Changelog

What shipped, and when.

Taken from the repository's own changelog. One tagged release so far; the rounds beneath it are build history, labelled as such rather than dressed up as versions.

v0.3.0

history

unreleased

Current source candidate. Not on PyPI: everything below is in this checkout and may not be in an installed wheel.

CHANGELOG.md:8

What the candidate adds

Reachable product surfaces: `neo -p "sentence"` for one-shot headless agent work, `neo -` for piped context, `neo serve` for the local agent server, and `neo acp` for editor integration over ACP v1.
A `neo capabilities` probe that compares the runtime registry against what the installed distribution actually provides, so a thinner wheel reports itself instead of failing silently.
A documentation truth gate (`python -m scripts.docs_truth`) that fails on site/doc version drift and on a feature-matrix claim with no test or report reference.

Honest limits of this candidate

No Docker-backed verification lane and no live-provider lane was run for these surfaces. The evidence is a scripted model through the real engine, plus the real server and ACP handshakes.
The agent server and ACP surfaces are new; the optional ACP filesystem, terminal, and MCP proxy methods are not implemented, and permission callbacks are supported rather than enforced by a policy.

v0.2.0

tagged

2026-09-21

The current published release. This is what `pip install neo-agent-cli` gives you today.

CHANGELOG.md:116

What the public release contains

The four layers end to end: planner, Docker sandbox, verifier-gated completion, scheduler, memory, and MCP.
The product shell: full-screen TUI with a rich REPL fallback, sessions, checkpoints, resume, undo, approval mode, skills, plugins, and connectors.
CLI citizenship: `--help`/`--version`, NO_COLOR, dynamic shell completions, self-update, uninstall, and a documented exit-code contract.

What the public release does NOT contain

The packaged `agent_sdk`, `acp`, `integrations`, `recipes`, and `extensions` payload landed after 0.2.0 and is source-only in the 0.2.1 candidate.
The `neo serve`, `neo acp`, `neo capabilities`, and `neo -p` surfaces do not exist in 0.2.0.

v0.1.0

tagged

2026-09-09

Initial tagged release: the full four-layer system, built and validated end to end.

CHANGELOG.md:8-104

The four layers

Harness — planner, step agent, and verifier-gated completion, with checkpoint/resume at step granularity and git-native output on verified fixes.
Execution — Docker sandbox with a fresh container per command, read-only rootfs, no network, dropped capabilities, and three-valued verification with flake detection.
Runtime — process-per-task scheduler proven at 45 concurrent tasks with 8 simultaneous mid-run kills, plus checkpoint/resume, an approval gate, and a per-call cost ledger.
Memory + MCP — tree-sitter code graph and SQLite decision store, exposed as five MCP tools over stdio, plus a client for consuming external servers.

Adaptive model routing

5 fixture bugs: same 5/5 success at 45% of baseline cost.
16-task expanded set: 16/16 in both arms at 39% of baseline cost, 3.3× faster wall-clock.
5 real OSS repositories: adaptive 3/5 against always-expensive 2/5, at 24% of the cost.

Validation beyond the fixture set

jaraco/path — full end-to-end run on an unfamiliar repository: verifier-gated success in one attempt, 6 calls, $0.053. The first attempt failed honestly and exposed a real harness bug, which was fixed.
python-semver — module-level definition of done, 15/15 checks.
Sandbox adversarially confirmed: 24/24 sequential plus concurrent attack suites held.

Security hardening

One real data leak found live — a task-id path traversal in task_status and the status CLI — fixed via a shared semantic guard and pinned by 101 adversarial tests across the MCP and CLI boundaries.

Known limitations, documented rather than hidden

Failure classification is deliberately not implemented; the chosen novel mechanism was routing, not repair strategy.
Free-tier endpoints made some multi-repo runs flaky. The harness refused to claim success in every such case.
SWE-bench Lite is deferred. Python-only. stdio-only MCP transport. No web UI beyond the read-only dashboard.

Rounds 1–5

history

pre-release

Build history before the tag: contracts and stubs, each module's first milestone, integration of the real stack, the resume contract, and the routing ablations at two scales.

CHANGELOG.md:106-115

What this covers

Initial cross-module contracts in INTERFACES.md, and the stub phase that let four workstreams build in parallel.
Integration of the real stack: real Docker sandbox, real scheduler, real cloud models.
The checkpoint and resume contract, verified with real process kills.
Adaptive-routing ablations at two scales, including the honest negative that drove the v2 predictor redesign.

The authoritative record is CHANGELOG.md in the repository. Per-module detail lives in each module's AGENTS.md, and the cross-module contracts in INTERFACES.md.