test: establish Phase 0 context migration baselines

[JasonW404]

Summary

• freeze legacy Jinja2 and managed ContextManager prompt/message snapshots for zh/en, empty/full sources, tools, builtin tools, skills, memory, KB, managed/external agents, and special characters
• capture speed/full identity, same-tenant multi-user, cross-tenant, database filter, and ContextManager cache isolation baselines
• freeze current context budget defaults and compression trigger behavior
• production behavior is unchanged

Behavior and risk

This PR intentionally records current behavior, including known gaps: conversation and KB lookup lack an explicit tenant parameter, and ContextManager caching is keyed only by conversation ID. These assertions are migration guards for later phases, not endorsements of those behaviors.

Risk is limited to test maintenance. Rollback is removal of the four added baseline files.

Local UT

• backend/.venv/bin/pytest -q test/backend/context_baseline test/sdk/core/agents/test_agent_context/unit/test_phase0_budget_baseline.py — 22 passed
• standard runner on the three added test files — 22 passed, 100%
• test/backend/utils/test_context_utils.py test/backend/utils/test_context_component_types.py — 179 passed
• test/sdk/core/agents/test_agent_context — 243 passed
• relevant agent creation, conversation, and run manager suites — 150 passed
• Ruff and git diff --check passed

The repository-wide runner was also attempted with its default 20 file workers. The added Phase 0 files passed, while unrelated backend app files hit the runner's 600-second per-file timeout under resource contention; no Phase 0 assertion failed.

Functional verification

SDK / real model / Langfuse

Model: qwen3.6-plus through the configured DashScope-compatible endpoint.

• speed: output SPEED_OK; Langfuse trace 012f22b87747bb0bf91e638f2d3f1881
• full: final context included system + memory user + KB user messages; output FULL_OK; Langfuse trace 238080b9272f63f17c55e45ef582ecb4

Traces were inspected for model input/output, final message ordering, usage, latency, and errors.

Backend API / logs

• speed /api/agent/run: SSE completed with final_answer=API_SPEED_OK
• full /api/agent/run with signed local JWT: identity resolved to user_id/tenant_id; SSE completed with final_answer=API_FULL_OK
• logs confirmed conversation creation, identity/tenant resolution, memory lookup, context component assembly, ContextManager creation, real model input, token accounting, and cleanup

No frontend production code is changed, so browser validation is not applicable.

Follow-up

Phase 1 will use these guards to tighten the backend-to-SDK authorized snapshot boundary and fix tenant/user/cache isolation without silently changing prompt content or ordering.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!