated, bot-driven code reviews without human involvement. There is currently no agent that assists a *human* reviewer — one who wants to conduct a structured review themselves but needs the LLM to read the code, surface findings, ask clarifying questions, and ultimately submit a formal Forgejo decision (approve or request changes) on their behalf. This issue creates `human-reviewer.md` in `.opencode/agents/`: a human-facing, interactive review agent that embeds the project's review playbook criteria and guides the reviewer through a structured evaluation. ## Expected Behavior A human invokes the agent via OpenCode. The agent: 1. Reads the PR diff, changed files, linked issue, and all prior review comments. 2. Evaluates the code against the embedded review criteria (focus areas, checklists, and priority matrix from `docs/development/review_playbook.md`). 3. Tags each finding with a severity label in the conversation output: `P0:blocker`, `P1:must-fix`, `P2:should-fix`, or `P3:nit`. 4. Asks the human clarifying questions about the code where context is ambiguous. 5. Posts a formal Forgejo review decision — **APPROVED** or **REQUEST_CHANGES** — based on the findings and the human's input. ## Acceptance Criteria - The file `/app/.opencode/agents/human-reviewer.md` exists. - The agent is `mode: agent` (human-callable directly from OpenCode). - The agent is permitted to ask the human questions (`question: allow`). - The agent reads the PR diff, changed files, linked issue, and prior comments before beginning its review. - The agent tags every finding with a severity level (`P0`–`P3`) in its output. - The agent can post a formal Forgejo review (APPROVED or REQUEST_CHANGES) via curl using credentials provided in the conversation. - The agent does **not** need Behave/Robot tests or a separate YAML config file — the agent file itself is the deliverable. ## Supporting Information - Existing automated reviewer for reference: [`.opencode/agents/pr-reviewer.md`](.opencode/agents/pr-reviewer.md) - Review criteria source: [`docs/development/review_playbook.md`](docs/development/review_playbook.md) - ⚠️ **No parent Epic linked yet.** A parent Epic must be added once identified. ## Subtasks - [ ] Create `/app/.opencode/agents/human-reviewer.md` - [ ] Write the frontmatter: `mode: agent`, permissions (allow `question`, `sequential-thinking*`, read-only Forgejo tools, `curl *` for posting the formal review; deny everything else) - [ ] Embed key review criteria from `review_playbook.md` directly in the agent instructions: focus areas, skip rules, P0–P3 priority matrix, per-type checklists (architecture, CLI, DB migration, security), escalation rules, and routing table - [ ] Write instructions for reading prior comments and PR diff before reviewing - [ ] Write instructions for tagging each finding P0–P3 in the conversation output - [ ] Write instructions for asking the human clarifying questions when context is ambiguous - [ ] Write instructions for posting a formal Forgejo APPROVED or REQUEST_CHANGES review via curl ## Definition of Done This issue is complete when: - The file `/app/.opencode/agents/human-reviewer.md` is created and the agent is functional when invoked from OpenCode. - A Git commit is created where the first line matches the Commit Message in Metadata **exactly**: `feat(agents): add user-driven review agent` - The commit is pushed to `feature/m7-user-driven-review-agent`. - The commit is submitted as a PR to `master`, reviewed, and merged. --- **Automated by CleverAgents Bot** Agent: new-issue-creator

Rows
Columns