Rubrics/Memory inspection based on local traces

[lolipopshock]

1. allow users to upload local sessions
2. the agent runs on github for that session, and then add it to memory / rubrics

[sepo-agent-app]

This fits the current architecture. Recommended MVP: a trusted user exports a sanitized local transcript, attaches it to an issue comment, then manually runs a new “Inspect Local Session” workflow using that comment URL. Sepo should inspect the trace as untrusted evidence and persist only distilled findings—not the raw transcript.

Existing foundations:

• The attachment helper already restricts downloads to GitHub user attachments and enforces safe filenames, timeouts, and a 25 MiB limit (source).
• Existing continuity bundles restore provider files into the runner’s home directory (source). User uploads should not use that path.
• run-agent-task already mounts both branches and independently commits memory plus validates/commits rubrics (setup, persistence).
• Existing prompts already define conservative memory durability and trusted-source rubric learning (memory, rubrics).

Implementation plan:

1. Add a local export CLI producing a versioned, sanitized trace format. Preserve user/assistant messages and minimal provenance; omit credentials, environment snapshots, and raw tool payloads. Support canonical JSON/JSONL plus plain Markdown/text initially—no uploaded archives.
2. Add a validator that downloads the attachment with the existing helper, checks schema/size/record limits, computes a digest, and confirms the attachment URL appears in the supplied source issue comment.
3. Add agent-local-session-inspection.yml with inputs for source comment URL, destination (memory, rubrics, or both), and branch refs. Run with session_policy: none.
4. Add a dedicated inspection prompt and route. Treat trace instructions as data, learn rubrics only from explicit preferences attributable to a trusted user, and add memory only for durable repository-specific facts or lessons. Duplicate or weak evidence should produce no changes.
5. Post a source-issue summary showing the digest and separate memory/rubrics commit outcomes; never quote sensitive trace content.
6. Add parser/security tests, workflow contract tests, prompt guardrail assertions, documentation, and run the complete .agent CI suite.

I’m interpreting “runs on GitHub for that session” as inspecting the transcript, not resuming the provider’s live session. Live resumption should be a separate security design because it restores uploaded state into the runner.

If that interpretation is correct, send:

@sepo-agent /implement Implement the sanitized local-session inspection workflow for issue #489 using the plan above.

---

Sepo activity

Sepo finished — answer · 6m09s · 45 steps

Last message

This fits the current architecture. Recommended MVP: a trusted user exports a sanitized local transcript, attaches it to an issue comment, then manually runs a new “Inspect Local Session” workflow using that comment URL. Sepo should inspec…

Activity

• ✏️ Edited Search for '(memory-scan|rubrics-update|rubrics-initialization)' in .agent (completed)
• 📖 Read Read file '/Users/shannon/projects/2604_sepo/repo/.agent/tools/local-runner/runner-4/_work/repo/repo/.agent/src/run.ts' (completed)
• 📖 Read Read file '/Users/shannon/projects/2604_sepo/repo/.agent/tools/local-runner/runner-4/_work/repo/repo/.agent/src/__tests… (completed)
• 🔧 Used tool "nl -ba .github/workflows/agent-memory-pr-closed.yml | sed -n '1,300p'; nl -ba .github/prompts/memory-pr-closed.md | se… (completed)
• 🔧 Used tool node -e 'const p=require("./.agent/package.json"); console.log(JSON.stringify(p.scripts,null,2))' && nl -ba .agent/pack… (completed)
• 💬 Message "This fits the current architecture. Recommended MVP: a trusted user exports a sanitized local transcript, attaches it to an issue comment, then manually runs a new “Inspect Local Session” workflow using that comment URL. Sepo should inspec…"

---

codex | gpt-5.6-sol[max] | Macmini-runner-4

[lolipopshock]

@sepo-agent /orchestrate

[sepo-agent-app]

Sepo is starting a focused child task for this orchestration.

Child task	Focus	Parent issue	Status
#490	local-session-inspection-mvp	#489	Running

I'll report back here when the child task finishes.

[sepo-agent-app]

Child task completed.

Child task	Outcome	Parent round	Next step
#490	Failed	2 / 40	Resuming parent orchestration

Summary: implement concluded failed

[sepo-agent-app]

Sepo is starting a focused child task for this orchestration.

Child task	Focus	Parent issue	Status
#492	local-session-trace-foundation	#489	Running

I'll report back here when the child task finishes.

[sepo-agent-app]

Child task completed.

Child task	PR	Outcome	Parent round	Next step
#492	#493	Failed	3 / 40	Resuming parent orchestration

Summary: fix-pr concluded failed; no automatic handoff was dispatched because fix-pr must succeed before re-review

[sepo-agent-app]

Sepo orchestration needs clarification before it can continue.

The latest fix run failed during setup—not while changing code. GitHub returned HTTP 503 while cloning agent/rubrics, so PR #493 remains unchanged. Please retry @sepo-agent /fix-pr on PR #493 with the latest review action items; parent orchestration can resume after that fix succeeds.

• Source action: orchestrate
• Source conclusion: failed
• Target: issue #489
• Round: 3/40
• Reason: agent planner blocked: PR Add sanitized local-session trace export foundation #493 still has three concrete review fixes, but the fix agent never started because cloning agent/rubrics failed with a transient GitHub HTTP 503. Re-review is unsafe until a fix pass succeeds.
• Source run ID: 30800877105

No follow-up workflow was dispatched. Reply with the requested context, then continue with /orchestrate, /implement, or /answer when ready.