fix(codex): use a stable prompt_cache_key instead of a per-request uuid4 to enable Codex prompt-cache reuse

[sumleo]

What

In the Codex provider, prompt_cache_key is set to a fresh uuid.uuid4() on every request:

# hindsight-api-slim/hindsight_api/engine/providers/codex_llm.py
"prompt_cache_key": str(uuid.uuid4()),   # in both call() and call_with_tools()

prompt_cache_key is OpenAI/Codex's explicit hint for routing a request to a cached-prefix backend, so it needs to stay constant across calls that share the same instructions + tools[] prefix. A new random value per request means that prefix is never matched, giving a ~100% prompt-cache miss on the hottest LLM path (recall/reflect/retain all funnel through these two methods) with no functional upside.

Fix

Derive the key once per provider instance from (account_id, model) so it is stable across requests but still distinct per account/model, and use it at both call sites. Tiny, self-contained change; the now-unused uuid import is dropped.

@functools.cached_property
def _prompt_cache_key(self) -> str:
    seed = f"{self.account_id}:{self.model}"
    return hashlib.sha256(seed.encode("utf-8")).hexdigest()[:32]

If you'd prefer the key to be scoped more tightly (e.g. per bank/session/reflect-mission) rather than per provider instance, happy to adjust — I kept it to identifiers already in scope to stay minimal.

How this was found

Spotted via static analysis of prompt-cache anti-patterns (a tool I'm experimenting with, CacheLint) and then confirmed by hand on main at 9dafadc (lines ~391 and ~714). I have not been able to run the full integration suite against a live Codex account, so a maintainer sanity-check on the cache-key semantics would be appreciated.

[sumleo]

Hi @nicoloboschi — gentle follow-up on this one (open since mid-June). It swaps the per-request uuid4 prompt_cache_key for a stable sha256(account:model) key so Codex prompt-cache reuse actually kicks in. CI is green and it's a small, self-contained change. Would you have a moment to take a look, or let me know if you'd like anything adjusted? Happy to rebase. Thanks for your time and for the project!