Skip to content

fix: remove parallel_sub_agents experiment from smoke-pi workflow #37344

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
pelikhan merged 2 commits into from
Jun 6, 2026
Merged

fix: remove parallel_sub_agents experiment from smoke-pi workflow #37344

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
feb6e22
Initial plan
Copilot Jun 6, 2026
dd7e10d
fix: remove parallel_sub_agents experiment from smoke-pi workflow
Copilot Jun 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
126 changes: 1 addition & 125 deletions .github/workflows/smoke-pi.lock.yml
View file
Open in desktop

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

31 changes: 0 additions & 31 deletions .github/workflows/smoke-pi.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,24 +17,6 @@ permissions:
issues: read
pull-requests: read
name: Smoke Pi
experiments:
sub_agent_decomposition:
variants: [single_agent, parallel_sub_agents]
description: "Test whether decomposing smoke tests into parallel sub-agents reduces token cost"
hypothesis: "H0: no change in effective token consumption. H1: parallel sub-agents reduce tokens by 15-25% by eliminating unnecessary context sharing"
metric: effective_token_count
secondary_metrics: [run_duration_seconds, test_pass_rate, false_failure_rate]
guardrail_metrics:
- name: test_completion_rate
threshold: ">=0.95"
- name: overall_pass_rate
threshold: ">=0.80"
min_samples: 20
weight: [50, 50]
start_date: "2026-05-22"
analysis_type: mann_whitney
tags: [cost_optimization, smoke_tests, pi_engine]
# issue: PLACEHOLDER_ISSUE_NUMBER
engine:
id: pi

@github-actions github-actions Bot Jun 6, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[/diagnose] The id: pi config gives no indication that Pi runs in single-pass mode (--print --mode json --no-session) and therefore cannot support experiments requiring task/read_agent tools. Without this being documented here, the next engineer authoring an experiment for this workflow faces the same trap.

💡 Suggested addition

Add a comment directly below id: pi:

engine:
  id: pi
  # Pi runs in single-pass mode (--print --mode json --no-session).
  # It has no task/read_agent tools and cannot receive async notifications.
  # Experiments MUST NOT require background sub-agents.
  model: copilot/gpt-5.4

This makes the constraint discoverable at the point where future experiments are authored rather than only surfacing in production failures.

model: copilot/gpt-5.4
Expand Down Expand Up @@ -96,26 +78,13 @@ timeout-minutes: 10

## Test Requirements

{{#if experiments.sub_agent_decomposition == 'parallel_sub_agents'}}
Launch five parallel `task` agents using mode: "background" to execute each smoke test independently. Use the `task` agent type with `description` field for each:

1. **GitHub MCP Test Agent**: Fetch 2 merged PR titles from ${{ github.repository }}
2. **Web Fetch Test Agent**: Fetch https://github.com and verify "GitHub" in response using web-fetch MCP
3. **File I/O Test Agent**: Create `/tmp/gh-aw/agent/smoke-test-pi-${{ github.run_id }}.txt` with timestamp
4. **Bash Test Agent**: Verify file creation with `cat` command
5. **Build Test Agent**: Run `GOCACHE=/tmp/gh-aw/agent/go-cache GOMODCACHE=/tmp/gh-aw/agent/go-mod make build`

Wait for all five agents to complete (you'll receive notifications). Read each agent's result using `read_agent`. Aggregate the results into a unified report with ✅/❌ status for each test.

{{else}}
Execute the following tests sequentially in a single turn:

@github-actions github-actions Bot Jun 6, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[/tdd] The bug was caught in production (~50% failure rate, ~17K tokens wasted per run). There is no regression test or compile-time check to detect if someone re-introduces an incompatible experiment variant for Pi.

💡 Possible guard options
  1. Schema validation in gh aw compile: Reject experiment variants that reference task/read_agent when targeting the pi engine.
  2. CI lint rule: A test that parses smoke-pi.md and asserts no experiments: block uses parallel_sub_agents or any variant whose prompt body contains mode: "background".
  3. Guardrail comment in the .md source (as suggested on the engine: block): Even a prose comment near experiments: documentation noting the Pi constraint would reduce rediscovery cost.

Option 1 is most robust; Option 3 is the lowest-friction starting point.


1. **GitHub MCP Testing**: Use GitHub MCP tools to fetch details of exactly 2 merged pull requests from ${{ github.repository }} (title and number only)
2. **Web Fetch Testing**: Use the web-fetch MCP tool to fetch https://github.com and verify the response contains "GitHub" (do NOT use bash or playwright for this test - use the web-fetch MCP tool directly)
3. **File Writing Testing**: Create a test file `/tmp/gh-aw/agent/smoke-test-pi-${{ github.run_id }}.txt` with content "Smoke test passed for Pi at $(date)" (create the directory if it doesn't exist)
4. **Bash Tool Testing**: Execute bash commands to verify file creation was successful (use `cat` to read the file back)
5. **Build gh-aw**: Run `GOCACHE=/tmp/gh-aw/agent/go-cache GOMODCACHE=/tmp/gh-aw/agent/go-mod make build` to verify the agent can successfully build the gh-aw project. If the command fails, mark this test as ❌ and report the failure.
{{/if}}

## Output

Expand Down
Loading