Install Sepo agent infrastructure

AS
2d ago
+62857-0

Summary

Adds the Sepo agent backend to this repository so @sepo-agent can answer, implement, review, and fix work on GitHub. This PR adds agent infrastructure only — no application source, secrets, or branch protection are changed.

Installed from the latest stable Sepo release v0.5.0 (self-evolving/repo).

Required setup after merge

  1. Install the Sepo GitHub App on this repository, or choose another auth path from the setup guide.
  2. Add at least one model-provider secret under Actions secrets: OPENAI_API_KEY, CLAUDE_CODE_OAUTH_TOKEN, and/or ANTHROPIC_API_KEY.
  3. Run the Agent / Onboarding / Check Setup workflow.
  4. Review the Sepo setup check issue it opens and complete any remaining setup it reports.
  5. If agent/memory is missing, run Agent / Memory / Initialization.
  6. Optionally run Agent / Rubrics / Initialization if this repo wants rubric steering.

Also confirm Issues and GitHub Actions are enabled (Settings → General → Features, Settings → Actions → General).

Source revision

  • self-evolving/repo @ v0.5.0 (5f7d2032080e9c0578d0e4cb3b98b2f695f33547)

Installed files

  • .agent/ — agent backend (265 files), excluding generated .agent/dist/ and .agent/node_modules/.
  • .github/ — agent workflows (28), composite actions, prompts, issue templates, and assets, merged into the existing .github/.
  • .gitignore — appended .agent/dist/ and .agent/node_modules/.

Preserved / skipped

  • All target-owned .github/ files preserved; the conflict audit found no source path that collides with an existing target path.
  • Root README.md, AGENTS.md, and .skills/ were not copied (install defaults).
  • No application source changed.

Validation

  • npm ci, npm run build, and npm test pass on the source at v0.5.0 (run once against the identical payload).
  • git diff --cached --check is clean.
  • Conflict audit: no source .github path collides with an existing target path.

Skipped checks

  • Per-repo .agent build/test not rerun; the payload is byte-identical to the validated source revision.

This PR was created by an AI agent (OpenHands) on behalf of @ashcoft.


Devin Review


Summary by cubic

Installs the Sepo agent backend so @sepo-agent can answer, implement, review, and fix work on GitHub. This adds agent infrastructure only - no application source, secrets, or branch protection are changed.

The backend comes from self-evolving/repo at v0.5.0 and lands as .agent/ plus merged .github/ workflows, composite actions, prompts, issue templates, and assets. Generated outputs (.agent/dist/, .agent/node_modules/) are gitignored. Existing .github/ files are preserved; no source path collides with a target path.

Setup after merge

  • Install the Sepo GitHub App on this repository.
  • Add at least one model-provider secret (Actions secrets): OPENAI_API_KEY, CLAUDE_CODE_OAUTH_TOKEN, and/or ANTHROPIC_API_KEY.
  • Run the Agent / Onboarding / Check Setup workflow and complete any remaining setup it reports.
  • Run Agent / Memory / Initialization if agent/memory is missing.
  • Optionally run Agent / Rubrics / Initialization for rubric steering.
  • Confirm Issues and GitHub Actions are enabled in repository settings.

This change is Reviewable

RetriggerView in GreptileConfidence Score: 0/5

Fix approval request trust and the workflow failures before merging.

Fix All in CodexFindings

  1. P1 Security Approval starts the wrong work ▶
  2. P1 Older PRs cannot run fixes ▶
  3. P1 Overlapping runs lose memory ▶
  4. P1 Failed publishing looks successful ▶
Fix with agent prompt
### Issue 1
.agent/src/approval.ts:145-151
`findPendingRequestById` trusts approval markers from any comment. A public commenter can edit an earlier comment to copy a real request ID and replace its hidden target and request text. Since comments are scanned oldest first, a maintainer's approval then dispatches the forged request instead of the agent's request.

Keep comment authors and accept markers only from the trusted agent. Also bind the target to the current thread.

**How this was verified:** Comment loading drops the author, matching checks only the request ID, and the approval workflow dispatches the matched marker's target and text.

### Issue 2
.github/workflows/agent-fix-pr.yml:153-159
The workflow switches the runtime checkout to the PR branch before loading `run-agent-task`. An existing same-repository PR opened before this installation has no `.github/actions/run-agent-task/action.yml`, so `/fix-pr` fails before the agent starts. Implementation on an older stacked base has the same problem.

Keep the installed runtime checkout unchanged and put the target branch in a separate worktree, as the `add-rubrics` path already does.

### Issue 3
.github/actions/run-agent-task/action.yml:549-559
Memory writes have no retry after a rejected push. Two answers on different issues, or an answer overlapping the scheduled sync, can clone the same `agent/memory` head and both make edits. The first push succeeds; the second fails and its checkout is discarded after only a warning.

Retry by fetching and merging the latest branch before pushing. The shared rubric branch needs the same protection.

### Issue 4
.github/workflows/agent-implement.yml:341
Final results still use the status computed before publishing. If an implementation push or PR creation fails, the issue comment reports success. In `fix-pr`, the comment correctly reports a failed push, but the handoff still sends `success`. Orchestration can then review the unchanged remote branch instead of stopping.

Compute one final status after commit, push, and PR creation, and use it for both comments and handoffs.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Installs Sepo's runtime, GitHub workflows, prompts, documentation, and local-runner tools without changing application code.

  • GitHub requests now run through the Sepo agent workflows.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[GitHub mention or label] --> B[Router]
  B --> C[Answer]
  B --> D[Approval request]
  D --> E[Implementation]
  B --> E
  B --> F[Review or fix]
  E --> G[Verify and publish]
  F --> G
  G --> H[Comment and optional handoff]
  C --> M[Shared memory branch]
  E --> M
  F --> M

Reviews (1) · Last reviewed commit: "chore: install Sepo agent infrastructure" · Reviewed by Greptile