Pine 0.13.0 · Documentation

Pine 0.6.1: five recorded workflows

Back to Pine

These five recordings run Pine 0.6.1 with real Pi 1.4.0, herdr 0.9.5, Pine's renderer and disposable project files on Linux. Only the model responses are scripted, with a 0.18-second response delay. Commands, saved agents, feedback checks, browser interactions and the generated app's local smoke checks run for real.

Waits are shortened. The terminal streams keep their recorded output in order while long waits and repeated spinner ticks move faster. The page plays at 1.1× speed. Displayed process timers still show the original run's time; they do not become a prediction of model speed. A successful scripted trial demonstrates the workflow, not general agent quality.

Each topic has its own complete recording. The terminal is captured at 96 columns by 28 rows, with native herdr pane borders enabled from startup, and fits inside the desktop frame at 15–16px. There is no camera crop or sideways pan. On a phone, the whole recording fits inline beneath its topic; the captions and this walkthrough describe what happens. Reduced-motion mode starts with a still and requires a Play action. Playback, pause and the timeline remain available.

Widgets

The project already contains a note titled Delivery promise: Retries preserve the original delivery ID. In a clear Pi conversation:

/pine open notes

Pine opens the real notes pane below the conversation. The selected note, its text and the pane's actions are all visible. Pi remains ready above it. The captured session lasts 1.436 seconds; the delivered recording lasts 1.432 seconds, and the note appears at 00:00.519.

A note card on the page appears at that same playback cue. It is a page illustration: playing the recording does not create or edit a note on the visitor's computer. Revert hides the card and remembers that choice for the browser session. Reduced motion keeps this page effect off.

Build

The project contains notes.txt: delivery retries preserve the original delivery ID, and the retry count has not yet been specified.

Typed in the parent Pi conversation:

/pine new agent Review delivery notes. --kind review --model p03/scripted --name release-reader --access read-only

The command names the review kind, local scripted model and read-only access. Confirm the access and review the saved-agent plan. The separate builder saves the executable, checks startup and runs a bounded trial. The parent reports the current step once. The result names ./agents/release-reader, the trial result, and the workspace and pane. The agent opens in its own conversation, ready for a new request.

The captured session took 11.545 seconds. The delivered recording is 6.839 seconds after shortening waits. The trial's process exit is shown with the reminder to review its answer before using the agent for decisions.

Dashboard

Typed in the parent Pi conversation:

/pine dashboard

The browser dashboard shows the project's actual widgets. The recording begins in its Notes view, with Select multiple already enabled. Choose only Delivery promise, then Send 1 selected to Pi. The one-item selection reference is added to Pi's unsent draft. The browser displays the acceptance receipt; the person still decides when to submit the message.

The receipt reads Added to Pi draft. Press Enter in Pi when ready. It is fully visible in the delivered video at 00:01.600, and the independent Pi draft contains one selection reference with only that note. The captured browser session took 2.355 seconds; the encoded video lasts 2.383 seconds after frame-rate quantization.

The recording contains actual browser frames at 920 × 538 pixels. The page shows the complete browser frame. It does not paint a terminal receipt or illustrative text over the recording.

Facts

Typed in the saved agent's own conversation:

/feedback fact: The maximum retry count is three.
/fix

Pine first saves one pending fact. The real maintenance workflow evaluates the current and proposed agent, preserves the earlier revision and applies the correction. The running agent reloads and reports 1 applied, 0 pending.

Then ask:

What is the maximum retry count?

The answer is The maximum retry count is three. The scripted endpoint returns that answer only when the saved fact is present in the model's loaded context. The question alone is insufficient.

The captured workflow took 31.257 seconds. The delivered recording is 4.070 seconds after shortening waits. Use representative tasks and your chosen model to assess correction quality beyond this controlled example.

Durable

Typed in the parent Pi conversation:

/pine durable release-reader --model openai/gpt-5-mini

Pine exports a separate app to ./durable/release-reader and explains what needs an online adapter. A local herdr pane does not become a hosted pane. The generated app includes team chat and a session inspector, while its README names the hosting and integration work that remains.

In the exported directory, using Node 24:

npm install --ignore-scripts
npm run check
node --import tsx scripts/demo-smoke.ts > smoke-result.txt
node scripts/show-smoke.mjs

The demo copies the generated smoke script and changes only its scripted model delay from 2.2 seconds to 0.18 seconds. The checks remain the same. The final shell command reads the actual smoke result and prints its summary. The local smoke run exercises real Harness turns, request signatures, acknowledgement, replies, deduplication and worker recovery against isolated loopback servers. Its model endpoint is scripted. No model provider, GitHub or Slack account is contacted by the smoke run, and this recording does not deploy an application.

The visible result is PASS: 25 checks, 7 replies. It reports the original 60ms Slack acknowledgement and 3662ms crash recovery measurements. Those are measurements of this local smoke run. The captured session took 49.051 seconds; the delivered recording is 4.722 seconds after shortening waits. The final caption reminds you to choose hosting separately.

Setup and data

Follow the agent setup guide for the pinned public package or an exact local workspace copy. Success requires the loaded package version, a passing doctor result and visibly rendered notes and tasks inside herdr. All records in the recordings are disposable examples. No model credential or private project is recorded. Credits and provenance describe the player, source recordings and poster captures.