Skip to content

Calliope 3.0 — The Great Simplification

Latest
Latest

Choose a tag to compare

View all tags
@ragelink ragelink released this 09 Jul 05:40
v3.0.0
a70d4de
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Calliope 3.0 — The Great Simplification

The private-AI agent CLI. One terminal agent for any model backend — including the ones you run yourself. MIT, small core, no lock-in.

v3 removed more than it added, on purpose. Then it added the things nobody else has.

By the numbers

v2.5 v3.0
Slash commands 106 22 (+ flag-gated /fleet)
Config keys 56 16
CLI flags ~20 11
Source 47k LOC (+27.6k theme package) ~35k LOC, themes archived
Cold start — 75ms (compiled binary), budget-enforced in CI
Keystroke-to-paint full-tree re-render p95 2.5ms, render-isolated
Tests 4,602 claimed 3,771 hermetic, 94% coverage, 90% floor enforced

Removed

Theme packs (~180), companions/personas, HUD skins, multi-agent orchestration, an embedded API server, terminal recordings, conversation branching, bookmarks, templates, todos, profiles, background jobs, tmux integration, and a complete duplicate legacy UI. The full ledger with rationale is in the CHANGELOG. v2 configs migrate automatically.

Added

  • Single-binary installs — macOS (arm64/x64) and Linux (x64/arm64), checksummed, no Node required. packaging/install.sh or Homebrew.
  • Local models as a first-class target — tool-schema simplification, a malformed-call repair loop with grammar-constrained retries, hash-anchored edits so small models can't misapply stale diffs, compact prompts, capability probing. Verified live against gemma4:31b on Ollama.
  • Governance built in, not bolted on — every session writes a tamper-evident, hash-chained audit log (secrets redacted, on by default). calliope replay re-renders any session read-only and verifies the chain (exit 4 on tampering). calliope cost reports where money and tool time went. Budget caps halt the loop cleanly (CI-friendly exit codes). A fail-closed pre-tool policy hook plugs in external policy engines.
  • ACP agent mode — calliope acp speaks the Agent Client Protocol: use Calliope inside Zed and JetBrains, operating on live editor buffers.
  • Agents that show receipts — plan mode requires reading before proposing; plans built on nothing get marked "unverified" in the transcript. Approvals bind to execution: "go" means tools run, not prose.
  • Fleet mode (off by default, zero cost when off) — coordinate agents and operators over self-hosted IRC; the channel doubles as a live audit trail.

The launch story worth telling

v3's final week was a live-testing gauntlet where the product debugged itself:

  1. A tester asked the agent to "get situated." It promised to investigate — and did nothing. The audit log convicted it in seconds: twelve read-only calls, zero writes, then a claimed "Implemented." over an empty directory.
  2. That catch became two shipped features: plans now require evidence, and approvals now bind to execution.
  3. The retest: the agent built a complete audit-analysis tool — grounded plan with line-number citations, honest test results, real cost figures accurate to six decimals.
  4. The boss level: we flipped one byte in an audit log. The agent's tool and Calliope's own verifier convicted on the same line, independently. The audit format is now third-party-verified — by a third party the product built.

That loop — agent misbehaves, audit log catches it, harness improves, agent succeeds — is the product thesis in one story: trust is a feature you can ship.

Coming in 3.1

Unified permission decisions with legible deny reasons · cross-provider conformance testing against captured wire traffic · release signing and provenance · ACP registry listing · the return of multi-agent, riding on fleet mode.

npm install -g @calliopelabs/cli · MIT · github.com/calliopeai/calliope-cli

Assets 7
Loading