DottaandClaude Opus 5
e72925a848
Adds a chat surface to the Phase 7 browser explorer where the board sends prompts to a scenario and sees exactly how the mock Paperclip control plane was used on every turn — the point of the phase is that the activity stays visible rather than hidden behind the conversation (PAP-16916). Runtime owns the turn model, so the UI derives nothing: - `Phase7ChatSession` keeps one long-lived mock core across turns, so the state a turn changes is still there for the next one. Turn 0 is the control-plane seed (checkout, wake routing, exposure) that exists before anyone can type. - Timeline entries gain a `turn`; the artifact gains `turns[]` with per-turn exposure, authorization records, state diff, reconciliation events, counts, and a turn-scoped verdict. A new `user_message` kind joins the five existing ones. - `phase7TurnParity` judges a turn on what that turn did. A session-level expectation about work nobody has asked for yet would read as a failure of turn 1 rather than as work still to come. - Recording machinery moves to `scenario-execution.ts`, shared with the 7F single-shot runner so a scenario reads the same either way. Two scenarios carry authored multi-turn scripts: `ap-mcp-gate-01` (allowed turn, then a denied one) and `ix-checkbox-01` (open an interaction, then the board answers it — a control-plane-owned action with no agent tool, which is what makes the control plane schedule the continuation wake). Every other corpus scenario falls back to its 7F plan as one scripted turn. The surface implements the approved 7I interaction map: turn strips bridging each turn to its evidence, the six-section activity stream, the mobile Chat/Activity segments with a deny count chip, reset and scenario-switch confirmation, and deterministic deep links including a staged streaming hold. Deviations taken during implementation are recorded in §11a of the map rather than left for a reviewer to find. No ACPX implementation, no real control-plane integration, no credential in the browser: Codex mode stays disabled with a named reason and scripted mode drives the same in-page mock core offline. Tests: 30 new runtime/component tests and 18 browser tests; 79 phase-7 unit and 43 phase-7 browser tests pass. Ten §11 screenshots recorded by `record:phase7i`, which fails the run on any horizontal page scroll. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
5.2 KiB
Phase 7I Tutorial: Chat With the Mock Control Plane
Time to first success: about 3 minutes. One command opens a chat where you send prompts to a Phase 7 scenario and watch exactly how the mock Paperclip control plane was used on every turn. The tutorial runs from the repository root. It starts no Paperclip service, contacts no Paperclip control plane, and holds no provider credential.
Phase 7I extends the Phase 7 browser surface; it does not integrate the runner into Paperclip. Real integration is Phase 8 and requires separate approval. See the future binding boundary reference.
What you need
- Node.js 20 or newer and pnpm 9 or newer. Verified with Node 22.22.2.
- No Rust toolchain and no network access after
pnpm install.
Install the package workspace from the repository root:
pnpm install --filter @paperclipai/paperclip-runner --lockfile=false --offline --ignore-scripts --dev
1. Open the chat (about 1 minute)
pnpm --filter @paperclipai/paperclip-runner run demo:phase7
Open http://127.0.0.1:4183/phase7-explorer/#/chat/ap-mcp-gate-01.
Before you type anything, the conversation already has content: a session seed strip reading "Control plane checked out the task and opened the run — no agent tool exists for this," and a Session seed activity strip counting what it did. That is the point of turn 0. Work happened, and no agent tool was involved.
The session status chip reads Idle · seeded. The right rail — Activity on desktop, the Activity segment on a phone — shows the seed group with its exposure decision, state diff, and verdict.
2. Run a turn that succeeds
Type into the composer and press Enter:
Where does the MCP tool approval for this task stand?
The turn appends: your message as a right-aligned bubble, two semantic calls
(get_task_context, list_approvals) as tool cards with their disposition
badges, the agent's reply, and a Turn 1 activity strip:
Turn 1 activity ⚙ 2 tools ⌘ 0 control plane Δ 0 domains ✓ Pass
Click the strip. The Activity stream opens turn 1's group with the six evidence sections in order — exposed tools, calls and results, authorization, control plane, state diff, parity. Nothing changed state, and the diff says so.
3. Run a turn that is denied
Raise the approval request yourself and get the tool switched on.
request_approval is denied. The denial renders in place on the danger surface
with the failing operation, the code policy_denied, the reason
required_claim_missing, and an explicit line that no protected state was
returned. The turn strip counts it (✕ 1 denied), and on mobile the Activity
segment carries a · 1 deny chip so you can see it without opening the segment.
The session status still reads Scripted · deterministic and the session
parity still passes. A denial is a correct outcome, not a broken session — this
run simply does not hold governance:approvals:request. Open the turn's
Authorization section to see the claims considered and the missing one.
4. See control-plane-owned work with no agent tool
Open http://127.0.0.1:4183/phase7-explorer/#/chat/ix-checkbox-01?replay=fake.
replay=fake plays the recorded three-turn conversation.
Turn 2 opens a checkbox interaction. Turn 3 is the board answering it — and that answer is performed by the control plane, not by a tool:
⌘ Control plane recorded the board's answer and scheduled the continuation
wake — no agent tool exists for this.
Open turn 3's State diff section. The interactions domain shows
status: pending → answered, the unchanged domains collapse to one line, and a
Control plane badged row records wake.scheduled. The wake is scheduled by
reconciliation; the agent never asked for it.
5. Reset and replay
Reset session asks first, because it discards mutable mock state. Accept it and the transcript returns to the seeded turn 0 — no turns, no denials, a clean diff. Switching scenario from the rail mid-session asks the same question, since two scenarios never share mock state.
Replay scripted session plays the remaining recorded turns and needs no confirmation: it reproduces an identical timeline.
6. Run the tests and record the evidence
pnpm --filter @paperclipai/paperclip-runner run test:phase7
pnpm --filter @paperclipai/paperclip-runner run test:browser:phase7
pnpm --filter @paperclipai/paperclip-runner run record:phase7i
The first covers the session data contract and the chat components, the second
drives the surface in a real browser at both viewports, and the third writes the
ten acceptance screenshots to knowledge/evidence/phase-07/ — failing if any
route scrolls horizontally.
What this does not do
- No ACPX implementation and no real control-plane integration.
- No provider credential in the browser. Codex (bounded) stays disabled with a named reason unless the local provider relay is running; Scripted (deterministic) drives the same mock core offline.
- Nothing you type is persisted. A reload drops an unsent draft, which is the accepted trade against storing conversation content.
Reference: Phase 7I scenario chat.