sruja / rules

---
description: Wire Phase 3/4 context tools (drift injector, session prune, memory search) when using Sruja MCP
alwaysApply: true
---

# Sruja context host (Phase 3 + 4)

When the **sruja** MCP server is enabled (see `.cursor/mcp.json`), follow this host contract. Sruja suggests; you apply.

## Tool Profiles

Sruja MCP server supports tool profiles to control which tools are available:
- `minimal` (~10-12 tools): Core ladder + focus briefing + essential utilities
- `coding` (~15-18 tools, default): Minimal + hybrid query + critique + context pruning
- `arch`: Coding + read-only authoring helpers
- `full`: All tools (backward compatible)

The profile is controlled via:
1. Environment variable: `SRUJA_MCP_TOOL_PROFILE`
2. MCP initializationOptions (set in `.cursor/mcp.json`)
3. Defaults to `coding` profile

## Session start

1. After MCP connects, if you receive **`notifications/drift_state`** (or `SRUJA_MCP_WATCH_DRIFT=1` is set), treat `params` as authoritative session context:
   - `schema_version` must be `drift_state/v1`
   - Surface `truth_status`, `health_score`, and top `violations` before large edits
2. If no notification arrived, call **`sruja_get_drift_state`** once at the start of architecture-bounded work.
3. If `aidlc-docs/aidlc-state.md` or a workflow `inception/aidlc-docs/aidlc-state.md` exists, run **`sruja workflow status -r .`** (with `--id` when multiple workflows) to surface phase gates and `aidlc` stage before large SDLC edits.

## Before precedent / “why did we…” questions

1. Call **`sruja_search_memory`** with a short query (e.g. component name, decision topic).
2. Prefer hits with `trust: reviewed_truth` (decision records) over `hypothesis` (learnings, events).
3. Never treat memory hits as permission to edit `repo.sruja` without the normal proposal/review flow.

## Long sessions (context budget)

When the conversation has accumulated many architecture element references (roughly 8+ tool turns on the same task, or token pressure):

1. Call **`sruja_suggest_context_prune`** with:
   - `active_element_ids`: current focus (from `sruja focus`, `sruja_get_focus_briefing`, or scan-resolved module IDs — prefer graph IDs over bare C4 names when warnings appear)
   - `session_element_ids`: element/module IDs mentioned in the thread
2. **Omit** transcript blocks for ids in `compress_ids`; keep full detail for `keep_ids`.
3. After compressing, avoid re-expanding dropped ids unless the user asks.

## Ladder (unchanged)

`sruja_list_architecture_index` → `sruja_get_topology` → `sruja_get_elements` → `sruja_get_task_context` (`cache_friendly: true` when caching).

## End of session

Before marking a task complete:

1. Run **`sruja verify-task --profile <coding|bugfix|review|arch> -r .`** (or call `sruja_verify_task` via MCP).
2. If verify failed:
   - Fix the issue
   - Re-run verify
   - Record a correction learning: `sruja agent record -r . -c "<context>" -H "<what was tried>" -o failed -g "<guardrail>"`
3. If verify passed:
   - Optionally record an affirmation learning

Prompt template: MCP `prompts/get` → `sruja_mcp_guide`.