chore(deps): bump posthoganalytics to 7.38.3

[fivestarspicy]

Problem

• Replay Vision reports inflated input cost in AI observability on cache-heavy scanner generations.
• Its scanners use Gemini explicit context caching, where the cached-content count and the prompt count come from separate measurements and can disagree by a few percent.
• Ingestion currently infers Gemini's cache accounting model from those counts, and that shape makes cache reads look like a separate token pool rather than a subset of input.
• The full input then bills at the prompt rate and the cache bills again at the cache rate. Gemini prices cache reads at a tenth of the prompt rate, so a nearly fully cached prompt bills several times over.

Changes

• Bumps posthoganalytics from 7.30.1 to 7.38.3.
• 7.38.3 ships fix(gemini): declare cache reporting as inclusive on generations posthog-python#860, which makes the Gemini wrapper declare $ai_cache_reporting_exclusive on generations that report cache reads, so ingestion stops inferring it.
• No application code changes. Replay Vision already calls the wrapped client, so it picks this up on upgrade.

The lock update touches posthoganalytics only, with no transitive dependency changes.

Note

This spans 24 releases, so it is wider than Replay Vision. I read the changelog across the range: no breaking changes, but several behavior changes land with it. The two worth a reviewer's attention are below.

get_feature_flag_payload() now JSON-decodes payloads for locally-evaluated flags, so its return type no longer depends on where the flag resolved. Every non-test call site in this repo already handles both shapes, either through an isinstance(..., str) check before json.loads or by accepting dict | None, so this removes a latent inconsistency rather than creating one. Also in the range: alias() and group_identify() now drop empty or missing identities with a warning instead of enqueuing unusable events, flush() no longer waits out flush_interval before delivering a partial batch, and $exception_list is emitted root-cause-last.

How did you test this code?

• Ran uv lock, then uv sync. Resolution is clean and the lock diff is a single version change.
• Verified against the installed 7.38.3 that the Gemini converter sets the flag and that posthoganalytics.ai.utils maps it onto $ai_cache_reporting_exclusive.
• Imported posthoganalytics.ai.gemini.genai and confirmed the AsyncClient that Replay Vision uses resolves.
• Reviewed the changelog for all 24 releases in the range and checked this repo's call sites for the payload, identity, and flush changes named above.
• Not run: the backend test suite, which needs a database this sandbox does not have. CI covers it. I also did not exercise a live Gemini scanner run, so the end-to-end effect on reported cost is unverified here.

Automatic notifications

• Publish to changelog?

Docs update

None.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

I (the PostHog Slack app, running Claude) traced the underlying cost issue from a Slack thread, wrote the SDK fix in PostHog/posthog-python#860, and a PostHog engineer asked for this bump once it published. I did not set an assignee because I could not verify their GitHub handle in session.

An earlier attempt fixed this in ingestion instead, by adding a tolerance to the inference. That was closed in favour of declaring the accounting model at the source, which avoids tuning a threshold constant against observed token shapes.

Worth flagging for whoever reviews the blast radius: teams on older SDK versions are still subject to the inference, and it does not cover OTel or manual capture at all. This bump fixes this repo, not the general case.

Public artifact: no customer or session material reached this PR.

---

Created with PostHog from a Slack thread

[trunk-io]

😎 This pull request was merged.

[trunk-io]

   

View Full Report ↗︎ ⋅ Docs

[github-actions]

🤖 CI report

✅ Playwright — all passed

All tests passed.

View test results →

❌ Hobby preview — Health check did not pass within 35 minutes

Failing fast because: Health check did not pass within 35 minutes

---

_{Run 31206202443 | Consecutive failures: 2}

[deployment-status-posthog]

Deploy status

Environment	Status	Deployed At	Workflow
dev	✅ Deployed	2026-08-10 14:25 UTC	Run
prod-us	✅ Deployed	2026-08-10 15:04 UTC	Run
prod-eu	✅ Deployed	2026-08-10 15:09 UTC	Run