Updates

New features in Zo Co-browse, newest first. Each entry explains what shipped; open How to use it for the hands-on steps.

When Zo narrates, the plan still runs NEW

2026-09-27 · PR #427 · v0.3.7.1

A bug hunt on a real session (the "multi-hop navigate using Jev" conversation) found that when Zo explained its plan in prose before returning the action JSON, the extension dropped the whole plan: nothing executed, and the raw JSON appeared in your chat as the answer. Fixed three ways: the plan now executes no matter how much Zo narrates around it (narration moves to the collapsible thinking trace); if an action turn genuinely produces nothing, you get an honest ⚠️ warning instead of fake success; and raw JSON never streams into the chat mid-turn.

Zo now knows its own manual NEW

2026-09-27 · PRs #417–#421 · v0.3.7.0

The extension now ships a versioned skill into your Zo workspace and keeps it in step with the extension itself — installed when you install the extension, updated when the extension updates. Nothing to set up.

• Zo natively knows the product. The skill catalogues every surface — Modes, bang commands, Recipes, handoff runs, tab references, the composer pickers — plus the full action protocol. When you ask "can you automate this weekly task?", Zo consults its own manual and points you at !recipe record instead of guessing. Deep details live in reference files Zo reads only when it needs them, so your prompts stay lean.
• Leaner turns. The per-turn system line for each builtin Mode moved server-side too (a Zo persona carries it), and read-only turns got a one-line pointer to the manual — both verified live, both fail-safe: if the skill or persona can't be verified, your prompts carry everything inline exactly as before. Your own Mode tweaks and persona routing always win.
• Nothing changes if you never look. The skill lives at Skills/zo-cobrowse in your Zo workspace; the Settings → About card shows its install state if you're curious.

Diagnostics share, repaired NEW

2026-09-26 · PR #407 · v0.3.6.2

Same-day hotfix to the share button that shipped this morning: dpaste.com rejected uploads (the form was sent without a content-type header — now fixed), and the 0x0.st fallback host has shut down entirely, so the fallback is now paste.debian.net (Debian Pastezone) — same anonymous, 24-hour expiry, and pastes there are hidden by default. Nothing about what is or isn't included in the bundle changed.

Navigation that really navigates — plus a shareable debug link NEW

2026-09-26 · PR #403 + PR #404 · v0.3.6.1

A stabilization release around one honest-to-goodness bug and one new way to help us debug the rest:

• "Navigate to …" actually navigates. Navigate actions had been silently failing in the side panel since the first release — the tab stayed put while the chat still said "Navigated." The panel now tells the background exactly which tab to drive, a failed navigation shows a persisted error line instead of a false success, and an end-to-end test guards the whole path.
• Share a 24-hour debug link. With Debug mode on, Settings → Features has a new 🔗 Share diagnostics (24h link) button: one click uploads an anonymous, metadata-only timing bundle to a public paste host that auto-expires in 24 hours and copies the link to your clipboard. Nothing is ever sent unless you click; page text, tokens, browsed URLs and identifying settings are never included.
• Jev leaves evidence. With Debug mode on, the exported diagnostics now show every Jev fast-path decision (click rescues, handoff done-gates) with confidence and latency — next to the ⚡ marker you already see on action cards.
• Accessibility round 2. The write-assist popover shows keyboard focus rings on all its buttons; a background chat that is still generating announces "generating" to screen readers instead of a color-only pulse; history controls and the per-turn context chip carry real accessible names; small buttons meet the 24px touch target.

Everyone can drive it — accessibility round 1 NEW

2026-09-25 · #392–#393 · v0.3.6.0

The first dedicated accessibility pass over the panel, Settings and the write-assist popover:

• Keyboard-complete. Every control — history cards, chat-tab close, pickers, popups, dialogs — is reachable and operable from the keyboard, with visible focus rings and Esc where you expect it.
• Screen readers hear what happens. Mode changes and saves announce; sections are real tabs; toggles and Show/Hide buttons state themselves.
• Motion and contrast respect your settings. "Reduce motion" stills every animation; all text now meets WCAG AA contrast in all five themes.
• The usability walkthrough finally ran (it had been waived since v0.3.0.0) — verdicts in docs/qa/usability-walkthrough-0.3.6.md.

A security pass you can read NEW

2026-09-25 · #383–#384 · v0.3.5.0

Everything shipped since the last security audit got the same treatment, plus a document you can hold us to:

• One secret-rule for capture and recording. The recipe recorder now uses the same sensitive-field detection as page capture — a field labeled "Credit card number" never records what you typed, even when the page gives the field a meaningless name. More patterns count as sensitive (card expiry, security code, IBAN, routing number…).
• Healed recipes keep their placeholders. Fixing a step's targeting mid-run no longer saves a copy of the recipe with your real typed values baked in.
• Exported playbooks mask URLs. The SKILL.md export no longer prints full recorded URLs — anything after ? (where tokens live) is masked.
• A threat model in the repo. docs/qa/threat-model.md maps, surface by surface, what a hostile page can and cannot make the extension do — and what we accept as residual risk. Updated every release.
• Privacy page is honest about Jev. The optional Jev fast path sends small redacted page-state to TypeSafe AI; PRIVACY.md now says so (it is off by default and never carries form values).

Handoff runs you can trust NEW

2026-09-22 · #368–#372 · v0.3.4.3

A reliability pass over !handoff closed five ways an unattended run could waste your time or tokens:

• No more stuck runs. If a turn comes back as prose instead of actions (a login wall, a question), the run now stops honestly and asks for you with ▶ Resume — it used to sit "working" forever.
• Resumes see the page. Resuming a paused run (or a compose park you just resolved) re-attaches the full page — the run used to re-plan from the URL alone and crawl.
• One run per chat. A second !handoff in the same chat asks you to stop the first instead of silently fighting it over the same tab.
• Read-only runs stay read-only, cheaply. Delegate runs never spend Jev fast-path calls on clicks they aren't allowed to make.
• Honest badge. The toolbar ▶ clears when a restart pauses your run.

Leaner prompts — cheaper, faster turns NEW

2026-09-22 · #355–#358 · v0.3.4.2

A prompt-efficiency audit of everything the extension sends to Zo trimmed dead weight from several turn types — same behavior, smaller bills:

• Read-style questions in Co-browse ("Summarize this page…") no longer carry the Jev click-vocabulary block, which only makes sense on turns that can act on the page.
• Visual mode turns no longer attach the page's clickable-element and form-field lists (~1.3k tokens per screenshot turn) — a screenshot mode can't click anything, so the lists were pure cost.
• Jev users' action turns carry a tighter instruction block (every safeguard kept), and the internal save-page/save-conversation writes use one shared prompt instead of three copies that could drift.

Jev — a fast path for agentic browsing NEW

2026-09-21 · #341–#343 · v0.3.4.0

Optional, off by default: connect a Jev API key (TypeSafe AI's "System One" model) in Settings → Connection → ⚡ Jev, and the extension's mechanical decisions stop waiting on full Zo turns. Jev does not generate text — it answers bounded questions about the page in a few hundred milliseconds (live-probed at 40–74× faster than a Zo decision turn), and Zo still plans every turn:

• Done-gate. Before a chained !handoff continuation, Jev checks whether the goal is already achieved on the page — a confident yes completes the run without the extra Zo round-trip.
• Click picks. When a click's target can't be found (or Zo plans an ambiguous one as pick: {question}), Jev chooses among the page's actual clickable elements; only a confident answer executes.
• Zo stays the planner, and the rails stay in code. The sensitive-submit probe, the post-fill no-click rule, and handoff boundaries all run on the resolved click; low-confidence answers fall back to the exact previous behavior, and every fast-path decision is labeled on the action card.

Page state (never form values) goes to TypeSafe AI in addition to Zo only while the toggle is on.

Settings: just a username and a token NEW

2026-09-21 · #339–#340 · v0.3.4.0

The Connection pane now asks exactly two things — your Zo username (the slug in https://<slug>.zo.space) and your access token — and derives both the Zo.space endpoint and the Zo web origin from it as you type. The old endpoint fields moved under an Advanced section (hand-edited values still win), new installs no longer inherit anyone else's space endpoint, and there is exactly one Save Settings — a sticky bar visible from every settings tab.

A more legible, more reachable panel NEW

2026-09-20 · #296–#315 · v0.3.3.0

Twenty fixes from a full UX audit of the panel, options, and onboarding — the theme is legibility and reachability:

• Everything is readable in every theme. Footer text, chips, labels, and reasoning prose now clear WCAG AA contrast in all six themes — including sepia and the light default that had been failing quietly.
• Keyboard and screen-reader users get real affordances. Every icon button announces what it does; the connection status has a text form; a consistent focus ring shows keyboard users where they are; all 24-pixel hit targets.
• Nothing vanishes. A failed turn keeps its error card (with working Retry) across reloads; skipping parked actions leaves a note in the chat saying exactly what was skipped.
• Unattended turns are transparent. Handoff continuation turns show the context-tier chip — what Zo could see — right on the turn.
• Details that were almost right. Same-titled tabs are now distinguishable everywhere (including in the prompt Zo receives); the first-run tour points at a real Open-settings button instead of a gear icon that never existed; the prompt inspector has a Copy button; the theme setting lives under About → Appearance.

Teach by stating: Zo composes your recipes NEW

2026-09-19 · #289 + #290 · v0.3.2.0

The strongest recipe-authoring act is the one you already did: watching Zo complete a flow. Two new ways to turn that into a reusable artifact:

• Save a Zo-run as a recipe. When a !handoff run finishes, the done card offers ↧ Save as recipe. The executed steps — minus everything Zo was refused — become a draft: refused actions turn into human checkpoints, and anything Zo filled becomes a parameter you fill on every run (Zo-invented values are discarded, never saved). A cleanup pass tidies the steps; the artifact must pass the same validator every recipe answers to.
• !recipe compose <goal>. Zo walks a flow to compose a recipe and is never allowed to fill or submit — enforced in code, not by politeness. When a form blocks the path, a park card asks you to fill it on the page; when two controls are ambiguous, you pick; submits stay yours. Your on-page actions become the draft's defaults. !recipe compose stop ends it.
• The first run verifies. Composed drafts are marked 🤖 composed · unverified. Their first replay is a rehearsal — checkpoints cannot be skipped — and passing it marks the recipe verified. A failed or abandoned rehearsal leaves it a draft, honestly.

Recipes go portable — workspace write-back, library panel, skill export NEW

2026-09-19 · #256 + #257 · v0.3.1.0 / v0.3.1.1

The recipes story from September 14 closes its roadmap — your learned flows are no longer stuck in one browser:

• Save to workspace. Any learned recipe writes back to /home/workspace/recipes/<name>.json — plain, inspectable JSON that follows you across machines. The learned-recipe card offers the save; an existing target asks before overwriting; drifted content bumps the patch version so the newest artifact always wins.
• Healed cues persist. When a run self-heals a recipe that came from the workspace, the completion card offers ↥ Save healed cues — the fix is patched into the source file (parameters intact), so the next run from that file doesn't pay the heal cost again.
• 🧾 Recipe library panel. A new button by the composer opens every saved flow: run it, rename it, delete it (local only — workspace files are never touched), import a workspace JSON (invalid files show the validator's exact errors), or export it.
• SKILL.md export. One click bundles a recipe to /home/workspace/Skills/<name>/SKILL.md following the Zo skill format, so Zo itself can suggest a saved flow when it recognizes the task. Documentation only — flows still execute in the extension's deterministic player, never Zo-side, and captured values are redacted before anything leaves the browser.
• v0.3.1.1 polish (stabilization round): !recipe list shows where each recipe lives and its last-run status; library rows badge the last run; the popup links to the guide.

How to use it

Save a learned recipe to the workspace — right after recording, the "🧠 Learned locally" card offers the one-click save, or from the command line:

!recipe save my-flow                    ! → /home/workspace/recipes/my-flow.json
!recipe save my-flow recipes/custom.json --force   ! custom path, overwrite

Open the library: press 🧾 Recipes by the composer's option row. Each row shows name, version, steps, required parameters (*), where it lives (💻 local vs 🌐 workspace), and its last-run status. ▶ Run starts it (the params card appears for required parameters), ✎ Rename edits inline, 🗑 Delete removes the local entry after a confirming second click.

Import from the workspace: paste a path into the popup's /home/workspace/recipes/… footer and press ＋ Import — the file is validated before it lands, and bad files explain exactly which rule they broke.

Export a skill: ⤓ Export on a row writes /home/workspace/Skills/<name>/SKILL.md (+ references/recipes.md) — give it a skim before sharing: values are redacted, but labels and checkpoint text come from the pages you recorded. Full details in the recipes guide.

Recipes — repeatable multi-page workflows

2026-09-14 · #220

A recipe is a saved, multi-page workflow you can replay anytime — the extension drives your tab step by step (navigate, fill, check, attach, extract), and pauses at the steps only a human should do: captcha, OTP, payment. Nothing is improvised: playback is deterministic, so the same recipe does the same thing every run.

Highlights:

• Human checkpoints by design. Payment/OTP/captcha steps are declared human steps — the run parks, shows you a card with instructions, and resumes only after verifying you actually landed where you said you would (e.g. the post-payment page). Payment submits are never auto-clicked — that's enforced in the recipe format itself, not just promised.
• Typed parameters. Fields that change per run (name, department, financial year) are ; a card prompts for them when the run starts.
• Self-healing. If a page changed and a step's cues no longer match, the run asks Zo once to repair the cues, bumps the recipe's version, and retries.
• Learn by doing. Record a flow manually once — the extension watches, writes the recipe draft, and Zo cleans it up. Sensitive pages you visited while recording become human checkpoints automatically, and sensitive field values are never recorded.
• Evidence. Steps can extract values (a registration number, an application ID) that are reported in the chat when the run finishes — and can feed later steps.
• Attach files. Steps can attach a workspace file (PDF, image, up to ~1 MB) to a form's file input.
• Zo drafts field values at run time. A fill step can be marked generated: when the run reaches it, Zo writes the field's content fresh — from a prompt in the recipe and optionally a workspace document — while your contact details stay static. Generated values can require your review before they're filled.

How to use it

Run a recipe (a JSON file in your Zo workspace, e.g. recipes/rti-filing.json):

!recipe run recipes/rti-filing.json

• If the recipe has parameters, a card appears — fill them and press Start run.
• The 🧾 progress line shows where the run is; ✕ stop aborts.
• At a human checkpoint, do the manual part on the page (pay, enter the OTP), then press Done — verify. If the automatic check can't confirm your step, Skip check continues anyway.
• Extracted evidence lands in the chat when the run finishes.

Learn a recipe from a manual run:

!recipe record my-flow

Click through the flow yourself (multiple pages are fine), then press ✕ stop on the ⏺ recording line. The learned recipe is saved locally — replay it with !recipe run my-flow. Sensitive pages you visited while recording become human checkpoints automatically.

Manage:

!recipe list        # what's saved (workspace recipes load by path)
!recipe stop        # stop the live run (or an in-progress recording)

A recipe is plain JSON — steps, cue arrays (label/question/aria/selector — never a single brittle selector), and . Author one in /home/workspace/recipes/ and run it by path.

Have Zo draft a field at run time (#228) — static where you want it, generated where you don't:

{
  "type": "fill",
  "cues": [{ "strategy": "question", "value": "RTI Application text" }],
  "evidenceKey": "application_text",
  "generate": {
    "prompt": "Draft an RTI application to the PIO of {{department}} requesting the FY {{fy}} annual report. Formal, first person. Applicant: {{applicant}}.",
    "maxChars": 3000,
    "contextFile": "/home/workspace/notes/draft-points.md",
    "review": true
  }
}

review: true shows you the draft (editable) before it fills; generated text lands in the run's evidence either way.

0.2.9 — quick wins v0.2.9

2026-09-14

Four quality-of-life features shipped together:

• Read workspace files in chat — Zo can pull a file from your workspace into the conversation on demand (read_file), so "summarize the notes I saved" just works.
• Pinned chat tabs — pin the chats you keep coming back to; they stay first, survive restarts, and restore where you left them. Chats also export as Markdown.
• Export everywhere — !export saves the conversation as Markdown, the page as a note, or opens a reader-view PDF; chats can be written straight to your Zo workspace.
• Write-assist, live — the ✎ Zo icon on text fields now streams its suggestion as it's written, and follow-up chips (Shorter / Formaler / your own) rework the result in place.

How to use it

• read_file: just ask — "pull up notes/rti-checklist.md and summarize it". Zo fetches it from the workspace itself.
• Pin a chat: right-click a chat tab → pin. Pinned tabs sort first and survive browser restarts.
• Export: !export (Markdown download) · !export page (page as note) · !export pdf (reader-view PDF) · !export recipes/my-note.md (write to your workspace).
• Write-assist: focus any text field, click the ✎ Zo icon, describe the improvement — the result streams in live. Use the chips to iterate (Shorter, Formaler, or a custom instruction), then Accept to fill.

0.2.8 — the stabilization bash v0.2.8.x

2026-08 → 2026-09-14

Fifteen point releases (0.2.8.0 – 0.2.8.15) that verified and hardened everything shipped in 0.2.0–0.2.7: honest screenshot failures, Zo web deep links (open any conversation on zo.computer), streaming and form-fill fixes, and a QA-agent pass that closed every finding it filed. No new features by design — this was the "trust what's there" release train.

0.2.7 — handoff: delegate a goal to Zo v0.2.7

2026-09-03

!handoff <goal> gives Zo a goal it works unattended and read-only: it navigates your tabs, extracts, compares — and reports back a digest. Boundary rules park anything it shouldn't do (clicks, fills) for you to run from the review card; a ▶ badge shows while it works; runs survive service-worker restarts and can be resumed.

How to use it

!handoff compare the pricing across these five product tabs

• Zo works the pages read-only and posts a digest when done (or when it needs you).
• ✕ stop on the progress line aborts; ▶ Resume continues a paused run.
• Anything parked by the read-only boundary lands in the review bar — you run those yourself.

Earlier highlights v0.2.0–0.2.5

• Form-fill with review (0.2.0) — Zo batches form fills by question text; sensitive forms (password/card/CVV) gate behind a review card, and Zo never clicks submit for you.
• Chat tabs (0.2.2) — several conversations open at once; streams survive switching.
• Modes, rationalized (0.2.5) — five built-ins (Co-browse, Ask, Extract, Visual, Lean 🪶), editable in Settings, plus the !context one-turn attach and the prompt inspector.
• Vision gate (0.2.4) — the 📷 toggle sends a real screenshot; models without vision are flagged honestly.