fix(ci) - Agent v1 shell command CI break

[Siddhant-K-code]

Summary

Fixes the current CI break caused by Agent v1 shell hook parsing, plus the follow-on CI failures exposed after that compile fix.

PreBashCall and PostBashCall require a command field, but the Agent v1 preset only stored shell commands in context metadata. That made main fail to build before tests could run.

Main CI context

Recent main merges around the break:

• fix: resolve nightly agent test failures for droid and opencode #1725: fix: resolve nightly agent test failures for droid and opencode — green baseline before the break.
• Add agent-v1 shell command attribution #1636: Add agent-v1 shell command attribution — first merge where main CI started failing.
• Fix Cursor hook stdin decoding for CJK content #1637: Fix Cursor hook stdin decoding for CJK content — merged after main was already red.
• Fix Copilot VS Code selected model extraction #1638: Fix Copilot VS Code selected model extraction — merged after main was already red; also introduced test-only SQLite opens that failed the connection policy check once builds were fixed.

Related PRs checked

• feat(agent-v1): add pre_bash_call/post_bash_call types #1514: feat(agent-v1): add pre_bash_call/post_bash_call types — historical Agent v1 shell-event context.
• fix(copilot): VS Code Copilot always reports the account default model instead of the user-selected mode #1560: fix(copilot): VS Code Copilot always reports the account default model instead of the user-selected mode — related public-contributor Copilot context, not the compile break.
• feat: HTTP notes backend migration #1604: feat: HTTP notes backend migration — by Siddhant-K-code; had a “Tests started to fail after rebase” comment, but appears separate from this Agent v1 compile failure.
• fix: parse Droid settings.json as JSONC during hook install #1515 / Fix compact stash recovery edge cases #1721 / feat: add CPU/memory resource telemetry to daemon heartbeats #1724: Windows test break was called out as caused by fix: parse Droid settings.json as JSONC during hook install #1515; the latest Windows failure on this PR was SQLite error: no such table: metrics, fixed here by making metrics fallback DBs schema-initialized.

Changes

• Populate PreBashCall.command and PostBashCall.command from Agent v1 shell hook payloads.
• Keep the existing metadata value for compatibility.
• Add parser assertions so the typed command field cannot be omitted again.
• Use the SQLite memory-limited helper in Copilot OTEL/model-extraction test DB setup.
• Initialize the metrics fallback database schema so Windows test daemons cannot poison the metrics singleton with an empty SQLite DB after fallback initialization.
• Skip the PR smoke benchmark with a GitHub warning when the current branch builds but the main comparison baseline does not build.

Validation

• task test TEST_FILTER=test_fallback_database_initializes_schema
• task test TEST_FILTER=test_reset_hard_with_c_flag
• task test TEST_FILTER=sqlite_connections_go_through_memory_limited_helpers
• task test TEST_FILTER=copilot_otel
• task test TEST_FILTER=agent_v1
• task lint
• task build
• task fmt
• python3 -m py_compile scripts/benchmarks/git/benchmark_nasty_modes_vs_main.py
• Benchmark skip-path sanity run against current broken origin/main

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.