fix: require explicit RAG scope before generators use docs

[madara88645]

Summary

• Fixes RAG: require explicit context scope before generator requests use persisted documents #1155. Generator and compile flows no longer treat leftover indexed documents as automatically “in use.” Prior-session files can stay in the library, but they are not sent into a new request until the user attaches them for this session (or uploads/ingests during this session).
• Context Manager now shows library availability separately from active scope, including document counts/source and a clear Detach context action. The misleading “Connected = docs attached” signal is replaced with backend health vs. active-scope status.
• Backend agent/skill generation now defaults enable_context_retrieval to off (same opt-in pattern as compile), so cloud-backed generation cannot silently inherit stale local index content.

Small related improvement

• Compiler, Agent Generator, and Skill Generator now share one session-scoped context state (instead of separate hook instances), so the “RAG Context Active” banner and the generate request stay in sync.

Test plan

• python -m pytest tests/test_generator_rag_opt_in_api.py tests/test_context_generation.py -q
• cd web && npm run test -- --run app/hooks/useContextManager.test.tsx app/components/__tests__/ContextManager.test.tsx app/__tests__/page-rag-context-scope.test.tsx app/agent-generator/page.test.tsx app/skills-generator/page.test.tsx
• Manual: open Compiler / Agent / Skill pages with a pre-populated local index and confirm no “active” banner until Use library for this session
• Manual: attach, generate, detach, and confirm later generate requests do not re-enable retrieval without attaching again
• Manual: upload a file in-session and confirm it auto-attaches for that session only

Made with Cursor

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
compiler	Ready	Preview	Aug 4, 2026 3:36pm

[cursor]

Stale comment

PR Risk Assessment (Automation) — Re-evaluation

Risk level: Low (unchanged)
Code review required: No
Decision: Prior approval retained — no re-approval posted

Trigger

PR synchronized after prior approval (98c1c91). New HEAD: 2b41a12.

Delta since last approval

Only test files changed in the new commit:

• tests/test_agent_generator.py — add enable_context_retrieval=False to mock expectations
• tests/test_hybrid.py — pass enable_context_retrieval=True where RAG retrieval is exercised
• tests/test_multi_agent.py — expect context=None when retrieval is off; add opt-in flag to API mocks
• tests/test_skills_generator.py — add enable_context_retrieval=False to mock expectations

No production code, infra, auth, or schema changes in the delta.

Full PR evidence (production surface)

9 non-test files modified:

• api/routes/generators.py — enable_context_retrieval defaults to false on agent/skill generator requests
• app/llm_engine/hybrid.py — gates context_strategist.process() behind opt-in flag (mirrors compile pattern)
• web/app/hooks/useContextManager.ts + ContextManager.tsx — session-scoped attach/detach; library visible but inactive until attached
• web/app/hooks/useCompiler.ts, page.tsx, agent-generator/page.tsx, skills-generator/page.tsx — wire contextAttached into requests; shared context state
• web/lib/api/types.ts — type annotation only

Why Low (risk did not increase)

• Conservative behavioral change: defaults to less context sent to LLMs, not more
• Established pattern: extends existing compile RAG opt-in design to generators
• Limited blast radius: scoped to RAG context scoping across three related UI flows; no auth, billing, infra, or deployment changes
• Strong test coverage: dedicated API tests, context-generation updates, frontend scope tests; latest commit tightens mock alignment

Reviewer assignment

Not required at Low risk. No CODEOWNERS file. 0 reviewers currently assigned.

CI

All checks passing (Smoke, Analyze, CodeQL, Snyk, GitGuardian, VS Code Extension, PR Tests).

---

Automated risk assessment — conclusions derived from diff evidence only.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Stale comment

PR Risk Assessment (Automation) — Re-evaluation

Risk level: Low (unchanged)
Code review required: No
Decision: Prior approval retained — no new approval posted (already approved on 98c1c91)

Trigger

PR synchronized (synchronize). New HEAD: 31dc335 (merge main into fix/rag-explicit-context-scope-1155).

Delta since last assessment (2b41a12)

The new commit is a merge from main only — it brings upstream CI/tests/security fixes already on main into the branch. Comparing origin/main...HEAD still shows 29 files / +614 −46, identical in scope to the prior assessment (RAG explicit scope work only). No additional PR-authored production logic in this sync.

Evidence (production surface)

Area	Change
api/routes/generators.py	enable_context_retrieval defaults to false for agent/skill generator requests
app/llm_engine/hybrid.py	Gates context_strategist.process() behind opt-in flag (mirrors compile)
web/app/hooks/useContextManager.ts + ContextManager.tsx	Session attach/detach; library visible but inactive until attached
web/app/hooks/useCompiler.ts, page.tsx, generator pages	Wire contextAttached → enable_context_retrieval / compile flag
web/lib/api/types.ts	Type annotation only

Why Low (risk did not increase)

• Conservative default: less persisted RAG content sent to LLMs unless the user opts in
• Cohesive, limited domain: RAG scope across compiler + agent/skill generators; no auth, billing, infra, or schema changes
• Substantial automated tests: test_generator_rag_opt_in_api.py, context-generation updates, frontend scope tests
• Merge sync: does not expand the merge-to-main diff beyond the above

Reviewers

Not required at Low risk. No CODEOWNERS file. 0 requested reviewers (under the 2-reviewer cap).

CI

Checks were in progress at assessment time (Smoke, CodeQL Analyze, VS Code Extension); Snyk and GitGuardian already green.

Slack

Automation has no Slack channel configured for this run (slackChannelId null); summary delivered via this PR review only.

---

Automated risk assessment — conclusions derived from diff evidence only.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

PR Risk Assessment (Automation) — Re-evaluation

Risk level: Low (unchanged)
Code review required: No
Decision: Prior approval retained — no new approval posted (already approved on 98c1c91)

Trigger

PR synchronized (synchronize). New HEAD: 5146e1e (merge main into fix/rag-explicit-context-scope-1155).

Delta since last assessment (31dc335)

New commit merges upstream main into the branch. Comparing 31dc335..5146e1e touches 9 files (+186 lines), including merged-from-main generator error-artifact handling (_raise_if_generator_error_artifact in api/routes/generators.py, web/lib/generatorErrorArtifact.ts, client asserts on agent/skill pages) plus matching tests.

origin/main...HEAD remains 29 files (+614 −46) — same PR-authored scope as prior runs (explicit RAG scope for compile + generators). The merge does not widen the merge-to-main diff.

Evidence (production surface)

Area	Change
api/routes/generators.py	enable_context_retrieval defaults to false; passes flag into hybrid generators
app/llm_engine/hybrid.py	Gates context_strategist.process() behind opt-in (mirrors existing compile path)
web/app/hooks/useContextManager.ts + ContextManager.tsx	Session attach/detach; library visible but inactive until attached
web/app/hooks/useCompiler.ts, page.tsx, generator pages	Wire contextAttached → retrieval flags on compile/generate requests
web/lib/api/types.ts	Type annotation only

Why Low (risk did not increase)

• Conservative default: less persisted RAG content sent to LLMs unless the user explicitly attaches context
• Established pattern: extends existing compile enable_context_retrieval design to agent/skill generators
• Cohesive domain: RAG scoping across related UI flows; no auth, billing, infra, deployment, or schema changes
• Merged safety fix: error-as-markdown generator responses are rejected server- and client-side (from main; reduces mistaken export of failures)
• Strong automated tests: test_generator_rag_opt_in_api.py, context-generation updates, frontend scope tests

Reviewers

Not required at Low risk. No CODEOWNERS file. 0 requested reviewers.

CI

CodeQL, Analyze, Snyk, GitGuardian, and Vercel green; Smoke and VS Code Extension still pending at assessment time.

Slack

No Slack destination is configured for this automation run; summary delivered via this PR review only.

---

Automated risk assessment — conclusions derived from diff evidence only.

  

_{Sent by Cursor Automation: Assign PR reviewers}