feat(product): complete durable Braid workflows and terminal UX

[drewstone]

Product

Braid is one Pi-based terminal interface and JSONL CLI over agent-runtime.

An AgentProfile selects the model, runner, instructions, tools, skills, and permissions.

The runtime routes that profile through CLI Bridge, Tangle inference, or Tangle Sandbox.

Braid owns the durable conversation, branches, approvals, activity, graph, trace analysis, and terminal presentation.

What changed

• complete retained admission, exact reconnect, replay, cancellation, interaction, analysis, graph, and terminal workflows
• adopt current published Agent SDK, Runtime, Eval, CLI Bridge, Tangle provider, and Pi TUI packages
• keep one Runtime-owned execution path and remove local package overrides
• present transcript-first terminal states without invented cost, latency, machine, or placement values
• reject npm version collisions unless the published tarball is byte-identical
• update the retained Sandbox documentation from obsolete failure history to current production proof

Production proof

The default retained Tangle cohort passed 3 of 3 proofs at two-way concurrency on 2026-08-15.

Route: Braid 4562961 → agent-runtime → Tangle Sandbox → OpenCode → tangle-router/glm-5.2.

• 3/3 original Braid processes ended with SIGKILL
• 3/3 fresh processes restored exactly one local run from the provider reference
• 3/3 replays had unique fresh event IDs and zero acknowledged-event overlap
• 3/3 follow-up turns used the same provider session
• 3/3 cancellations were accepted, replayed for identical input, rejected for changed input, and remotely observed as cancelled
• 3/3 exact Sandbox resources were confirmed deleted; cohort active-resource delta was zero
• proof duration: 82.255 s minimum, 89.016 s median, 96.213 s maximum
• six completed turns: 9,231 input tokens, 1,992 output tokens, $0 provider-reported cost
• three cancelled turns: token and cost data unavailable
• memory samples: 152–157 MB current, 630–633 MB peak, 2,048 MB limit
• unavailable from the provider: machine ID, IP, effective CPU, disk, storage, placement, region, and per-Sandbox cost

Repository and release proof

• format: 745/745 files
• lint: 745/745 files
• strict TypeScript: pass
• unit scope: 251/251
• protected-live self-tests: 26/26
• live CLI Bridge adversarial matrix: pass
• registry collision cases: 4/4
• live npm: 0.1.2 available; changed 0.1.1 rejected; exact 0.1.1 accepted with SHA-256 match
• merge-tree against current main: clean
• pre-commit and pre-push secret checks: pass

The local performance-only stage was not accepted because the shared host load was 133.76 on 32 CPUs.

The clean GitHub runners below own the final full-suite performance result.

Visual proof

The recording uses a packed Braid artifact and a real Product engineer profile.

It sends a coding task through Local CLI Bridge to Claude Code and Opus, then runs /ask through the trace analyst.

Static frame: profile, model, runner, connection, transcript, and trace analyst