bodek: a beautiful Bubble Tea TUI for the odek agent

[jkyberneees]

Summary

bodek is a beautiful Bubble Tea terminal interface for the odek agent.

It is a pure front-end: it launches (or attaches to) an odek serve instance and renders the agent's live stream over odek's WebSocket + REST protocol. All agent behaviour — tools, the danger approval engine, the Docker sandbox, skills, memory, sessions — is provided by odek itself; bodek never re-implements any of it.

┌──────────────┐   WebSocket (RFC 6455, JSON) + REST   ┌──────────────────┐
│    bodek     │ ◄───────────────────────────────────► │   odek serve      │
│ (Bubble Tea) │   tokens · tools · approvals · @refs   │  (ReAct engine)   │
└──────────────┘                                        └──────────────────┘

What it does

• Streaming chat — token-by-token answers rendered as Markdown (glamour).
• Live tool activity — every tool_call/tool_result with a per-tool glyph, spinner, argument preview, and one-line result.
• Security approvals — odek's danger prompts surface as an inline panel; answer a/d/t.
• @ file/session attachments — live autocomplete via /api/resources; references resolve server-side through odek's untrusted-content boundary.
• Session browser (^R) — resume, replay, and delete past conversations (/api/sessions).
• Model switcher (^O) — /api/models.
• Cancellation (Esc) — abort a running turn via /api/cancel.
• Sandbox aware — header shows 🛡 sandboxed / ⚠ host access; --sandbox opts into Docker isolation.
• Context-aware progress — 🧪 running tests, 📖 reading client.go, 🚀 pushing, with a live elapsed timer.
• Fluent UX — gradient wordmark/hairline, smooth braille spinner, smart autoscroll, scroll indicator.

Per-session auth tokens are persisted in ~/.bodek/sessions.json so resume/cancel/delete work across runs (mirroring the Web UI's localStorage).

Project layout

Path	Responsibility
cmd/bodek	CLI entry: flags, lifecycle, wiring
internal/server	Launch/attach to odek serve, resolve the auth token
internal/client	odek serve WebSocket + REST protocol
internal/tokens	Local persistence of per-session auth tokens
internal/tui	Bubble Tea model, update loop, panels, view

Quality

• ~99% internal-package statement coverage (client 100%, tui 99%, tokens 98%, server 95% — the remainder is unreachable OS-error handling), via unit + integration tests against an in-process odek serve stand-in.
• golangci-lint v2 clean, go vet clean, gofmt clean.
• CI (build, vet, race tests + coverage, lint) and a GoReleaser release workflow on version tags.

Verification

End-to-end validated against a locally built odek serve: token auth, WebSocket dial, prompt/event streaming, and the resource/session/model/cancel REST endpoints all work. The interactive TUI itself was not run on a TTY in CI and a full LLM turn was not exercised (no API key / network egress in the build env).

🤖 Generated with Claude Code

https://claude.ai/code/session_012rWqu7pktbd2ejfNw3a7Vf

---

Generated by Claude Code

[jkyberneees]

🔎 Verification Certificate

Applied The AI Verification Protocol v5.2.7 to this PR. This was a single-agent pass (no multi-provider B/C/D/E pipeline, no replay-sandbox/fuzz infra), so η is reported qualitatively with coverage as a proxy rather than a full mutation/oracle derivation — unverified gaps are listed explicitly per §9.

Classification: GeneratedCode (AI-authored) · Verdict: HumanReviewRecommended

Axes (✅ pass · ⚠️ note · 🔴 fail→fixed)

Axis	Status	Finding / Resolution
2.1 Semantic Correctness	✅	Logic exercised by unit + integration tests against an in-process odek serve stand-in.
2.2 Behavioral Contract	✅	Event/REST field mapping verified against cmd/odek/serve.go (the spec source), not inferred from bodek.
2.3 Security Surface	🔴→✅	Fixed: untrusted agent/tool output rendered to the terminal without stripping ANSI/control escapes.
2.4 Structural Integrity	✅	Clean client/server/tokens/tui separation; no cycles.
2.5 Behavioral Exploration	⚠️	Disconnect/approval/cancel/empty-size paths covered by tests; a live LLM turn was not exercised (no key/egress).
2.6 Dependency Integrity	✅	All deps pinned in go.mod (no */floating ranges).
2.7 Generator Provenance	✅	N/A (verifier side); commits attributed.
2.8 Adversarial Surface	🔴→✅	Fixed: terminal-injection via ESC/OSC 52 (clipboard exfil) / cursor & screen manipulation in streamed tokens, tool args/results, reasoning, notices, and resumed transcripts.
2.9 Documentation Coverage	✅	README updated with every new feature, key bindings, badges, dev/CI docs.

Remediation applied (auto-repairable, additive — §7.1)

fee8bbe — sanitize()/isControl() strip C0 control bytes + DEL (incl. ESC) while keeping newlines/tabs, applied to all untrusted display paths. Added unit + end-to-end defang tests. This is purely defensive (no behavior change to legitimate content), so it was auto-applied rather than queued for human-only review.

Also fixed: CI Lint job (golangci-lint-action v6 → v7, required for golangci-lint v2).

Quality signals (η proxy)

• Internal-package statement coverage 98.7% (1122/1137) — client 100%, tui 99%, tokens 98%, server 95% (residual is unreachable OS-error handling).
• go vet, gofmt, golangci-lint v2 all clean; race-enabled tests pass.

Unverified gaps (§9 — honest disclosure)

• No end-to-end LLM turn (no API key / network egress in the build env) — streaming render verified by construction + simulated events, not a real provider round-trip.
• No TTY run of the interactive Bubble Tea screen.
• η signals m (mutation), f (fuzz), and the independent-oracle o were not computed — out of scope for a single-agent pass.

Rationale: no residual 🔴 after remediation; verdict is HumanReviewRecommended (not AutoApprove) because this is AI-generated code verified by a single agent, and the live behavioral layer (2.5) could not be exercised here — that's the layer the protocol says AI cannot self-evaluate.

🤖 Generated with Claude Code

---

Generated by Claude Code