Roadmap and issues

tau's issues are Markdown files in the repository, under roadmap/, one file per issue, grouped by layer. There are no GitHub issues. A small script, roadmap/board.py, reads the header lines of every file, prints the board and validates the tracker. This page is how to read it and how to work an issue.

Layers

tau is built in layers; each becomes useful on its own before the next starts. The layer plan itself is in ARCHITECTURE.md.

Layer	Goal
L0 · Remote access	Phone → Tailscale → mosh → tmux
L1 · Agent	Strands agent, persona, tiers, sessions, TUI, hub daemon
L2 · Models	Role routing through a LiteLLM proxy with hard budgets
L3 · Access and data	Tunnel + Access in front of the web UI, Telegram by direct polling, local settings, memory and KB, encrypted R2 backups (the directory keeps its old name L3-cloud)
L4 · Interface	TauBar, the web UI, tau-voice, Claude Code integration, ambient mode
L5 · Mesh	Zenoh + MCP nodes, MQTT bridge, reflex routines
L6 · Robot	SO-101 and the workout coach
L7 · Lab	AgentCore and other experiments
L8 · Distribution	PyPI releases and the docs site
Backlog	Waiting on hardware or decisions

Each layer directory has a README.md with the goal, a "done when" sentence and a table of its issues.

The board

python3 roadmap/board.py               # everything, grouped by layer
python3 roadmap/board.py --next        # unblocked ready-for-agent issues
python3 roadmap/board.py --human       # unblocked ready-for-human issues
python3 roadmap/board.py --check       # validate the tracker (run after editing issues)
python3 roadmap/board.py --layer L3    # one layer; also `backlog`

L1-agent
  003  ready-for-human  Provision model credentials with cost guards
  004  done             Tracer bullet: `tau chat` answers in the Jarvis persona
  006  done             Tool permission tiers with approval and a loop guard
  ...

--next prints only issues whose every blocker is done, so an agent (or you) can pick the top one without reading the graph.

An issue file

roadmap/<layer>/<NNN>-<slug>.md. Ids are global, three digits, never reused; a new issue takes the highest existing id plus one and gets a row in its layer's README. The header lines are machine-read and must keep this form:

# 051 · Short imperative title

Status: needs-triage
Type: AFK
Layer: L5 · Mesh
Packages: tau-mesh
Blocked by: 032, 033

## What to build

End-to-end behavior of this vertical slice, not layer-by-layer steps.

## Acceptance criteria

- [ ] Observable, testable outcome

## Blocked by

- [032 · Title](032-slug.md)

## Comments

• Type is AFK (an agent can do it end to end), HITL (needs a human in the loop) or needs-hardware.
• Packages names the distributions or directories the work touches.
• Blocked by lists ids; the board refuses to offer an issue whose blockers are not done.
• Comments is an append-only log with dated notes.

Statuses

Status	Meaning
needs-triage	new, not evaluated yet
needs-info	blocked on a question, written under Comments
ready-for-agent	fully specified; an agent can do it end to end
ready-for-human	needs the user: a decision, an account, or physical work
needs-hardware	specified, but the hardware is not here yet
in-progress	someone, human or agent, is on it
done	all acceptance criteria met
wontfix	dropped; the reason is in Comments

The five triage roles (needs-triage, needs-info, ready-for-agent, ready-for-human, wontfix) match the vocabulary of the engineering skills the repository uses; needs-hardware, in-progress and done are lifecycle states on top.

Working an issue

1. Pick one with --next (agents) or --human (you). Never start an issue whose blockers are not done.
2. Set Status: in-progress.
3. Build the slice, keep the tests green, commit in small Conventional Commits that reference the id (Refs: 051).
4. Tick the acceptance criteria, set Status: done, and append a dated note under ## Comments: what changed, which commits, anything the next issue should know.
5. If the spec is wrong or unclear, set needs-info, write the question under ## Comments, and stop.
6. Out-of-scope discoveries become new issues with Status: needs-triage.
7. Run python3 roadmap/board.py --check after editing any issue file.

Two rules for agents working alone: never take a ready-for-human or needs-hardware issue, and never bypass the non-negotiables (tier-2 approval, simulation-first robot, no secrets in tracked files) to close one faster.

Where the plan is explained

• ARCHITECTURE.md: every decision with its reasoning, the principles, the layer roadmap and a dated change log.
• ADRs: one file per significant decision; 0006 is the private-first topology this site is organized around.
• CONTEXT.md: the glossary. Issues, code and docs use its terms exactly.
• CLAUDE.md: the instructions coding agents follow in this repository.