opencode-learn

Pi learn system for OpenCode — teaching that locks in, not just delivers.

Port of the pi learn harness to OpenCode. Original system by Amos Blomqvist from How I Use AI to Learn Things (thumbnail: pi by Mario Zechner / earendil-works/pi).

Quiz TUI — multi-select Which of these are rare earth elements? with ☐ 1.Cerium → ☑, I don't know — genuine gap, Note (optional) Tab-to-edit, 0 selected → ↳ Submit (Space toggle, ↓ to Submit → Enter). Durable pendingDir survives kill.

Why

Two brains can hold the same facts — one as disconnected lone facts, the other as a dependency graph where every fact is derivable from a few core truths. Teaching here builds that graph: nodes (unconditional truths) → edges (motivated discovery, 3Blue1Brown-style how could I have discovered this?).

• Probe → Plan → Teach — every session locates your edge (graded quiz probes), scopes goal (native question), plans DAG, then teaches node-by-node with quiz-check.
• Verifies, not hallucinates — researcher subagent (task subagent_type=researcher) before any shaky claim.
• Logs to Obsidian — md_log mirrors YOU/OPENCODE/Quiz/Question as > [!quote]/> [!abstract]/> [!question|success|failure] callouts, LaTeX $…$ rendered in Obsidian. viz/ PNGs embed as ![[viz-*.png|500]].

Keywords: opencode opencode-plugin learn teach quiz md-log obsidian 3blue1brown socratic

What's inside

| Pi original amosblomqvist/learn | Opencode port personal/.opencode | Notes | |---|---|---| | skills/teach/SKILL.md | skills/teach/SKILL.md | quiz + native question (custom ask removed to cut hallucination, question is single source), task researcher | | skills/visualize/SKILL.md | skills/visualize/SKILL.md | task subagent_type=mermaid-maker/svg-maker + ![[filename|500]] | | extensions/quiz | plugins/learn.ts:quiz + quiz_batch | Graded single|multi + I don't know + note, shuffled, correctAnswer by value, TUI durable pendingDir opencode-visual-tools | | extensions/ask-user-question | removed — use native question | One less tool → less overhead/hallucination | | extensions/md-log | plugins/learn.ts:md_log/md_unlog opencode.json:md_log | > [!quote] YOU / > [!abstract] OPENCODE + backfill client.session.messages + chat.message/experimental.text.complete/tool.execute | | extensions/visual-tools mermaid/svg | plugins/learn.ts:write_mermaid/edit/render + write_svg/edit/render | STAGING_ROOT=tmp/opencode-visual-tools, Chrome+mmdc / rsvg-convert→magick, viz/ publish | | agents/researcher mermaid-maker svg-maker | agents/{researcher,mermaid-maker,svg-maker}.md | researcher safe_bash → bash:ask in opencode; makers get write_*/render_* at primary (* :allow) per user choice |

Plus plugins/learn-tui.tsx — modal TUI QuizDialog/QuizBatchDialog (4-state hit/miss/false-alarm/correct-rejection success/error/warning solid bg inverted, 0a/15 lighter, ○/✓/✗), durable pendingDir .opencode/learn-pending watch+poll 700ms + hb 2s .tui-alive.

Honour & reference

• Pi — Mario Zechner — pi coding agent / earendil-works/pi — the runtime this port targets away from.
• Learn system — Amos Blomqvist — amosblomqvist/learn — the teaching philosophy (unconditional truths, how could I have discovered this?) and the quiz/ask/md-log/visual-tools harness. This port keeps the probe→plan→teach DAG and researcher verification verbatim.
• Video — How I Use AI to Learn Things (assets/thumbnail.png in original) — watch first.

This is a personal port, shared as-is. Original learn is for one learner (Amos) — edit skills/teach/SKILL.md to fit you.

Install

Plugin — simple (no installer)

Add to both configs (opencode needs server + TUI):

~/.config/opencode/opencode.jsonc — server (learn + visual):

{ "plugin": ["@bojackduy/opencode-learn"] }

~/.config/opencode/tui.json — TUI (learn-tui):

{ "plugin": ["@bojackduy/opencode-learn"] }

Use the bare package name in both files, not @bojackduy/opencode-learn/tui//server — opencode already picks the right export (./tui vs ./server) based on which host loads it. A scoped-package spec with a /tui or /server suffix is silently broken: npm-package-arg parses the second slash as a local directory reference instead of a package+subpath reference, so the plugin never resolves or activates, with no error logged anywhere.

Restart OpenCode. Verify /md_log, quiz, write_mermaid appear in tool list.

Local checkout:

// opencode.jsonc
{ "plugin": ["./path/to/personal/.opencode"] }
// tui.json
{ "plugin": ["./path/to/personal/.opencode/plugins/learn-tui.tsx"] }

Installer — skills + agents + plugins in one command

npx -y @bojackduy/opencode-learn@latest

Installs:

• plugin → opencode.jsonc/tui.json (package spec @bojackduy/opencode-learn@<version>)
• agents/ → ~/.config/opencode/agents/{researcher,mermaid-maker,svg-maker}.md
• skills/ → ~/.config/opencode/skills/{teach,visualize}/SKILL.md + marker-pdf-parser/notebooklm-lecture-notes
• commands/ → ~/.config/opencode/commands/{md_log,md_unlog}.md (if present)

Re-run to update. Then restart OpenCode.

Uninstall:

npx -y @bojackduy/opencode-learn@latest --uninstall

Or global:

npm i -g @bojackduy/opencode-learn@latest
opencode-learn            # install
opencode-learn --uninstall # remove

Manual

Copy plugins/, agents/, skills/teach, skills/visualize into ~/.config/opencode/ and add plugin entries above.

Usage

Mirror to Obsidian

/md_log /path/to/note.md   # file must exist — backfills history, mirrors YOU/OPENCODE/Quiz as callouts, LaTeX $…$
/md_unlog

View in Obsidian — viz/ PNGs embed as ![[viz-*.png|500]], mermaid $$…$$.

Quiz — probe & check

quiz(question="What is 2+2?", options=[{label:"3"}, {label:"4"}], correctAnswer="4", explanation="…")
quiz_batch(quizzes=[{question:"…", options:[…], correctAnswer:["Red","Blue"], explanation:"…", multiSelect:true}])

Single → TUI popup QuizDialog (single/multi + I don't know + note). Batch → deck 1/3→3/3 QuizBatchDialog (same 4-state hit/miss/false-alarm/correct-rejection solid bg inverted). Both durable pendingDir .opencode/learn-pending — kill opencode mid-popup → re-show on restart.

Quiz dialogs are asynchronous. The agent must call quiz or quiz_batch alone, end that assistant turn once the dialog is displayed, and continue only after the TUI injects the learner's answer.

Open forks: native question (single/multi Other).

Visual — one correct picture

task(subagent_type="mermaid-maker", prompt="graph TD: packet → ordering → reliable stream, 3 nodes only")
# → returns filename: viz-packet-*.png → embed ![[viz-*.png|500]] (md_log mirrors)

Publishing (template from @bojackduy/opencode-loopd)

This harness follows opencode-loopd's npm + installer template (~/Code/opencode-loopd):

• package.json — name:@bojackduy/opencode-learn version bin:opencode-learn=scripts/install-node.mjs files:[plugins,agents,skills,commands,dist] publishConfig access public peerDependencies @opencode-ai/plugin
• scripts/install-node.mjs — idempotent opencode.jsonc/tui.json plugin array rewrite (keeps comments, formatPluginArray), OPENCODE_CONFIG_DIR override, commands/*.md + skills/* copy, --uninstall purge, ensureDependency @opencode-ai/plugin.
• tsconfig.json + bun build → dist/ for tui (solid) if needed, prepack: bun run typecheck && bun test && bun run build.

Release:

bun run typecheck && bun test && bun run build
npm version patch -m "chore: release %s" && git push && git push --tags
# GH Action .github/workflows/npm-publish.yml on v*.*.* → npm publish --access public + gh release

See ~/Code/opencode-loopd/{package.json,scripts/install-node.mjs,.github/workflows/npm-publish.yml,README.md#Install} as template.

Requirements

• OpenCode >=1.18 + bun >=1.1
• ask-user-question not needed — native question used
• Visual: Chrome (/Applications/Google Chrome.app/...) + @mermaid-js/mermaid-cli (bundled) or rsvg-convert/magick for SVG. viz/ inside vault.

License

AGPL-3.0-or-later — see LICENSE. Original learn is personal, shared as-is by Amos. Original pi remains MIT.

Credits

• Mario Zechner — pi agent, TUI, extension API
• Amos Blomqvist — learn teaching system, quiz/md-log/visual-tools, video + repo
• Port to OpenCode — personal/.opencode learn+learn-tui (visual STAGING_ROOT opencode-visual-tools, OPENCODE block, question dedup, quiz_batch beautiful Quiz i/N)