Skip to content

roadmap.md

Latest commit

 

History

History
101 lines (87 loc) · 5.22 KB

roadmap.md

File metadata and controls

101 lines (87 loc) · 5.22 KB
Raw

Roadmap

Done

Real cluster access

  1. client-go + informer cache — internal/k8s replaced the mock as the default backend. Both implement domain.Source; the UI is unchanged and untouched by which one it has. See backends.md.
  2. Streams — logs follow (l), a real interactive exec session with a raw resize-aware TTY (s), and real SPDY port-forward (p).
  3. Real AI — internal/ai posts to the OpenAI-compatible or Anthropic endpoint per /settings, with cluster context injected into the prompt.
  4. Config file — see config.md.
  5. Self-update — internal/update installs the newest GitHub release over the running binary: checksum-verified, atomic rename, offer to restart. A once-a-day startup check reports a newer version and nothing else. internal/version carries the -ldflags stamp. See update.md.

Performance

The first live version was unusably slow. Fixed by lazy per-kind informers, a render path that does no I/O, and cheap background counts — startup went from seconds to microseconds. The causes and the regression guards are written up in performance.md, because each fix is easy to undo by accident.

UX

  • Search — ctrl+p palette over kinds and objects; f finds rows in the open table; the resource list is type-to-filter. Neither side pane spends rows on a permanent search box any more.
  • Namespace / context pickers — ns … ▾ in the banner opens the Namespaces table (enter switches and shows pods); :ns and :ctx open compact popups. Neither asks you to type a name.
  • Theme picker — /theme previews live before committing.
  • No setup screen — first run opens the cluster. The CLI name shown in hints (kubectl/k8s/k/custom) defaults to kubectl and is changed with /settings.
  • Copy mode — ctrl+s releases the mouse so the terminal can select text, the only way to copy out of a mouse-capturing TUI.
  • Table — dim row numbers in the gutter, natural A→Z sort by default (events stay newest-first), an honest loading state instead of "no resources found".
  • Actions pane — lists only the actions that apply to the selected kind.
  • Panes — clicking blank space or scrolling selects a pane; enter or a double-click opens logs, falling back to describe.
  • Command prefixes — / for the cluster, : for k10s itself, each with its own popup listing exactly what it can do.

Known limits

  • Cmd+K cannot be bound. macOS terminals consume Cmd themselves and never write it to the TTY; bubbletea has no Super/Cmd key at all. iTerm2, WezTerm and Ghostty can be configured to send \x10 (ctrl+p) for Cmd+K; Terminal.app cannot.
  • The palette's object search covers loaded kinds only. Kinds not yet opened match by name. Searching their objects would mean watching the whole cluster — see performance.md.
  • Updates are checksum-verified, not signature-verified. checksums.txt proves the download matches what the release published; it proves nothing about who published it. Shipping a public key and checking a cosign or minisign signature is the fix, and is worth doing before this is used anywhere that matters.
  • No releases published yet. Updates come from p10node/k10s (update.DefaultRepo), but this tree has no tags, so /update reports "no published releases" until just tag v0.1.0 pushes one. A fork can point elsewhere with K10S_UPDATE_REPO / update.repo.
  • The API key is stored as plain text in ~/.k10s/config.yaml (mode 0600), masked only in the UI. Swapping in an OS keychain behind the same AI.APIKey field is the fix if that's not acceptable.
  • Untested against a real cluster by its author. Every live path is implemented against the real client-go/kubectl APIs and covered by tests using fake clientsets, but no run against an actual API server has been performed here. Smoke-test on a kind cluster before trusting it in prod.

Open decisions

# Question Current
1 Pane ratio — widen either side? 22 / auto / 24
2 Default theme tokyo-night

Possible next steps

The backlog now lives in plan.md as task cards — each one carries its own goal, files, design notes and acceptance criteria, so it can be picked up (by a person or an agent) without further context. Twenty-six cards across four phases:

  • P0 — trust & correctness. A kind-cluster e2e suite, a container picker (logs currently always read Containers[0], which is the wrong container on any pod with a sidecar), per-object events, CLI flags, and a read-only mode.
  • P1 — daily driver. Column sort, multi-select bulk actions, log grep/--previous, a port-forward manager, secret decode, saved views.
  • P2 — differentiators. An owner tree, an AI that actually reads describe and events (with redaction before anything leaves the machine), a pulse dashboard, can-i, diff-before-apply, Helm.
  • P3 — security, polish, distribution. Signature verification, keychain, custom keybindings and columns, export, packaging.