Roadmap

The releases planned from 0.16.0 to 0.23.0, agreed with the operator on 2026-10-01, have all shipped, and so have 0.23.1 and 0.24.0 (the inquiry verb), on 2026-10-12, and 0.24.1, a patch of small items, and 0.24.2, command time limits, on 2026-10-14. What each carried is in CHANGELOG.md. An approved run outside the sandbox, chosen by the operator on 2026-10-14, and the documentation site are built for the next release, which the operator has chosen to call 0.25.0 rather than a patch, 0.24.3. Each release carries one or two larger items and a few smaller ones; a small item sits with the larger one it touches. These are plans, not commitments: an item may move when its work shows it should, and the page is edited when it does. When an item ships it leaves this page for the changelog, and anything not yet scheduled stays in backlog.md. Each release’s preflight runs the eval’s floors, as a guard against regressions, not a measurement.

0.25.0

• An approved run outside the sandbox (ADR 0064, chosen by the operator on 2026-10-14): run_command’s outsideSandbox, asked of the person every time for that exact line, once, behind the operator’s commandPolicy.sandbox.allowOutsideSandbox. Probed on 2026-10-14: cargo install, rustup update, and every Homebrew write are refused by the sandbox, and no profile can fix Homebrew, whose own sandbox cannot nest in wisp’s. Amended the same day: the person runs one such command in chat as !!COMMAND, which wisp-tui shows in a muted red input mode.
• The documentation site (ADR 0065): an mdBook over docs/, at https://wisp.pidster.com, with new guides ahead of the reference.
• Homebrew’s not writable lines read as the sandbox’s refusal, so the model is told it may ask to run the command outside the sandbox (ADR 0054, amended 2026-10-14).

Not scheduled

Every item in backlog.md not yet given a release, grouped by what it waits on.

Waiting on data:

• a small specialised distiller, and a tool-choice classifier, once the audit log holds enough labelled pairs (Falcon-H1-Tiny-Tool-Calling-90M, pulled on 2026-10-04, is the experiment for the second);
• classifiers from other providers: embedding nearest-neighbour over labelled examples, or an external process speaking JSON Lines; and classifiers for other tasks: personal data for redact, log-line categories.

Waiting on a measurement run:

• the rest of what ADR 0052 lists under “What 0.20.0 must measure”, beyond the Qwen3-1.7B-4bit comparison 0.20.0 ran: prefix reuse, the cache’s real cost per token, the windows sized for common models, the pull against the real Hub, and Core AI past its bundle’s window;
• llama.cpp and LM Studio against a real llama-server and LM Studio, neither installed here (ADR 0058).
• the verification pass’s model step (verification.mode: model), built and off by default: it is turned on only when, on the default model, it flags no honest reply and a verdict takes under 3 s; on the on-device model on 2026-10-12 it flagged 6 of 14 honest replies (p50 1.9 s). It waits on a more capable model, or a narrower prompt that leaves plans and explanations alone, measured with scripts/check eval compare (ADR 0060, “Measured, 2026-10-12”).

Waiting on the operator:

• the hand test of the client’s question form (ADR 0062, D8; the procedure is in mcp.md, “Questions”), still to run; once it passes, inquiry.elicitation may default on. Whether a question’s banner should then wait for the dialog as an approval’s does (approval.bannerDelaySeconds) is open: today it is posted at once, and waiting would need an amendment to ADR 0062.

Waiting on signing (a Developer ID or App Store signature):

• a notification helper app, so notifications come from wisp itself, with its icon and action buttons;
• a DMG, with the package declaring where its parts are;
• Private Cloud Compute, whose entitlement Apple grants only to signed App Store apps.

Waiting on someone asking:

• tools for embeddings and reranking;
• reverse delegation through MCP sampling: the local model asks the calling agent’s model when it is stuck on a sub-step;
• the deferred uses: bulk classification, and offline work (git chores shipped on 2026-09-24 as draft_change and wisp draft).