Fix the engine config for Daily Issues Report Generator: it is pinned to the sample Copilot Python SDK driver (.github/drivers/copilot_sdk_driver_sample_python.py), which the sandbox cannot run — python3 is absent and the copilot Python package is not installed — so every scheduled run hard-fails with zero tokens and produces no report. This has kept the workflow red for 7+ consecutive days.
Parent: #46170
Problem statement
The agent step (Execute GitHub Copilot CLI) never reaches a single turn. The copilot harness is invoked as a Python SDK driver:
copilot_harness.cjs python3 "${GITHUB_WORKSPACE}/.github/drivers/copilot_sdk_driver_sample_python.py" /usr/local/bin/copilot
Two blocking faults, both deterministic:
[copilot-harness] pre-flight: command not found: python3 (F_OK check failed — binary does not exist at this path)
...
Traceback (most recent call last):
File ".github/drivers/copilot_sdk_driver_sample_python.py", line 8, in <module>
from copilot import CopilotClient, RuntimeConnection
ModuleNotFoundError: No module named 'copilot'
The harness classifies this failureClass=partial_execution (not CAPI/model/auth) and burns all 4 attempts identically (tokenCount=0, retriesRemaining=0), so the step exits 1. This is distinct from #46326 (CAPI 400 claude-haiku-4.5 on Responses API) and #46172 (chroot EACCES) — here the agent process dies at import time, before any Copilot API call.
Affected workflows & run IDs
- Daily Issues Report Generator (
.github/workflows/daily-issues-report.lock.yml), engine copilot, model claude-sonnet-4.6
- Representative failed run: §29691397935 (2026-07-19T14:43Z) — 2 firewall requests, 0 tokens, 0 turns
- Comparator (prior failed run,
audit-diff): §29648548749 (2026-07-18) — identical: 0 tokens, ~10m duration, no drift
- Persistent streak (all
failure): 29590056273 (07-17), 29509837477 (07-16), 29426288647 (07-15), 29343594309 (07-14), 29262028117 (07-13)
Probable root cause
The workflow's engine/driver config points at the sample Python SDK driver (copilot_sdk_driver_sample_python.py) instead of the standard Copilot CLI path. The sandbox image provides neither a python3 on PATH inside the chroot nor the copilot Python SDK package, so the driver cannot import. Because it is a config/environment mismatch, retries cannot recover it.
Proposed remediation
- Repoint Daily Issues Report Generator off the sample Python SDK driver back to the standard Copilot CLI driver (drop the
copilot_sdk_driver_sample_python.py invocation), then recompile the lock file.
- If the Python SDK driver is intended, provision
python3 in the sandbox PATH and pip install the copilot SDK package in the driver's environment before invocation.
- Classify
ModuleNotFoundError / missing-interpreter pre-flight (command not found: python3) as non-retriable so it fails fast instead of burning 4 attempts and 10 minutes of runner time.
- Add a compile-time check that rejects a driver path whose required runtime/package is not present in the target image.
Success criteria / verification
- Re-run Daily Issues Report Generator: the agent reaches ≥1 turn with
tokenCount > 0 and emits its report (no ModuleNotFoundError: No module named 'copilot', no command not found: python3).
- The workflow returns green on
main for 2+ consecutive scheduled runs.
- No
failureClass=partial_execution with tokenCount=0 across all 4 harness attempts.
Analyzed run IDs: 29691397935, 29648548749. Related failure window: 2026-07-19 (last 6h).
Related to #46170
Generated by 🔍 [aw] Failure Investigator (6h) · 300.9 AIC · ⌖ 37.3 AIC · ⊞ 5.2K · ◷
Fix the engine config for Daily Issues Report Generator: it is pinned to the sample Copilot Python SDK driver (
.github/drivers/copilot_sdk_driver_sample_python.py), which the sandbox cannot run —python3is absent and thecopilotPython package is not installed — so every scheduled run hard-fails with zero tokens and produces no report. This has kept the workflow red for 7+ consecutive days.Parent: #46170
Problem statement
The agent step (
Execute GitHub Copilot CLI) never reaches a single turn. The copilot harness is invoked as a Python SDK driver:Two blocking faults, both deterministic:
The harness classifies this
failureClass=partial_execution(not CAPI/model/auth) and burns all 4 attempts identically (tokenCount=0,retriesRemaining=0), so the step exits 1. This is distinct from #46326 (CAPI 400claude-haiku-4.5on Responses API) and #46172 (chroot EACCES) — here the agent process dies at import time, before any Copilot API call.Affected workflows & run IDs
.github/workflows/daily-issues-report.lock.yml), enginecopilot, modelclaude-sonnet-4.6audit-diff): §29648548749 (2026-07-18) — identical: 0 tokens, ~10m duration, no driftfailure): 29590056273 (07-17), 29509837477 (07-16), 29426288647 (07-15), 29343594309 (07-14), 29262028117 (07-13)Probable root cause
The workflow's engine/driver config points at the sample Python SDK driver (
copilot_sdk_driver_sample_python.py) instead of the standard Copilot CLI path. The sandbox image provides neither apython3on PATH inside the chroot nor thecopilotPython SDK package, so the driver cannot import. Because it is a config/environment mismatch, retries cannot recover it.Proposed remediation
copilot_sdk_driver_sample_python.pyinvocation), then recompile the lock file.python3in the sandbox PATH andpip installthecopilotSDK package in the driver's environment before invocation.ModuleNotFoundError/ missing-interpreter pre-flight (command not found: python3) as non-retriable so it fails fast instead of burning 4 attempts and 10 minutes of runner time.Success criteria / verification
tokenCount > 0and emits its report (noModuleNotFoundError: No module named 'copilot', nocommand not found: python3).mainfor 2+ consecutive scheduled runs.failureClass=partial_executionwithtokenCount=0across all 4 harness attempts.Analyzed run IDs: 29691397935, 29648548749. Related failure window: 2026-07-19 (last 6h).
Related to #46170