Libr Tool E2E Harness Loop

Cursor-owned loop for live builtin MCP tools against a running LibrAgent:

E2E instruct → transcript extract → postmortem → improvement tasks → (patch) → re-run

Optimize tool contracts (schema, handler, hints, bridge/backend), not model choice.

When to use

• Verify a tool surface after a merge/rebuild (e.g. browser userChrome / sidecar)
• Turn a failed/partial E2E into a ranked fix backlog
• Drive short improvement cycles with measurable re-E2E

Dependencies

LibrAgent must be up (GET /api/health). Prefer executionMode: yolo for unattended E2E.

Hard rules

• Self-contained request text only — child session does not see Cursor chat.
• Always pass absolute workspacePath when the E2E needs the repo.
• Extract GET …/messages?limit=500 before terminate+delete.
• Artifacts under .libragent/work/tool-e2e-harness/<cycle-id>/ only (unless user asks otherwise).
• Separate fact (tool result text) / interpretation / hypothesis.
• One primary causal patch per cycle when implementing; ask before large product edits.
• Do not run pnpm refactor:validate unless the user explicitly asks.
• Stop looping when acceptance criteria pass, max cycles hit, or user aborts.

Cycle

Default max cycles: 3 (override if user says). Freeze a cycle-id like 20261003-browser-userchrome-1.

0. Freeze the harness contract

Write .libragent/work/tool-e2e-harness/<cycle-id>/contract.md:

• App version / branch / “fresh rebuild?” note
• Target tools + backends (e.g. browser__createSession with browser=userChrome|sidecar)
• Preconditions (extension Connected, bridge :3847, network)
• Pass/fail matrix + contamination rules (no silent backend swap)
• Out of scope

1. Instruct E2E via libr-delegate

Follow libr-delegate. Brief template: references/e2e-brief.md.

Requirements for every brief:

• Explicit tool names and required args (browser=userChrome etc.)
• Harmless fixtures only unless user authorizes otherwise
• Mandatory final markdown: Verdict / Runtime / Tool matrix / Failures / Contamination
• “No codebase archaeology” unless the cycle is analysis-only

Create session → poll to Idle/Error/Paused → on Paused resume once.

2. Extract transcript

PORT=$(cat ~/.libragent/http_port 2>/dev/null || echo 3030)
BASE="http://127.0.0.1:${PORT}"
curl -sS "$BASE/api/sessions/$SID/messages?limit=500" \
  > ".libragent/work/tool-e2e-harness/<cycle-id>/messages.json"
# save final assistant markdown as report-raw.md
curl -sS -X POST "$BASE/api/sessions/$SID/terminate"
curl -sS -X DELETE "$BASE/api/sessions/$SID"

3. Postmortem

Write postmortem.md using references/postmortem.md.

Rebuild intent → tool call → observation → next tool from messages (not memory).

Classify each failure:

Map owning layer (narrowest): schema → description/hints → MCP handler → service/bridge → extension/SW → docs/settings.

4. Derive improvement tasks

Write tasks.md using references/improvement-backlog.md.

Rank: blocker > major > nit. Prefer contract_lie / contamination / handler_bug over prompt-only tips.

Each task needs: evidence pointer, owning paths, proposed minimal change, re-E2E acceptance check, confidence.

5. Act (only if user approved patching this cycle)

1. Smallest fix for the top task (or the user-named task).
2. Focused tests (pnpm rust:test --test … / targeted lint).
3. Note reload requirements (e.g. Chrome extension Reload).
4. Bump cycle-id, return to step 0/1 with the same matrix + regression checks.

If user asked for planning only: stop after tasks.md and summarize.

6. Compare cycles

When a prior postmortem.md exists, write delta.md:

• matrix cells that flipped PASS↔FAIL
• new failures introduced
• tasks closed / still open

Outputs (per cycle)

.libragent/work/tool-e2e-harness/<cycle-id>/
  contract.md
  messages.json
  report-raw.md          # final assistant deliverable
  postmortem.md
  tasks.md
  delta.md               # from cycle 2+

Cursor summary to user

Keep short:

1. Verdict + contamination
2. Top 1–3 tasks (severity + path)
3. Whether another E2E cycle is warranted and what must be reloaded/rebuilt