[Refactor] Add an ordered context transformation plan to the existing IR

[ethanjyx]

Closes #3472

Purpose

Give reset, exact deduplication, turn selection, and the existing block compressor one request-owned transformation plan. Policies receive detached views and propose edits; shared validation preserves protected messages, complete selected turns, tool-exchange membership, and retained message order before committing a step.

Extend the existing RequestIR and MessageIR with trusted provenance, eligibility, protection reasons, and stable turn/exchange identity. Capture original semantic history before RAG and Memory enrichment so later topic evidence can read an immutable snapshot. Route configured compression through the same plan and record bounded receipts without request text or arbitrary callback errors.

No public configuration or new policy is enabled. RAG still precedes Memory; configured tool-result text compression remains supported while non-text blocks and tool structure stay intact. Reset triggers, deduplication equivalence, ranking, summarization, and token budgeting remain separate work.

Owner: wg/agentic-context.

Test Plan

• Run the contextcompression and protocolcodec suites, including the race detector.
• Fuzz history-removal invariants and verify rejected edits never partially apply.
• Verify byte-compatible disabled plans and transformed round trips for OpenAI Chat, Responses, and Anthropic.
• Run request-pipeline tests with native libraries, repository lint, and the local image startup checks.

Test Result

Passed on commit 7513f43d857d692e1e5b650ab60e9b7e23e36b04:

• make agent-validate.
• SKIP_MODEL_DEPENDENT_TESTS=true make agent-ci-gate AGENT_CHANGED_FILES_PATH=/tmp/context-plan-changed-files: native Candle/ML/NLP builds, router build, baseline pre-commit, Go structural lint, structure and architecture checks, and 55 Go package test results, including the complete pkg/extproc, pkg/contextcompression, and pkg/protocolcodec suites.
• go test -race ./pkg/contextcompression ./pkg/protocolcodec from src/semantic-router.
• go test ./pkg/contextcompression -run '^$' -fuzz '^FuzzTransformationRemovalInvariants$' -fuzztime=10s -parallel=2 (61,889 executions).
• Nine Linux request-pipeline tests in the final Dockerfile's go-builder image, with native Candle/ML/NLP libraries: original history capture, history selection without compression, tool-result compression, request controls, non-text preservation, RAG opt-in, neutral-request preparation, fail-closed handling, and recovery-tool name protection.
• make vllm-sr-dev: both final router and dashboard images built successfully, including the frontend Vite build.
• make agent-serve-local ENV=cpu AGENT_STACK_NAME=context-plan AGENT_PORT_OFFSET=400, then make agent-smoke-local with the same stack and offset: router, Envoy, and dashboard healthy.
• A live request through Envoy and the router to a local mock backend selected smoke-model, preserved the system instruction and conversation history, and returned the backend response.
• The feature-only test resolver reported no additional matching commands.

The Docker storage blockers are resolved after restoring host disk headroom and restarting Docker Desktop. Image-store reads, builds, network creation, and startup all completed without I/O errors. The temporary test stack and mock backend were stopped, and the test-only builder image was removed.

Validation scope: the core gate used the repository's supported SKIP_MODEL_DEPENDENT_TESTS=true option, excluding the model-dependent memory/tools package suites. External Milvus, Qdrant, Redis, Valkey, and Llama Stack test integrations retained the core gate's default skip settings. Model/GPU-dependent and Kubernetes E2E suites were not run. No public policy is enabled, so existing E2E coverage is retained for this internal refactor.

---

Semantic Router PR Checklist

• PR title begins with exactly one bracketed category.
• The PR links an accepted issue with exactly one Workgroup owner.
• Commits are signed off.
• Test results reflect completed validation and the remaining blockers.

[netlify]

✅ Deploy Preview for vllm-semantic-router ready!

Name	Link
🔨 Latest commit	66f028f
🔍 Latest deploy log	https://app.netlify.com/projects/vllm-semantic-router/deploys/6a9f073b0b6adc0008bb4f96
😎 Deploy Preview	https://deploy-preview-3521--vllm-semantic-router.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.
🤖 Make changes	Run an agent on this branch

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[Xunzhuo]

Thanks for putting the shared seam in place. The request-owned plan validates the complete step order before mutation, gives policies detached views, preserves complete turns and tool exchanges, and routes the existing compressor through the final commit stage. I do not see a code blocker in this draft. Please rebase onto current main and mark it ready when you want the final approval pass.

[ethanjyx]

@Xunzhuo updated

[Xunzhuo]

Thanks for taking this out of draft. The rebased head is patch-equivalent to the reviewed transformation-plan implementation, the history and protected-turn invariants remain intact, and the latest exact-head checks are green. Approved.

[mergify]

Queued — the merge queue status continues in this comment ↓.

[ethanjyx]

@Mergifyio queue

[mergify]

queue

☑️ Command disallowed due to command restrictions in the Mergify configuration.

Details

• sender-permission >= write

[wilsonwu]

LGTM

[mergify]

Merge Queue Status

• ✅ Entered queue — 2026-09-07 07:49 UTC · Rule: default · triggered by @wilsonwu with the merge queue checkbox
• 🚫 Left the queue — 2026-09-07 08:11 UTC · at faca2402f33b93417305785d7d9b5ae7da945a3e

This pull request spent 22 minutes 26 seconds in the queue, with no time running CI.

Reason

The pull request can't be updated

For security reasons, Mergify can't update this pull request. Try updating locally.
GitHub response: refusing to allow a GitHub App to create or update workflow .github/workflows/ci-changes.yml without workflows permission

Hint

You should update or rebase your pull request manually. If you do, this pull request will automatically be requeued once the queue conditions match again.
If you think this was a flaky issue, you can requeue the pull request, without updating it, by posting a @mergifyio queue comment.

Tick the box to put this pull request back in the merge queue (same as @mergifyio queue).

• Requeue this pull request

[ethanjyx]

@wilsonwu @Xunzhuo can you approve again ? I rebased.