agenttail v0.3.0

agenttail v0.3.0

Meet Tail, the console on stage, and honest numbers for Claude Code. The
logo's chevron grew an eye and moved into the editor: it reacts to what the
agent does, thinks out loud while the model thinks and plays games on a long
think. Still a local, read-only viewer by default.

Highlights

Tail, the mascot

• Tail lives in the editor corner while a session is in motion. Its eye
  follows the pointer; it hops when a command passes, sweats when one fails,
  cheers when a turn ends, dozes off when nothing happens and likes being
  petted (not too fast). Settings can turn it off.
• While the agent thinks, three dots rise into a thought bubble with a clock,
  the latest line of reasoning and some nonsense (catching kittens, what is
  42). Five quiet minutes and it falls asleep under a STALLED? warning.
• On a long think it plays one of ten games by itself: snake, tetris, a
  runner, chess (two famous games replayed), chomp, space invaders, a
  geometry-dash level with ship sections, breakout, flappy and a shell-command
  typing game. ▶ play hands you the controls (arrows and space; Esc gives
  them back, N for the next game), and the best score per game is kept.
• The logo is Tail now, in the README, the favicon, the wordmark and the boot
  banner.

Console on stage

• The new default: a running command brings the console up over the editor,
  an edit sends it back to its dock, and a quiet console shrinks to a rail so
  the agent pane gets the room. The console header switches back to the plain
  docked console.

Claude Code

• A cost for Claude sessions: what the same tokens would cost at Anthropic's
  API list prices, each reply priced by its own model, with cache reads,
  5-minute and 1-hour cache writes, Haiku 5.5's long-prompt rate card, Opus
  5.5 fast mode and web searches. It is labelled ≈ API: a subscription is not
  billed per token, so it is a yardstick, not your bill.
• Token and step counts were about 2.5× too high: the transcript repeats a
  reply's usage on every content-block line, and each reply now counts once.
• Claude sessions get turns, from a real prompt to end_turn or an interrupt,
  so the Log shows per-turn summaries and costs. The header names the model.
• A session stays live while the model composes a long reply: a turn is open
  until the transcript says end_turn, not just while the file grows.

Watch wall

• Every pane gets its own Tail in its agent's colour, reacting to that agent,
  with the thought bubble and the games (▶ play works there too; Esc ends the
  game before it closes the wall).
• The pane scoreboard shows TOK/S.

Everywhere

• TOK/S shows a number: output tokens per second over the last 30 seconds.
• Themes: red redrawn as crimson and bone, a new neon (synthwave) theme, a
  separate success colour so success never looks like the accent or an error,
  and editor bracket pairs in each theme's own colours. Seven in all.
• A very long running command no longer pushes the page wider than the window.
• The README has a Meet Tail section, a themes mosaic and fresh captures.

Compatibility

• Same Node 24+ requirement and log locations. Nothing to migrate.
• Claude token and step totals drop for existing sessions: that is the
  double counting going away, not lost data.

Requirements

• Node.js 24 or newer for the package/portable build.
• The Windows standalone exe embeds Node and needs no installation.
• Windows is tested. macOS and Linux are untested.

Stream-mode warning

Stream mode is on by default and redacts secrets, identity, network addresses and several opaque identifiers before anything reaches the browser. Redaction reduces risk; it is not a guarantee. Review the stream before sharing or recording it.

Known limits

• The Windows exe is unsigned, so SmartScreen may warn. Verify the SHA-256 checksum before running it.
• Only newer sessions are listed (60 dsh/Claude, 40 Antigravity).
• The Claude API-equivalent cost uses list prices (no batch, regional or long-context surcharges) and skips models without a list price.
• Peak/off-peak pricing ignores public holidays; GPU metering needs nvidia-smi.

Artifacts

• agenttail-0.3.0.tgz — npm package (not on the npm registry yet: npx github:ionut27001/agenttail).
• agenttail-0.3.0-portable.zip — source plus start.cmd / start.sh, for Node 24+.
• agenttail-0.3.0-win-x64.exe — standalone Windows x64 build (unsigned).

Verify the checksums

Get-FileHash .\agenttail-0.3.0.tgz, .\agenttail-0.3.0-portable.zip, .\agenttail-0.3.0-win-x64.exe -Algorithm SHA256

Compare the output with SHA256SUMS.txt.

License

MIT © 2026 ionut27001. agenttail is not affiliated with DeepSeek, Anthropic or Google. Product names are used only to describe supported sources.

agenttail v0.2.0

agenttail v0.2.0

Watch several agents at once, opt-in control for DeepSeek Harness, and a
deeper read side. The default is still a local, read-only viewer; control is
off unless you pass --allow-control --no-stream-mode.

Highlights

Several agents

• Agents overview (Alt+A): every agent active in the last 30 minutes across
  dsh, Claude Code and Antigravity, grouped by git repository (worktrees
  included), each with its own colour, branch, model and current action. The
  activity bar and tree show the other agents on the repo you are viewing.
• Watch wall (Alt+W): one live pane per agent with its own editor, console,
  scoreboard and context meter. The stage follows the agent, replies run as
  captions, and a conversation drawer keeps your messages and the replies.
• A shared file tree for the team, with edit counts, a dot per agent, each
  agent's current file, clash warnings and click-to-open; a timeline of every
  action per agent underneath.
• Quiet agents shrink to a slim column; a step with no output says
  waiting on the model with a clock, and stalled? after five minutes.

Control (opt-in)

• Start and own dsh --profile acp sessions: workspace allowlist, live model
  catalog (including custom providers such as Ollama), permission preset and a
  first message.
• Message input with queueing, exact-command approval cards, reject reason,
  Stop and End, all backed by ACP permissions and cancellation.
• Owned sessions are marked ◈; sessions started elsewhere stay view-only.
• Switch the model between turns from the input row; pasted images are sent
  only to agents that declare image prompts (dsh does not).

Read side

• Model/provider, permission preset and approval-policy history, per-turn
  tokens and cost in Log, a cost-per-turn sparkline, and a background job list
  that jumps to console output.
• Context meter: how much of the context window the session uses.
• Agent replies render as markdown; each turn in Log has a summary (model,
  tools, files changed, failures, tokens, cost, duration).
• The agent pane can dock as a full-height column on the right
  (Ctrl+Shift+.); repeated identical actions collapse into one console line;
  editors wrap long lines.
• Six colour themes (matrix, amber, red, ice, dracula, paper), a Settings modal
  behind a gear, and the agenttail wordmark in the header.

Security

• Per-start bearer token, --allow-control gate, stream-mode lockout, no
  auto-approve, workspace allowlist and an audit log without message bodies.
• Stream mode now also redacts file paths outside the workspace (they could
  carry the Windows user name).
• New control, agents and control-UI test suites; all existing suites
  still pass.

Compatibility

• Same Node 24+ requirement and ~/.dsh/sessions layout.
• Control needs an acp dsh profile (created on first use by dsh) and a
  control.json workspace allowlist.
• Existing read-only usage is unchanged: no flag, no control endpoints.

Requirements

• Node.js 24 or newer for the package/portable build.
• The Windows standalone exe embeds Node and needs no installation.
• Windows is tested. macOS and Linux are untested.

Stream-mode warning

Stream mode is on by default and redacts secrets, identity, network addresses and several opaque identifiers before anything reaches the browser. Redaction reduces risk; it is not a guarantee. Review the stream before sharing or recording it.

Known limits

• The Windows exe is unsigned, so SmartScreen may warn. Verify the SHA-256 checksum before running it.
• Only newer sessions are listed (60 dsh/Claude, 40 Antigravity).
• Control starts one dsh --profile acp process per owned session; sessions started outside agenttail are view-only.
• Peak/off-peak pricing ignores public holidays; GPU metering needs nvidia-smi.

Artifacts

• agenttail-0.2.0.tgz — npm package (not on the npm registry yet: npx github:ionut27001/agenttail).
• agenttail-0.2.0-portable.zip — source plus start.cmd / start.sh, for Node 24+.
• agenttail-0.2.0-win-x64.exe — standalone Windows x64 build (unsigned).

Verify the checksums

Get-FileHash .\agenttail-0.2.0.tgz, .\agenttail-0.2.0-portable.zip, .\agenttail-0.2.0-win-x64.exe -Algorithm SHA256

Compare the output with SHA256SUMS.txt.

License

MIT © 2026 ionut27001. agenttail is not affiliated with DeepSeek, Anthropic or Google. Product names are used only to describe supported sources.

agenttail v0.1.0

agenttail v0.1.0

tail -f for AI coding agents. A local, read-only live viewer for DeepSeek Harness, Claude Code and Antigravity CLI sessions.

Highlights

• Live editor typing, tree heat and per-file edit/failure tooltips.
• Console with commands, job progress ([n/N]), collapsible output and a running spinner.
• Agent chat plus a full Log trajectory view.
• Activity bar with one current state, colour-coded animation and finished/error transients.
• Replay for finished sessions.
• Stream mode redaction with fail-closed behaviour.
• No telemetry and no external requests. Monaco, xterm.js and JetBrains Mono are vendored; the page works offline.
• Read-only, localhost only, with Host/Origin checks against DNS rebinding and CSRF.

Supported sources

Source	Logs
DeepSeek Harness	~/.dsh/sessions/**/session.v4.jsonl.zstd
Claude Code	~/.claude/projects/**/*.jsonl
Antigravity CLI	~/.gemini/antigravity/**

Requirements

• Node.js 24 or newer for the package/portable build.
• The Windows standalone exe embeds Node and needs no installation.
• Windows is tested. macOS and Linux are untested.

Stream-mode warning

Stream mode is on by default and redacts secrets, identity, network addresses and several opaque identifiers before anything reaches the browser. Redaction reduces risk; it is not a guarantee. Review the stream before sharing or recording it.

Known limits

• Windows is the tested platform; macOS and Linux builds are untested.
• The Windows exe is unsigned, so SmartScreen may warn. Verify the SHA-256 checksum before running it.
• Only newer sessions are listed (60 dsh/Claude, 40 Antigravity).
• Peak/off-peak pricing ignores public holidays; GPU metering needs nvidia-smi.
• .dsh logs with corrupted zstd frames are skipped.

Artifacts

• agenttail-0.1.0.tgz — npm package.
• agenttail-0.1.0-portable.zip — source plus start.cmd / start.sh, for Node 24+.
• agenttail-0.1.0-win-x64.exe — standalone Windows x64 build (unsigned).

Verify the checksums

Get-FileHash .\agenttail-0.1.0.tgz, .\agenttail-0.1.0-portable.zip, .\agenttail-0.1.0-win-x64.exe -Algorithm SHA256

Compare the output with SHA256SUMS.txt.

License

MIT © 2026 ionut27001. agenttail is not affiliated with DeepSeek, Anthropic or Google. Product names are used only to describe supported sources.