[cache-strategy] Fix cache miss in Smoke Gemini

[github-actions]

Cache Strategy Problem: Smoke Gemini

Severity: 🟠 High

Workflow: Smoke Gemini
Analysis Date: 2026-05-14
Issue Type: Cache miss / No persisted cache files

---

Problem Description

The last 24 hours show repeated cold starts for Smoke Gemini despite cache-memory being enabled. The workflow uses the expected run-specific key plus restore-key fallback pattern, so the key shape itself is not the problem; the actionable issue is that no useful cache state is being persisted for later runs.

Evidence

• Miss streak: 4 cache-memory runs observed today with misses
• Miss rate (14d): 100% in the available cache-strategy history
• Last cache hit: never recorded in cache-strategy history
• Log signals: 2026-05-14T17:17:19.1313321Z Cache not found for input keys: memory-none-nopolicy-smokegemini-25874070421, memory-none-nopolicy-smokegemini-

Recent Run Evidence

Run	Cache Result	Signal
§25863489227	miss	2026-05-14T13:50:57.6467062Z Cache not found for input keys: memory-none-nopolicy-smokegemini-25863489227, memory-none-nopolicy-smokegemini-
§25863982749	miss	2026-05-14T13:55:16.9656117Z Cache not found for input keys: memory-none-nopolicy-smokegemini-25863982749, memory-none-nopolicy-smokegemini-
§25864876786	miss	2026-05-14T14:20:36.1001836Z Cache not found for input keys: memory-none-nopolicy-smokegemini-25864876786, memory-none-nopolicy-smokegemini-
§25874070421	miss	2026-05-14T17:17:19.1313321Z Cache not found for input keys: memory-none-nopolicy-smokegemini-25874070421, memory-none-nopolicy-smokegemini-

Recommended Fix

1. Persist meaningful state or disable cache-memory — If Smoke Gemini does not need cross-run memory, remove cache-memory from the workflow so it stops paying cache setup/restore overhead. If it does need memory, write structured state under /tmp/gh-aw/cache-memory/<workflow>/ before the agent exits.
2. Keep the restore-key fallback — The observed key pattern includes a stable restore-key fallback, which is the valid last-write-wins design. Do not replace it with a run-specific key that lacks restore-keys.
3. Add a lightweight verification step — After the agent runs, verify that /tmp/gh-aw/cache-memory/ contains at least one useful file before uploading the cache-memory artifact.

Expected Impact

Fixing this should prevent repeated cold starts for this smoke workflow and reduce cache setup overhead and repeated work on future runs.

References: §25863982749, §25864876786, §25874070421

Generated by Daily Cache Strategy Analyzer · ◷

• expires on May 21, 2026, 7:12 PM UTC