Triage Agent Corrections

# Triage Agent Corrections

This issue tracks corrections to the triage agent system. When assigned to
Copilot, analyze the corrections and generate an improvement PR.

## Instructions for Copilot

When assigned:
1. Read each linked correction comment and the original issue for full context
2. Identify patterns (e.g., the classifier frequently confuses X with Y)
3. Determine which workflow file(s) need improvement
4. Use the `agentic-workflows` agent in this repo for guidance on workflow syntax and conventions
5. Open a PR with targeted changes to the relevant `.md` workflow files in `.github/workflows/`
6. **If you changed the YAML frontmatter** (between the `---` markers) of any workflow, run `gh aw compile` and commit the updated `.lock.yml` files. Changes to the markdown body (instructions) do NOT require recompilation.
7. Reference this issue in the PR description using `Closes #`
8. Include a summary of which corrections motivated each change

## Corrections

| Issue | Feedback | Submitted by | Date |
|-------|----------|--------------|------|
| [#1027] workspace.yaml missing git related info | Agent claims to have confirmed a bug, but does not show actual executable repro code. It also says "CLI *likely*" did something different in a prior version, but hasn't objectively confirmed this. As such the agent's output might all be wrong. The agent shouldn't tell us what will likely happen - it should create actual repro code and tell us what actually happens. | @SteveSandersonMS | 2026-04-07 |
| [#1042] Copilot CLI release/version mismatch makes Copi... | The agent shouldn't say "the behavior observed is a defect" because it hasn't actually observed that behavior itself. The agent should only make claims based on what it has been able to reproduce itself. Also from the "what I found" list, neither items 2 or 3 are bugs in any sense. All we're left with is speculation from the agent. A better conclusion from the agent in this case would be that it doesn't know. | @SteveSandersonMS | 2026-04-09 |
| [#1050] uvicorn main:app --reload | This is a clear case where the agent should have closed the issue. There's no feature request or bug report. | @SteveSandersonMS | 2026-04-09 |
| [#1052] API review and ongoing API quality controls | In this case since it's a work tracking issue posted by one of the project maintainers, you don't need to comment about it at all. | @SteveSandersonMS | 2026-04-09 |
| [#1064] [Bug] BYOK Anthropic provider missing multiple... | Agent claims "This is a confirmed bug." but I don't think it's possible for the agent to confirm this since it doesn't have a BYOK key for Anthropic. Agent should not claim it's confirmed something unless it does so empirically. | @SteveSandersonMS | 2026-04-13 |
| [#1146] Test harness CLI v1.0.32 incompatible with perm... | The agent claims to have investigated the issue and proposes a fix, but its fix is wrong (changing that package version will not affect the e2e test execution). So the agent must not have actually run the tests for real. The agent must not make claims about what works or doesn't work except exclusively by empirical observation. | @SteveSandersonMS | 2026-04-28 |