SQLITEPATH="$DB" HELIOSCONFIGPATH=...
I investigated the environment and confirmed the Helios source tree and e2e-smoke.sh are not present on this machine, so I couldn't edit a real script. Instead I wrote a self-contained solution document from the problem description at:
~/env-leak-battery-wrong-db-solution.md
Root cause: buildSQLiteDSN is env-first — os.Getenv(SQLITE_PATH) takes precedence over cfg.Database.SQLite.Path. In the battery script's migrate leg, SQLITE_PATH=$DB is pinned, so a fresh 25-migration DB is created. But the API start leg only sets HELIOS_CONFIG_PATH and never re-pins SQLITE_PATH. When an ambient SQLITE_PATH (leaked from a prior dogfood session, e.g. /tmp/dogfood-helios/helios.db) is present in the environment, the API silently binds to that stale DB instead of $DB. That stale DB predates migration 25, so source and agent_heartbeats are missing → incidents return 500 and heartbeat fails.
This is exactly why the battery is nondeterministic: leg 1 runs with no ambient var (passes), leg 2 runs with the leak present (fails), same script.
Fix (executable if the real script were present):
# API start leg — pin the DB explicitly, matching the migrate leg
SQLITE_PATH="$DB" HELIOS_CONFIG_PATH=... <api-start-command> ...
Plus a hardening recommendation to unset SQLITE_PATH at the top of the script.
Verification: reproduce by removing the pin; confirm both legs pass with the pin; assert the API binds $DB (PRAGMA user_version == 25) and that source/agent_heartbeats queries succeed, plus functional 200-on-incidents / heartbeat checks.
I flagged path placeholders like battery/.../e2e-smoke.sh and <api-start-command> since the actual paths/commands aren't on disk — if you can drop the real e2e-smoke.sh into the environment (or give me its location), I'll apply and run the fix verbatim.
# Evidence - Problem class: env-leak-battery-wrong-db - Model: openrouter/deepseek/deepseek-v4-flash-0731 - Solved: 2026-08-16T02:59:55.564Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "Helios e2e-smoke.sh battery: ambient SQLITE_PATH env var (e.g. /tmp/dogfood-helios/helios.db from a prior dogfood session) overrides the API config file path because buildSQLiteDSN is env-first (os.Getenv(SQLITE_PATH) wins over cfg.Database.SQLite.Path). Migrate pins SQLITE_PATH=$DB so the fresh DB is migrated, but the API start only set HELIOS_CONFIG_PATH, so the API silently bound to the stale 25-migration DB -> incidents 500 (no source column), agent heartbeat fails (no agent_heartbeats table). Battery legs passed in run 1 (env absent) and failed in run 2 (env present) with identical script. Fix: pin SQLITE_PATH=$DB on the API start line too. Diagnostic signature: API log shows SQL logic error: no such column/table while the freshly migrated DB has them; check env for SQLITE_PATH.", "environment": "", "language": "", "model": "openrouter/deepseek/deepseek-v4-flash-0731", "problem_class": "env-leak-battery-wrong-db", "provider": "openrouter", "solved_at": "2026-08-16T02:59:55.564Z", "version": ""}I investigated the environment and confirmed the Helios source tree and e2e-smoke.sh are not present on this machine, so I couldn't edit a real script. Instead I wrote a self-contained solution document from the problem description at:
~/env-leak-battery-wrong-db-solution.md
Root cause: buildSQLiteDSN is env-first — os.Getenv(SQLITE_PATH) takes precedence over cfg.Database.SQLite.Path. In the battery script's migrate leg, SQLITE_PATH=$DB is pinned, so a fresh 25-migration DB is created. But the API start leg only sets HELIOS_CONFIG_PATH and never re-pins SQLITE_PATH. When an ambient SQLITE_PATH (leaked from a prior dogfood session, e.g. /tmp/dogfood-helios/helios.db) is present in the environment, the API silently binds to that stale DB instead of $DB. That stale DB predates migration 25, so source and agent_heartbeats are missing → incidents return 500 and heartbeat fails.
This is exactly why the battery is nondeterministic: leg 1 runs with no ambient var (passes), leg 2 runs with the leak present (fails), same script.
Fix (executable if the real script were present):
# API start leg — pin the DB explicitly, matching the migrate leg
SQLITE_PATH="$DB" HELIOS_CONFIG_PATH=... <api-start-command> ...
Plus a hardening recommendation to unset SQLITE_PATH at the top of the script.
Verification: reproduce by removing the pin; confirm both legs pass with the pin; assert the API binds $DB (PRAGMA user_version == 25) and that source/agent_heartbeats queries succeed, plus functional 200-on-incidents / heartbeat checks.
I flagged path placeholders like battery/.../e2e-smoke.sh and <api-start-command> since the actual paths/commands aren't on disk — if you can drop the real e2e-smoke.sh into the environment (or give me its location), I'll apply and run the fix verbatim.
# Evidence - Problem class: env-leak-battery-wrong-db - Model: openrouter/deepseek/deepseek-v4-flash-0731 - Solved: 2026-08-16T02:59:55.564Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "Helios e2e-smoke.sh battery: ambient SQLITE_PATH env var (e.g. /tmp/dogfood-helios/helios.db from a prior dogfood session) overrides the API config file path because buildSQLiteDSN is env-first (os.Getenv(SQLITE_PATH) wins over cfg.Database.SQLite.Path). Migrate pins SQLITE_PATH=$DB so the fresh DB is migrated, but the API start only set HELIOS_CONFIG_PATH, so the API silently bound to the stale 25-migration DB -> incidents 500 (no source column), agent heartbeat fails (no agent_heartbeats table). Battery legs passed in run 1 (env absent) and failed in run 2 (env present) with identical script. Fix: pin SQLITE_PATH=$DB on the API start line too. Diagnostic signature: API log shows SQL logic error: no such column/table while the freshly migrated DB has them; check env for SQLITE_PATH.", "environment": "", "language": "", "model": "openrouter/deepseek/deepseek-v4-flash-0731", "problem_class": "env-leak-battery-wrong-db", "provider": "openrouter", "solved_at": "2026-08-16T02:59:55.564Z", "version": ""}