feat(lasso): intent double-duty on the classify call

[orgersh92]

What

Adds intent double-duty to the LiteLLM Lasso custom guardrail. When an app funnels its LLM calls through LiteLLM and seeds per-turn intent headers (via the lasso-sdk GatewayIntent primitive), the plugin's existing /classify call now also carries the per-turn intent signals — so content-safety and intent ride one call, with no separate exporter or second Lasso API key in the app.

Inert when no x-lasso-trace-id is present (a plain guardrail call), so it ships dark until the app starts seeding.

Change

On each classify call the plugin reads the app-seeded headers off the proxy request and stamps the classify payload's messages with intent ids:

Header	Meaning
x-lasso-trace-id	per-turn trace id (a ULID); groups the turn's events
x-session-id	conversation id → sessionId
x-lasso-application-intent	seeds sessionInformation.applicationIntent
x-lasso-application-name	seeds sessionInformation.agenticAppName

(x-lasso-encoding: pct → percent-decode the baseline values.)

The server derives the intent signal from role × content-kind, so the plugin sends blocks, not signals — reusing the guardrail's existing tool translation (tool_use under model, tool_result under developer, input as an object). Model reasoning is harvested from reasoning_content / thinking_blocks into a reasoning block. Also now sends source={type:"litellm"} for the "Used By" attribution badge.

Ordering + idempotency (the interesting part)

The gateway re-sends the whole conversation history on every turn, so the server dedups on eventId. Ids must therefore be deterministic — derived from (traceId, eventIndex) — so a re-send at the same position collapses instead of duplicating. Reproducible events (text / tool_use / tool_result) are indexed on a stride (p × 10); COMPLETION-only reasoning (never re-sent in history) fills just above the previous reproducible event, off the stride grid, so it sorts before the answer it explains without shifting the re-sendable events. The PROMPT phase caches its reproducible-event count (keyed by litellm_call_id) so the COMPLETION phase continues after it. This mirrors the Kong ai-lasso-guardrail plugin (the reference intent implementation).

Verification

• Pure id/index logic (stride ordering, deterministic ids, cross-turn idempotency, reasoning gap-slotting, unslottable-leading-reasoning drop) verified against the identical scheme used by the Kong plugin and the server-side LiteLLM shim.
• Unit tests added in tests/guardrails_tests/test_lasso_guardrails.py (inert-without-header, placeholder-ignored, stride stamping + deterministic ULID ids, pct-decoded session information, completion index continuation + reasoning harvest). Run in CI.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: de07c1fc-ad3b-4461-86cc-3957d141552a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/intent-double-duty

---

_{Comment @coderabbitai help to get the list of available commands.}