Changelog

The full, versioned history lives in the repo's CHANGELOG.md. This page mirrors everything unreleased on dev.

Unreleased

Nothing unreleased — the latest is 0.3.7.1.

[0.3.7.1] — 2026-09-27

Fixed

• Narration-led action envelopes now execute (#426). A real user turn returned a perfect Jev 2-hop plan and executed nothing — the parser only unwrapped a fence spanning the WHOLE output. Fenced envelopes are now found anywhere in the output (last fence first); surrounding prose rides the reasoning channel.
• Action turns that do nothing warn inline (⚠️, persisted) instead of completing silently.
• Raw JSON no longer streams into the chat on narration-led action turns — suppression recognizes the opened ```json fence progressively.

[0.3.7.0] — 2026-09-27

Added

• Versioned product skill (#411, #412). The bundled skill went multi-file — lean SKILL.md overview + feature catalog, references/protocol.md (action protocol verbatim), references/features.md (full shipped-behavior canon). Installs/updates proactively at chrome.runtime.onInstalled (MCP write_file + canary read-back); the lazy first-action-turn check remains the fallback. Per-file install status in the About card + inspector; the prompt-budget gate pins no per-turn growth for feature knowledge.
• Read-turn skill pointer (#414), probe-decided. Live probe: Zo reads the skill when a turn mentions it; with no mention it never consults workspace skills (and answered a feature question with a path contradicting the human-checkpoint design). A one-line pointer rides read turns on a verified install only.
• persona_id lane (#415). One Zo persona per builtin Mode carries that Mode's exact systemPrompt server-side (spike-verified composition + rotation). Inline system drops only on a byte-equal verified match with no user routing — everything else rides inline (fail-closed). Inspector mirrors the decision.

Changed

• Legacy repo skill/ companion retired (#413) — absorbed into the versioned bundled skill; presets live on as the builtin Modes.

[0.3.6.2] — 2026-09-26

Fixed

• Diagnostics share transport (v0.3.6.1 follow-up). dpaste.com rejected the upload (HTTP 400) because the form body went out without the urlencoded content-type header — now sent. The dead 0x0.st fallback is replaced with paste.debian.net (anonymous JSON API, 24-hour expiry, pastes hidden by default). Both hosts live-verified.

[0.3.6.1] — 2026-09-26

Fixed

• Navigate actions now actually navigate. The sidepanel's direct NAVIGATE path (a navigate action outside handoff/recipe flows) has been dead since the first commit: the panel sent no tabId, the background could not infer one (the panel is an extension page — no sender.tab), and the rejection was swallowed while the panel still rendered Zo's "Navigated…" done text. Now the panel sends the capture's source tabId, the background falls back to the active non-extension tab when it is absent, and a failed navigation renders a persisted error line instead of the done-response lie. Regression-tested end-to-end (e2e/03-actions.spec.ts).

Added

• User-triggered diagnostics sharing (Settings → Features → Debug). Next to 📋 Copy diagnostics there is now 🔗 Share diagnostics (24h link): one click composes the anonymous, metadata-only bundle (lib/debug-share.js — debug ring + settings allowlist only; page text, prompts, tokens, browsed URLs, conversation ids, and identifying config never included) and uploads it to dpaste.com (fallback 0x0.st) with a 24-hour auto-expiry and copies the link. Nothing is sent unless you click; hosts need no account/auth.
• Observability: navigation and Jev now leave evidence. With Debug mode on, the diagnostics ring records every NAVIGATE outcome, each Jev hook invocation (pick / resolve-pick / done-gate — served or refused, confidence, latency), and diagnostics shares.
• Accessibility round 2 (bash). Ledger round 2 in docs/qa/accessibility-review.md (#13–#18, all fixed in-round): the write-assist widget's shadow-DOM controls now show a per-theme focus ring (the page ring can't cross the shadow boundary); a backgrounded chat that is still generating states "— generating…" in the tab's accessible name instead of a color-only pulse; history-card glyph buttons and the rename input carry explicit accessible names; the per-turn context-tier chip announces its decision + reason; .btn-sm floored to the 24px hit-target.

[0.3.4.3] — 2026-09-22

Fixed — !handoff bug bash (#368–#372, PRs #373–#377)

Second owner-directed 0.3.4 stabilization round, over the Lane E delegate loop (evidence: tests/test-prompts/probe-handoff-bash.ts).

• Prose-only turns no longer strand the run (#368, PR #373) — blocked with the prose as the reason; ▶ Resume continues.
• Resumed runs re-prime with full context (#369, PR #374) — no more tier-0 crawl on the resume path (#351 class).
• One live run per chat (#370, PR #375) — a second !handoff refuses, mirroring compose.
• No Jev click machinery on readonly runs (#371, PR #376) — prompt vocabulary + executor pick calls suppressed run-scoped.
• Orphan sweep clears the ▶ badge (#372, PR #377).

[0.3.4.2] — 2026-09-22

Fixed — prompt-efficiency bug bash (#355–#358, PRs #359–#362)

Owner-directed bash over every prompt the extension sends to /zo/ask (evidence harness: tests/test-prompts/probe-bloat-0342.ts).

• Jev-Assisted block gated on the post-downgrade decision (#355, PR #359) — no dead 148-token action block on read-downgraded Co-browse turns.
• Visual mode drops the tier-2 selector lists (#356, PR #360) — domSections knob omits ~1.3k dead Elements/Forms tokens from read-only screenshot turns.
• Jev-Assisted block tightened 148→123 tokens (#357, PR #361) — guardrails test-pinned.
• One-shot workspace-write prompt deduped (#358, PR #362) — three inline copies → buildWorkspaceWritePrompt(); byte-identical, anti-drift tested.

Evals: visual-describe cache refreshed live (PR #363); 24/24.

[0.3.4.1] — 2026-09-22

Fixed — compose creator stabilization 1 (#351, PR #352)

Diagnosed from a real !recipe compose file an RTI run (4.3-minute turn at tier 0): the sticky DOM toggle capped the compose priming turn to a URL-only pointer, Zo drove blind — fetching pages itself, reading the protocol skill via read_file (came back polluted with workspace index content), and reconstructing the action grammar from repo scrapes. The compose bang also double-added the user bubble.

• Run-priming turns bypass the sticky DOM cap — compose and handoff priming sends attach the Mode's full context regardless of the toggle; the prompt inspector mirrors it (preview parity). Manual chats keep the cap — token discipline stays the user's call.
• Compose turns keep the FULL action tail (noSlimTail) — no slim "go read the skill" pointer, so a polluted/failed workspace read mid-compose can't happen; the grammar rides in-prompt. Panel preview mirrors.
• Single user bubble for the compose bang (the standard send path already renders it).
• Compose continuations clamp capture to tier ≥ 2 — parks and cue re-planning always see elements + forms.

Tests: 1519 unit/integration across 64 files; e2e gate added (e2e/27-recipe-compose.spec.ts: DOM off + compose → the priming ask carries ## Elements, one user bubble).

[0.3.4.0] — 2026-09-21

Added — Jev support + settings rationalization (0.3.4 slate: #339–#343, PRs #344–#348)

Five lanes from the 2026-09-20 owner intake, planned on a live comparative probe (Jev 340–512 ms vs Zo 15–36 s per decision, 40–74×, 4/4 agreement) and shipped one full zo-loop cycle per lane; milestone 0.3.4 drained. Stabilization ships as 0.3.4.N points on bug reports.

• Settings: Zo username + token (#339, PR #344). The Connection pane asks exactly two things — your Zo username slug and your access token; deriveZoHosts() fills the Zo.space endpoint and Zo web origin live (hand-edited Advanced values win, override-not-rewrite). The three endpoint fields moved into a collapsed Advanced section, and the owner-specific default space endpoint is GONE — fresh profiles get none, and space-backed features say so instead of silently querying someone else's tenant.
• One save (#340, PR #345). Exactly one Save Settings — sticky at the form end, visible from every tab, carrying the dirty marker. The duplicate in-card submit and the Prompts editor's scoped Save are gone; the editor's draft persists via the global save, and switching Modes auto-persists the outgoing edited draft (a failing draft blocks the switch instead of losing edits).
• Jev foundation (#341, PR #346). lib/jev.js (pure): decide-request builder, never-throw response parser, per-type confidence routing (noul vs choice confidences are not comparable — vendor model notes), question builders (click-choice, done-gate, cue-match), and the redactStateForJev boundary. Config: key + endpoint ride storage.local; enable/model/thresholds ride storage.sync. The options ⚡ Jev card ships DARK — off, no key — with a one-question Test probe (JEV_TEST) that works before opting in.
• Jev fast path (#342, PR #347). Done-gate: before a chained handoff turn, Jev (noul ≥ jevDoneConfidence) answers "is the goal already achieved on this page?" — a confident yes completes the run without the Zo round-trip (compose runs exempt). Click-pick: a failed click gets ONE Jev choice over the tab's clickable candidates (≥ jevPickConfidence) — a confident winner executes; anything else falls back verbatim. Every state passes the redaction boundary; timeline cards carry ⚡ Jev pick / ⚡ Jev fallback provenance. With Jev off, every flow is byte-identical.
• The marriage — Zo drives Jev (#343, PR #348). A config-gated prompt section (## Jev-Assisted Steps) teaches Zo the {type:"click", pick:{question}} vocabulary (default prompts byte-identical — offline evals 24/24 from cache); the executor's two-pass structure resolves picks in-page over a fresh candidate inventory and re-enters EVERY rail (sensitive submit probe, post-fill backstop, handoff boundary) on the resolved click — a Jev resolution can never bypass a gate. Low confidence parks the step: runs re-plan via the next Zo continuation (siblings continue), plain chats show the honest failed card.

Adversarial review caught two real defects pre-merge: CI's #309 hit-target sweep flagged the Jev card's sub-24 controls (fixed), and the J3 review round closed the J2 raw re-execution bypass (resolved clicks now re-enter every rail). Tests: 1516 unit/integration across 64 files (~4966 expects); 121 Playwright e2e across 43 numbered specs + 4 demo-gated. Live probe: tests/test-prompts/probe-jev.ts.

[0.3.3.0] — 2026-09-20

Fixed — UX bash (0.3.3 slate: #296–#315, PRs #316–#335)

Twenty atomic UX/a11y tickets from the 2026-09-19 comprehensive audit (static walkthrough + a Playwright probe driving the real extension through ~45 states × themes with a computed CSS audit). Every ticket ran the full zo-loop factory cycle with adversarial review; milestone 0.3.3 drained. Known issues will be addressed in 0.3.3.N stabilization points.

• Header & chrome (#296, #297, PRs #316, #317). At dock width the header brand collapses to the icon so the page title keeps a ≥16-character budget (e2e/28); controls-bar labels bumped to 11px over a readable token, gated by a committed CSS-audit walker (e2e/helpers/css-audit.ts — computed WCAG contrast per theme, transitions frozen, color(srgb) parsing).
• Messages & footers (#298–#302, PRs #318–#322). A --text-soft readable floor (≥4.5:1) across all seven theme blocks with --zo-muted-foreground repointed — footer time/chips/reasoning/system notes clear WCAG AA in every theme (e2e/30); active-chat stream errors now persist as {role:'error'} records and re-render as the error card with working Retry after reload (e2e/31); handoff/compose continuation turns stamp the capture tier they actually used, so the context-tier chip appears exactly where users cannot watch the capture (e2e/32); the recipe run line is ONE live element (started → progress → terminal — inversion structurally impossible) with an inline ≥24px stop control; Skip on parked actions posts a persisted system note (count + action types; boundary wording for handoff parks).
• Composer & pickers (#303–#306, PRs #323–#326). The three set-once toggles render as one wrapping chip row instead of three permanent rows (e2e/33); all three picker popups carry a keyboard hint footer, an amber accent edge on the active row (≥3:1 non-text contrast), and aria-activedescendant wiring (e2e/34); same-title tabs get VSCode-style path disambiguation in the strip, the @ popup, AND the prompt manifest — unique titles byte-identical (e2e/35); the prompt inspector gains a Copy button (raw prompt to clipboard) with letter-spacing pinned normal (e2e/36).
• Accessibility (#307–#310, PRs #327–#330). Every icon-only control carries an accessible name, swept in e2e across panel + options (e2e/37); connection status mirrors into an sr-only text twin (WCAG 1.4.1, e2e/38); one 24px hit-target floor over 23 sub-24 controls — the sweep also caught the header theme button rendering the options switch's knob (e2e/39); one global :focus-visible ring, all outline: none strips removed (e2e/40).
• Settings & onboarding (#311–#313, PRs #331–#333). The prompts editor LIVE PREVIEW is pinned populated-on-load + Mode-switch (e2e/41); the Theme card moved to the About pane as "Appearance" with live-sync intact (e2e/42); the first-run tour now carries an in-card Open settings button and references no nonexistent gear icon or Test Connection control.
• Recipes surface & i18n readiness (#314, #315, PRs #334, #335). Recipe-library rows keep ▶ Run inline and move Save/Export/Rename/Delete into a per-row ⋯ overflow (Delete's two-click confirm with red styling lives in the menu); i18n surface (a) — onboarding, error cards, empty state — extracted via tOr(key, englishLiteral) (_locales/en: 14 → 35 keys) with a census gate pinned in bun run lint (the prompt-budget-gate pattern); surfaces (b)–(e) follow as 0.3.3.N points.

Adversarial review rounds per ticket; the CI e2e job caught and fixed one wrong fixture assertion in review. Tests: 1482 unit/integration across 63 files (~4834 expects); 106 Playwright e2e across 42 numbered specs + 4 demo-gated.

[0.3.2.0] — 2026-09-19

Added — Zo-composed recipes (#289 C1, #290 C2)

• Save a Zo-run as a recipe (#289). A completed !handoff run's flow becomes a validated draft recipe (↧ Save as recipe on the done card) — boundary parks become human checkpoints, Zo fills become params with no default.
• !recipe compose <goal> (#290). Zo walks the flow to compose a recipe and never fills or submits (enforced in code): park cards ask you for values, choices, and submits; your on-page actions become the draft's defaults.
• Rehearsal. Composed drafts verify on their first run — checkpoints cannot be skipped; passing the rehearsal marks the recipe verified.

[0.2.7.1] — 2026-09-05

Fixed — stabilization round (2026-09-05 end-to-end pass)

• Esc-to-stop works anywhere in the panel (#133). The "Press Esc to stop" gesture previously only fired while the composer was focused. Esc now cancels the in-flight stream from the panel level; autocomplete popups consume Esc first (closing a popup never cancels the stream).
• Chat-tab streaming dot is background-only (#135) — no longer renders on the tab the user is already watching.
• Handoff done no longer repeats the digest (#138) — the completion status line is compact; the deliverable renders once, as the turn's answer.
• "Fill 1 fields" → "Fill 1 field" (#139) — plural fixed on the review confirm button and the fill_form card meta.

Added — regression coverage (#137)

• e2e specs 19-error-paths, 20-chat-tabs-parking, 21-history-ops, 22-bang-commands; mock Zo unauthorized (401) + fill-slow scenarios.
• Drift: Zo MCP catalog pruned 93→79 tools upstream; baselines re-pinned (#136), snapshot floor test loosened (required-tools loop is the real gate).

[0.2.5] — 2026-08-30

Changed — mode surface rationalized (5 modes, leaner prompts)

• Lean Mode 🪶 (URL-only, no page interaction) — a new built-in Mode that sends only the URL + title (tier 0) and a fetch-it-yourself contract: Zo never sees the page, fetches the URL itself with its web tools when needed (and says so plainly when the page is inaccessible/geoblocked instead of guessing), never returns browser actions, and writes + cross-references notes when the request is note-shaped. Spec: docs/superpowers/specs/2026-08-29-lean-mode-design.md.
• Mode lineup 6 → 5 — Summarize and Research merged into Ask (they were tier-1 readers differing only in query phrasing). Chats with them active migrate to Ask on load; !summarize / !research keep working in Ask; per-mode Settings overrides migrate onto Ask only when Ask had none.
• Tier-0 honesty in every prompt — turns that attach only the URL/title no longer claim "using the page content provided"; the tail now says content was not attached and licenses Zo to fetch the URL itself (or read_page on action-mode follow-ups). Applies to all modes on tier-0 turns.
• Bang aliases trimmed — !qa (use !ask) and !dom/!ctx (use !context) removed; !help now lists the full, accurate canonical set.
• Quick-action chips fixed — chips now send their stored prompt (the click handler previously sent the display label and ignored prompt), and the defaults are cut to two non-duplicative entries (Fill forms with test data, Extract links); custom chips in Options are untouched.

Added — typed-schema coverage completed

• Every module in extension/lib/ now has a Zod contract under tests/schemas/: five new schema files (parse-output, mcp, vision, intent, zo-prompts) cover the Zo response parser's channel triple, the MCP JSON-RPC envelopes, the live /models/catalog entries + vision-gate outputs, the intent classification, and the generate-mode reply (external Zo data). Schema-conformance blocks were wired into the corresponding test files; the runtime stays plain JS by design.

Added — UX polish + context transparency

• 📷 Image toggle (send-once screenshot) — a chip at the end of the tab strip arms ONE turn with a page screenshot: no !context prefix, no Mode hunting. Arming flips the MODE dropdown to Visual (unchecking before send restores it); the send forces tier 3, shows a 📷 Screenshot pill on the user bubble, then auto-clears the toggle (Mode stays Visual). The prompt inspector mirrors the force before sending, and the capture itself stays truthful (the vision gate can still skip a non-vision model; the 📷 footer chip only lights when pixels actually rode the turn).
• Context-tier chip on every assistant footer — 🔗 URL only / 📝 Text / 🧩 Elements / 📷 Screenshot, tooltip = the context-policy decision reason; persisted on the message so history re-renders keep it. Makes the per-turn token story visible without opening the prompt inspector.
• Empty-state starter chips — a fresh chat shows four clickable starting points (summarize / !context peek / extract links / research) that prefill the composer; the card retires itself on the first message.
• Copy button on code blocks — every rendered fenced block gets a Copy button (clipboard write + label flip). Also fixes a double-escape bug where code blocks displayed literal ' entities.
• ⬇ Latest pill — appears when the chat log is scrolled away from the bottom (e.g. while reading during a stream); clicks snap back.

Fixed — 📷 screenshots never reached Zo on real Chrome

• <all_urls> host permission: chrome.tabs.captureVisibleTab requires the literal <all_urls> pattern (or an activeTab gesture) — the manifest's scoped wildcards (http://*/* + https://*/*) do NOT qualify, so every tier-3 turn silently failed capture with "Either the '<all_urls>' or 'activeTab' permission is required." and shipped text-only context while the UI implied pixels were attached. The manifest now declares <all_urls>.
• Honest failure surfacing: a failed or skipped capture on a tier-3 turn is recorded in pageContext.screenshotError (vision-gate skip, disabled setting, or the capture error itself). The 📷 Screenshot pill on the user bubble now renders only when the image actually rode the turn; otherwise an inline system warning explains what happened instead of silently degrading to text-only.

Fixed — follow-up context (token optimization)

• Send-once tab excerpts: referenced-tab manifests re-sent their 500-char excerpt on EVERY turn for unchanged pages. Tabs already sent at the same url+title now ride as a pointer-only manifest line ("already provided above") — the T-ref stays alive for read_tab escalation while the excerpt rides Zo's conversation threading. Dedup state persists per chat (tabManifestSent in the session context state); the prompt inspector preview mirrors it, so preview and send can't diverge.
• No-thread re-attach guard: same-page follow-up dedup trusted conversation_id threading even when the thread was never established (retry after a stream that died before the conversation_id echo → a fresh Zo thread holds nothing). decideTurn now takes hasThread; without a thread, action turns re-attach full context.
• Single-chunk streams rendered empty bubbles: a stream whose whole answer arrived in the PartStart event created the live bubble with no streaming span, and STREAM_DONE's markdown replace skipped it. Now rendered.
• Wrong mode chip after !mode bangs: the STREAM_DONE footer resolved the active Mode instead of the turn's (bang-overridden) mode.

Added — settings + chat-list usability

• Settings tabbed UI — the section-nav chips became real tabs (Connection / Model & Persona / Prompts / Features / Actions / About): one pane visible at a time, no page-long scroll. The last tab persists across visits; #card-* deep links still land on the right pane (hash clicks included); Save stays visible below the panes.
• Token Show/Hide — reveal button on the access-token field.
• Fixed status toast — Save feedback used to render at the very bottom of the page, invisible from the Save button; now a fixed toast.
• Unsaved-changes marker — editing flags both Save buttons with a • (autosave controls excluded); clears on save.
• Runtime version — the About card reads the version from the live manifest (was hardcoded "v0.0.1"); repo links repointed to LogicIncZo.
• Chat-list preview snippets — each history card shows a one-line preview of the conversation's opening ask (first user message, collapsed).
• Search highlighting — history search matches are <mark>-highlighted in titles and snippets.