# Release Round, Late September 2026: 57 Releases, a Small Model for Small Calls, and Developer Time Tracking

> Fifty-seven releases across nine products. Octomind 0.51–0.55.2 shipped tool authorization, Jev evaluation gates, and developer time tracking with human time and energy reports; Octoweb 0.15–0.16 added page memory and a browser terminal; Octolib 0.40 added Claude Opus 5.5, Grok 4.7 and GPT-6 Sol and Luna.

The [early September round](/blog/release-round-early-september-2026) covered August 19 through September 6. Since then: **57 releases across nine products**.

**Small calls went to a small model, and three products started counting the person, not just the tokens.** Octomind and OctoHub shipped the evaluation chain the Octolib 0.39.0 post said was next. Octomind, Timex and TypeTab each added a way to measure your side of the work: time and energy per session, a daily energy budget, keystrokes saved.

---

## The count

| Repo                                                                | Range               | Releases |
| ------------------------------------------------------------------- | ------------------- | -------: |
| [octolib](https://github.com/Muvon/octolib/releases/tag/0.40.3)     | 0.36.2 → **0.40.3** |       15 |
| [octomind](https://github.com/Muvon/octomind/releases/tag/0.55.2)   | 0.51.0 → **0.55.2** |       11 |
| [octohub](https://github.com/Muvon/octohub/releases/tag/0.9.4)      | 0.8.0 → **0.9.4**   |       10 |
| [octocode](https://github.com/Muvon/octocode/releases/tag/0.26.4)   | 0.25.1 → **0.26.4** |        6 |
| [octobrain](https://github.com/Muvon/octobrain/releases/tag/0.14.4) | 0.13.2 → **0.14.4** |        6 |
| [octofs](https://github.com/Muvon/octofs/releases/tag/0.16.1)       | 0.15.4 → **0.16.1** |        5 |
| [octoweb](https://github.com/Muvon/octoweb/releases/tag/0.16.0)     | 0.15.0 → **0.16.0** |        2 |
| [timex](https://gettimex.app/blog/timex-1-5-0-energy-you-can-spend) | **1.5.0**           |        1 |
| [typetab](https://typetab.app/blog/typetab-1-2-0)                   | **1.2.0**           |        1 |

Tagged September 6 through September 30, except TypeTab 1.2.0 (September 5), which the previous round missed.

---

## Octomind 0.51 → 0.55.2 — small calls and a human timesheet

[**github.com/muvon/octomind**](https://github.com/Muvon/octomind/releases/tag/0.55.2) · 11 releases · [full changelog](https://github.com/muvon/octomind/blob/0.55.2/CHANGELOG.md)

Each of these started with a complaint about the work around the work.

**“It did the one thing I asked it not to.”** In 0.51.0, tool authorization checks a proposed call against your request before it runs. It's opt-in; ordinary permissions remain the deterministic boundary. The follow-up [0.51.1 notes](https://octomind.run/blog/octomind-0-51-1-release) cover the decision rules and the separate compression fallback.

**“I had to leave the conversation to run a workflow.”** Octomind 0.52.0 brought AI agent workflows from taps into sessions: `/workflow` lists or inspects a definition, and a named run starts from the same command. The `tap` tool can launch detached workflow runs whose result returns to the session inbox. [The 0.52.0 notes](https://octomind.run/blog/octomind-0-52-0-release) explain the command behavior and its tap-only limit.

**“My main model is writing a paragraph to answer yes or no.”** In 0.54.0 the supervisor can hand eight of its quick judgments to an evaluation model through evaluation-driven gates; 0.54.1 then evaluated capability routing, added calibrated checks, and shipped a typed evaluation command. With Jev, these calls return typed yes/no, choice or score results; on our own cloud, the post reports 0.5 to 1.3 seconds round trip and $0.00001 to $0.0001 per decision. The gates are off by default in the CLI. Read [the 0.54 release post](https://octomind.run/blog/octomind-0-54-agent-instincts) and [the Jev explainer](https://octomind.run/blog/jev-system-one-model-ai-agents) for setup and limits.

**“I can tell you what shipped, not what reviewing it took.”** Octomind 0.55.0 brings developer time tracking to `/report`: human time and energy per request, which `/report day`, `/report week` and `/report month` roll into project timesheets. Approve a change of 50 or more lines in under 30% of the minimum time needed to read the agent's reply and type yours, and the row is flagged as a rubber-stamp. The numbers are estimates: reading and typing speeds and the 4-DHE daily budget come from published research, and the post calls its energy weights hypotheses, not measurements. Session timing is the release's one **breaking change**; on first run your config gains a `[timing]` section with six keys for calibrating it to yourself. [How every number is computed](https://octomind.run/blog/octomind-0-55-human-time-energy) is on octomind.run.

**“A learned rule says it helped because it fired on easy tasks.”** In 0.55.0, evolution compares shadow control samples with live trials and retires candidates that regress or remain inconclusive after eight uses. Its numeric defaults are configurable in `[supervisor.learning.evolution]`; the [0.55.0 post](https://octomind.run/blog/octomind-0-55-human-time-energy) explains the trial model.

**“Thinking is using the whole output budget before the answer starts.”** `reasoning_effort = "none"` and `/effort none` set reasoning effort to none, disabling reasoning for supported providers; new configs set it for `[compression.model]`. Existing configs keep their current setting, so add the option yourself when compression is the problem.

Smaller, still nice: 0.53.0 added FastEmbed ONNX models and shares one embedding model across processes; 0.53.1 keeps session state and cost tracking from leaking between sessions. 0.54 added `/copy`, broader tab completion and compression evaluation. 0.55 renders markdown links as clickable OSC 8 terminal hyperlinks, and its system prompt now tells the model that pasted text in `<pasted_content>` tags (three or more lines) is quoted material, not instructions from you. 0.55.2 lets a WebSocket session frame set its working directory (`cwd`).

Upgrade with the installer from octomind.run:

```bash
curl -fsSL https://octomind.run/install.sh | bash -s -- --version 0.55.2
octomind --version
```

---

## Octomind Cloud, briefly

**On September 24 the free plan became hub-only access**: you reach the models from your own tools with a hub key, while cloud machines, routines, connectors and live URLs moved behind subscriptions. [The plan change](https://octomind.run/blog/free-plan-is-the-hub) has the account details.

**On October 1, Cloud reorganized around octos**: agents attached to workstreams and folders. Octos rely on the `cwd` field added to Octomind 0.55.2’s WebSocket session frame; [meet your octos](https://octomind.run/blog/meet-your-octos) explains the new layout.

---

## Octoweb 0.15 → 0.16 — remember pages, then open a terminal

[**github.com/muvon/octoweb**](https://github.com/Muvon/octoweb/releases/tag/0.16.0) · 2 releases · [0.15 notes](https://octoweb.xyz/blog/octoweb-0-15-release) · [0.16 notes](https://octoweb.xyz/blog/octoweb-0-16-release)

0.15.0 added page-text search, a Later queue, keyboard link hints, private tabs, six browser workflows and two MCP tools (bringing the server to 31). Private tabs use separate in-memory storage and aren't written to history.

**0.16.0 adds a browser terminal** (xterm, with split panes), SSH tunnel and per-site proxy routing, appearance selection, hold-to-reorder workspaces and quick slots, and automatic Homebrew installation of Octomind when opening the sidebar. Its notes cover platform requirements and the terminal’s behavior; upgrade with `brew upgrade --cask octoweb`.

---

## Octolib 0.36.2 → 0.40.3 — evaluation underneath the stack

[**github.com/muvon/octolib**](https://github.com/Muvon/octolib/releases/tag/0.40.3) · 15 releases · [0.40.3 release](https://github.com/Muvon/octolib/releases/tag/0.40.3)

The [Octolib 0.39.0 post](/blog/octolib-0-39-0-evaluation-models-jev) covers 0.36.2–0.39.0, including `octolib::evaluation` and the Jev interface. It said OctoHub evaluation support and Octomind supervisor evaluation would land next; **both arrived in this round**.

0.39.1 rejects blank provider API keys for evaluation calls. Octolib 0.40.0 added explicit reasoning disablement, Claude Opus 5.5 metadata and Grok 4.7. 0.40.1 added GPT-6 Sol and Luna; 0.40.2 added Claude Sonnet 5.5 pricing and normalized GLM-5.3 reasoning effort; 0.40.3 added gpt-6.1-sol pricing and corrected model pricing and capability metadata.

---

## OctoHub 0.8 → 0.9.4 — evaluations have an alias

[**github.com/muvon/octohub**](https://github.com/Muvon/octohub/releases/tag/0.9.4) · 10 releases · [0.9.0 release](https://github.com/Muvon/octohub/releases/tag/0.9.0)

The [0.8.0 post](/blog/octohub-0-8-0-unified-media-api) covers that release’s unified media API. In 0.9.0, OctoHub **added evaluation support and routed automatic evaluations by alias**; it also rejects requests with schemas it can't enforce. 0.9.1 adds configurable owner-queue timeouts. The remaining 0.9 releases in this range are dependency updates.

---

## Octocode 0.25.1 → 0.26.4

[**github.com/muvon/octocode**](https://github.com/Muvon/octocode/releases/tag/0.26.4) · 6 releases · [0.26.4 release](https://github.com/Muvon/octocode/releases/tag/0.26.4)

- **0.25.1 and 0.26.1:** dependency refreshes.
- **0.26.0:** ONNX provider listing and parsing; embedding providers can be shared across processes.
- **0.26.2:** verbose indexing reports per-file and per-batch progress, with checkpoints after stored batches.
- **0.26.3:** MCP tool listings include cache hints; indexing and search were optimized.
- **0.26.4:** locked dependencies refreshed.

```bash
cargo install octocode --version 0.26.4
```

---

## Octobrain 0.13.2 → 0.14.4

[**github.com/muvon/octobrain**](https://github.com/Muvon/octobrain/releases/tag/0.14.4) · 6 releases · [0.14.4 release](https://github.com/Muvon/octobrain/releases/tag/0.14.4)

- **0.13.2 and 0.14.1:** dependency, setup and project-guidance updates.
- **0.14.0:** embedding models can be shared across processes; local providers are cached and the default local model changed.
- **0.14.2:** modern MCP clients receive cache hints.
- **0.14.3 and 0.14.4:** dependency and maintenance updates.

```bash
cargo install octobrain --version 0.14.4
```

---

## Octofs 0.15.4 → 0.16.1 — three fixes after the filesystem round

[**github.com/muvon/octofs**](https://github.com/Muvon/octofs/releases/tag/0.16.1) · 5 releases · [0.16.1 release](https://github.com/Muvon/octofs/releases/tag/0.16.1)

The [Octofs 0.16 post](/blog/octofs-0-16-faster-filesystem-for-claude-code-and-codex) covers 0.15.0–0.16.0, including the MCP cache-hints change for clients using the 2026-07-28 protocol revision. 0.16.1 fixes three things: shell child processes no longer attach to a controlling terminal, filesystem tools refuse results that are too large, and search output stops inserting separators between result blocks.

```bash
cargo install octofs --version 0.16.1
```

---

## Timex 1.5.0 and TypeTab 1.2.0 — time, energy and keystrokes

[**Timex**](https://gettimex.app) · [1.5.0 notes](https://gettimex.app/blog/timex-1-5-0-energy-you-can-spend) · [**TypeTab**](https://typetab.app) · [1.2.0 notes](https://typetab.app/blog/typetab-1-2-0)

Octomind wasn't the only product to start measuring the person. **Timex 1.5.0 adds Energy**, a daily budget for demanding screen work with adjustable capacity and editable category loads; the post calls it an allowance, not a health measurement.

TypeTab, our on-device autocomplete for Mac, **got a Stats pane in 1.2.0** for keystrokes, words and estimated time saved, counted against a conservative 40 words-per-minute baseline. The pane shows today, the last seven days or all time, with a 14-day chart.

A correction: the previous round said TypeTab held at 1.1.0. It didn't — the 1.2.0 post went out September 5, two days before that round.

---

## One sweep across MCP and embeddings

**Cache hints reached three MCP servers.** Octofs 0.16.0 sends them to clients on the 2026-07-28 protocol revision, which fixed Claude Code loading no octofs tools at all; Octocode 0.26.3 includes them in tool listings; Octobrain 0.14.2 sends them to modern MCP clients.

**Embedding work crossed four repos**: Octolib 0.39.0 added a local ONNX embedding provider, Octomind 0.53.0 added FastEmbed ONNX models and shared one model across processes, Octocode 0.26.0 learned to list and parse ONNX providers and share embedding providers across processes, and Octobrain 0.14.0 shares embedding models across processes.

On Octomind Cloud, that sharing is what stopped parallel agents from each loading their own copy and running small machines out of memory.

---

## Holding, on purpose

Vext holds at 1.5.1 and Synx at 0.1.4; Pulsora and Octomind Tap had no tagged release. Outside the repos, we introduced [Cointrapper](/blog/meet-cointrapper-ai-trading-agent), the private crypto trading pool we run on our own open-source stack.

---

## How the stack fits together today

- **[Octolib 0.40.3](https://github.com/Muvon/octolib/releases/tag/0.40.3)** — Rust provider library with evaluation support.
- **[Octomind 0.55.2](https://github.com/Muvon/octomind/releases/tag/0.55.2)** — agent runtime with evaluation gates and human-time reports.
- **[OctoHub 0.9.4](https://github.com/Muvon/octohub/releases/tag/0.9.4)** — model proxy with evaluation routing.
- **[Octocode 0.26.4](https://github.com/Muvon/octocode/releases/tag/0.26.4)** — semantic code search over MCP.
- **[Octobrain 0.14.4](https://github.com/Muvon/octobrain/releases/tag/0.14.4)** — agent memory over MCP, sharing one embedding model across processes.
- **[Octofs 0.16.1](https://github.com/Muvon/octofs/releases/tag/0.16.1)** — filesystem and shell tools for MCP clients.
- **[Octoweb 0.16.0](https://github.com/Muvon/octoweb/releases/tag/0.16.0)** — browser with a terminal and site proxies.
- **[Timex 1.5.0](https://gettimex.app/blog/timex-1-5-0-energy-you-can-spend)** — time tracking with an Energy budget.
- **[TypeTab 1.2.0](https://typetab.app/blog/typetab-1-2-0)** — on-device autocomplete for Mac, now with a Stats pane.

Last round was about not making you wait. This one routed small decisions to a smaller model and started measuring your side of the work. If a release still leaves a gap in your workflow, open an issue on the project's GitHub repository and tell us where.

— Don
