Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Lesson 4: Grounded harness and continual learning (host-owned)

Harness vs runtime

Sruja (harness)Editor / CI (host)
Validates .srujasruja lint, sruja fmtProposes DSL edits
Compares model to codesruja drift, sruja intent checkInterprets violations
Stores learnings.sruja/agent_memory.jsonReflects on runs (optional)
Open-source LLM--enrich-cmd, config.tomlRuns Ollama, vLLM, etc.

Sruja does not ship ContinualLearningAgent, --autonomous, or automatic skill generation. Those are patterns you implement in the host, using Sruja JSON as ground truth.

Full reference: Grounded harness and continual learning.

Checklist: grounded architectural task

  1. Sync evidence — sruja sync -r .
  2. Brief the task — sruja focus --file <path> -r . or MCP task context
  3. Propose — skill/MCP synthesis or sruja author propose --enrich-cmd '…'
  4. Lint loop — sruja lint repo.sruja --format json until clean; feed errors back to the model
  5. Drift gate — sruja drift -r . -a repo.sruja
  6. Promote — sruja propose approve <id> or human merge to repo.sruja
  7. Record — sruja agent record or MCP sruja_record_learning on success/failure

Bounded agent loop

sruja agent plan -r . --goal "Fix drift on Checkout" --element-id Shop.Checkout --print
# Review plan JSON, then:
sruja agent apply -r . --plan docs/plans/<run-id>.json

Optional MaTTS parallel attempts:

sruja agent run -r . --goal "..." --file src/checkout.rs --mode apply --trajectories 3

Replay bundle after apply: .sruja/agent/runs/<run_id>/facts_bundle.json.

Local inference (Ollama example)

# .sruja/config.toml
[integrations]
default_provider = "cmd"
cmd = "ollama run llama3"
timeout_ms = 15000
max_bytes = 20000

sruja inspect onboard -r . -f markdown --enrich-cmd 'ollama run llama3'

Memory and curation

Learnings are LearningEntry records—not separate buckets for “facts” vs “preferences”:

sruja agent record -r . \
  -c "Checkout boundary" \
  -H "Split cart from payment container" \
  -o success \
  -g "Run drift after any Checkout container change" \
  --hitl-kind precedent

sruja agent curate -r .
# After reviewing curate output, merge duplicates into one entry:
sruja agent merge -r . --ids id1,id2 -c "Merged context" -H "..." -g "Combined guardrail"

MCP ladder (editors)

  1. sruja_list_architecture_index
  2. sruja_get_topology
  3. sruja_get_elements
  4. sruja_get_task_context (cache_friendly: true when caching)

Setup: mcp_setup.md.

Reflect / learn (host-owned)

After agent apply, optionally:

  1. Read facts_bundle.json and recent agent_memory.json entries.
  2. Use your editor agent or --enrich-cmd to draft narrative lessons (markdown only).
  3. Persist with sruja agent record or sruja_record_learning.
  4. Run sruja agent curate before merging duplicate learnings.

Cursor users: run the project command Sruja: Reflect on agent run (see .cursor/commands/sruja-reflect-on-run.md).

What you learned

  • Sruja constrains probabilistic output with deterministic gates.
  • Continual learning lives in artifacts and memory, not weight updates.
  • The host owns Act/Reflect; Sruja owns Record/Learn storage and enforcement.

Next: Agentic orchestration patterns and Sruja for industry pattern mapping.