v1.5.0

Kairo Atlas — a local, read-only interactive architecture map. Kairo
already extracts precise structural signals (the collapsed module graph,
salience, risk, checkpoint/session activity) but only stores them as
.kairo/*.jsonl, Mermaid text, and markdown — formats a human cannot
scan to answer "what are the important modules, how is this connected,
what's risky, what did the AI change?". Atlas projects those existing
signals into a navigable picture so a person can see the architecture
instead of reading raw artifacts. It is a projection (ADR-0011): no new
analysis engine, no new MCP tools, no schema changes — a view in the
existing inspect surface at kairo inspect → /atlas. Design and
trade-offs: ADR-0019.

Added

• 2D architecture map (default view) — deterministic, self-authored
  canvas renderer over the cached module graph. Pan/zoom/click; nodes
  sized by salience (degree centrality), coloured by group, risk rings,
  changed-by-AI ticks. Seeded force layout (no third-party graph library).
• 3D architecture map — a hand-written perspective projection of the
  same payload (rotate / zoom / pan via Shift+drag / reset camera /
  depth-aware selection). No three.js, no WebGL library, no CDN.
• Search — / focuses the box; case-insensitive path/module matching
  with matched-node halos, neighbour highlighting, and a results list.
• Filters — hide docs/tests/examples/generated; focus source-only,
  high-salience, high-risk, changed, checkpoint-related, session-related.
  Filtering is layout-stable (no relayout) across 2D, 3D, and hit-testing.
• Node detail panel — salience, centrality, fan-in/out, group/risk and
  changed/checkpoint/session badges, plus clickable incoming/outgoing
  dependency lists to walk the graph. Built entirely client-side from the
  delivered payload (no new endpoint, no added fields, safe DOM APIs only).
• Deterministic graph payload at GET /atlas/graph.json?kind=&top=:
  schema-versioned, nodes sorted (−salience, id), edges (from, to),
  repo-relative ids only, top-N salience cap (default 50) with an explicit
  honest truncation block. No wall-clock time, no randomness, no absolute
  paths.
• Scoped CSP — only /atlas* responses carry
  default-src 'none'; script-src 'self'; style-src 'self'; img-src 'self' data:; connect-src 'self'; base-uri 'none'; form-action 'none'
  (no unsafe-inline, no unsafe-eval, no remote origins). The rest of the
  inspect surface keeps its stricter JS-free default-src 'none' policy.
• Documentation — docs/ATLAS.md usage guide and a
  README "Kairo Atlas" section. Screenshots are intentionally not
  committed (local-first; no binary assets, no remote/CDN images).
• Tests — Atlas projection, routes/CSP, 2D, 3D, search/filters, and
  detail-panel suites asserting payload shape, CSP equality, and the
  absence of eval/remote origins/innerHTML in the renderer.

Dogfooded on the Kairo repository itself during development.

Local-first, read-only, no remote assets. Atlas is served from the
loopback inspector, never writes .kairo/, and makes only a same-origin
fetch('/atlas/graph.json').

Honest limitations. The graph is built from statically-extracted
internal relative imports (JS/TS + Python — ADR-0005); dynamic imports and
external packages are excluded by design. Nodes are directory-collapsed for
readability, and large graphs are truncated to the top-N most salient nodes
(the truncation banner says so). Atlas visualizes Kairo's deterministic
signals; it does not claim to understand every codebase.