v0.12.2

Fixed

• Bundle analyzers work now. Added a browser field pointing at a
  dependency-free stub entry (index.browser.js) — tools like bundlephobia
  that webpack-bundle the package no longer fail on child_process and the
  platform binary packages the Node entry resolves. The stub exposes
  { version } and throws clearly if the CLI surface is called in a browser.
• Publish workflow: the crates.io "already published" check queried the local
  workspace instead of the registry (cargo info resolves workspace members
  locally), silently skipping real publishes. It now asks the sparse index.

What's Changed

• fix(release): crates.io existence check queried the local workspace, not the registry by @geaglin in #51
• release 0.12.2: browser stub entry so bundle analyzers build the package by @geaglin in #52

Full Changelog: v0.12.1...v0.12.2

Fixed

• Bundle analyzers work now. Added a browser field pointing at a
  dependency-free stub entry (index.browser.js) — tools like bundlephobia
  that webpack-bundle the package no longer fail on child_process and the
  platform binary packages the Node entry resolves. The stub exposes
  { version } and throws clearly if the CLI surface is called in a browser.
• Publish workflow: the crates.io "already published" check queried the local
  workspace instead of the registry (cargo info resolves workspace members
  locally), silently skipping real publishes. It now asks the sparse index.

What's Changed

• fix(release): crates.io existence check queried the local workspace, not the registry by @geaglin in #51
• release 0.12.2: browser stub entry so bundle analyzers build the package by @geaglin in #52

Full Changelog: v0.12.1...v0.12.2

Fixed

• Bundle analyzers work now. Added a browser field pointing at a
  dependency-free stub entry (index.browser.js) — tools like bundlephobia
  that webpack-bundle the package no longer fail on child_process and the
  platform binary packages the Node entry resolves. The stub exposes
  { version } and throws clearly if the CLI surface is called in a browser.
• Publish workflow: the crates.io "already published" check queried the local
  workspace instead of the registry (cargo info resolves workspace members
  locally), silently skipping real publishes. It now asks the sparse index.

What's Changed

• fix(release): crates.io existence check queried the local workspace, not the registry by @geaglin in #51
• release 0.12.2: browser stub entry so bundle analyzers build the package by @geaglin in #52

Full Changelog: v0.12.1...v0.12.2

Fixed

• Bundle analyzers work now. Added a browser field pointing at a
  dependency-free stub entry (index.browser.js) — tools like bundlephobia
  that webpack-bundle the package no longer fail on child_process and the
  platform binary packages the Node entry resolves. The stub exposes
  { version } and throws clearly if the CLI surface is called in a browser.
• Publish workflow: the crates.io "already published" check queried the local
  workspace instead of the registry (cargo info resolves workspace members
  locally), silently skipping real publishes. It now asks the sparse index.

What's Changed

• fix(release): crates.io existence check queried the local workspace, not the registry by @geaglin in #51
• release 0.12.2: browser stub entry so bundle analyzers build the package by @geaglin in #52

Full Changelog: v0.12.1...v0.12.2

v0.12.1

Supply-chain hardening: the npm install is now inert and auditable at a glance.

Changed

• Per-platform binary packages. The prebuilt binary ships as five
  buildwithnexus-<os>-<cpu> packages selected automatically via
  optionalDependencies (the esbuild pattern). The main package is ~7 readable
  files with no install scripts, no network code, no shell-outs (beyond
  spawning the CLI itself), no bundled sources, no eval — supply-chain
  scanners have nothing to flag. Binaries are SHA-256-verified when packaged
  and carry build-provenance attestations.
• Auto-update moved into the binary. The daily npm-registry check and
  silent npm install -g refresh now run inside the CLI (background thread,
  never blocks startup) instead of the npm wrapper. BWN_NO_AUTO_UPDATE=1
  still disables installs; an update notice prints on the next launch.
• Installs with --omit=optional skip the platform binary; point BWN_BIN
  at a self-built binary (documented in the launcher's error message and at
  buildwithnexus.dev/docs/install).
• Added a main entry point (index.js) with a tiny programmatic API
  ({ version, binaryPath, run }) so bundle analyzers stop erroring on the
  bin-only package.

What's Changed

• release 0.12.1: zero supply-chain flags, crates.io publishing (cargo install bwn), CI action bumps (supersedes #47) by @geaglin in #49
• release: ship-now/tokens-later publishing (npm OIDC day one) + inline image rendering in the TUI by @geaglin in #50

Full Changelog: v0.12.0...v0.12.1

Supply-chain hardening: the npm install is now inert and auditable at a glance.

Changed

• Per-platform binary packages. The prebuilt binary ships as five
  buildwithnexus-<os>-<cpu> packages selected automatically via
  optionalDependencies (the esbuild pattern). The main package is ~7 readable
  files with no install scripts, no network code, no shell-outs (beyond
  spawning the CLI itself), no bundled sources, no eval — supply-chain
  scanners have nothing to flag. Binaries are SHA-256-verified when packaged
  and carry build-provenance attestations.
• Auto-update moved into the binary. The daily npm-registry check and
  silent npm install -g refresh now run inside the CLI (background thread,
  never blocks startup) instead of the npm wrapper. BWN_NO_AUTO_UPDATE=1
  still disables installs; an update notice prints on the next launch.
• Installs with --omit=optional skip the platform binary; point BWN_BIN
  at a self-built binary (documented in the launcher's error message and at
  buildwithnexus.dev/docs/install).
• Added a main entry point (index.js) with a tiny programmatic API
  ({ version, binaryPath, run }) so bundle analyzers stop erroring on the
  bin-only package.

What's Changed

• release 0.12.1: zero supply-chain flags, crates.io publishing (cargo install bwn), CI action bumps (supersedes #47) by @geaglin in #49
• release: ship-now/tokens-later publishing (npm OIDC day one) + inline image rendering in the TUI by @geaglin in #50

Full Changelog: v0.12.0...v0.12.1

Supply-chain hardening: the npm install is now inert and auditable at a glance.

Changed

• Per-platform binary packages. The prebuilt binary ships as five
  buildwithnexus-<os>-<cpu> packages selected automatically via
  optionalDependencies (the esbuild pattern). The main package is ~7 readable
  files with no install scripts, no network code, no shell-outs (beyond
  spawning the CLI itself), no bundled sources, no eval — supply-chain
  scanners have nothing to flag. Binaries are SHA-256-verified when packaged
  and carry build-provenance attestations.
• Auto-update moved into the binary. The daily npm-registry check and
  silent npm install -g refresh now run inside the CLI (background thread,
  never blocks startup) instead of the npm wrapper. BWN_NO_AUTO_UPDATE=1
  still disables installs; an update notice prints on the next launch.
• Installs with --omit=optional skip the platform binary; point BWN_BIN
  at a self-built binary (documented in the launcher's error message and at
  buildwithnexus.dev/docs/install).
• Added a main entry point (index.js) with a tiny programmatic API
  ({ version, binaryPath, run }) so bundle analyzers stop erroring on the
  bin-only package.

What's Changed

• release 0.12.1: zero supply-chain flags, crates.io publishing (cargo install bwn), CI action bumps (supersedes #47) by @geaglin in #49
• release: ship-now/tokens-later publishing (npm OIDC day one) + inline image rendering in the TUI by @geaglin in #50

Full Changelog: v0.12.0...v0.12.1

Supply-chain hardening: the npm install is now inert and auditable at a glance.

Changed

• Per-platform binary packages. The prebuilt binary ships as five
  buildwithnexus-<os>-<cpu> packages selected automatically via
  optionalDependencies (the esbuild pattern). The main package is ~7 readable
  files with no install scripts, no network code, no shell-outs (beyond
  spawning the CLI itself), no bundled sources, no eval — supply-chain
  scanners have nothing to flag. Binaries are SHA-256-verified when packaged
  and carry build-provenance attestations.
• Auto-update moved into the binary. The daily npm-registry check and
  silent npm install -g refresh now run inside the CLI (background thread,
  never blocks startup) instead of the npm wrapper. BWN_NO_AUTO_UPDATE=1
  still disables installs; an update notice prints on the next launch.
• Installs with --omit=optional skip the platform binary; point BWN_BIN
  at a self-built binary (documented in the launcher's error message and at
  buildwithnexus.dev/docs/install).
• Added a main entry point (index.js) with a tiny programmatic API
  ({ version, binaryPath, run }) so bundle analyzers stop erroring on the
  bin-only package.

What's Changed

• release 0.12.1: zero supply-chain flags, crates.io publishing (cargo install bwn), CI action bumps (supersedes #47) by @geaglin in #49
• release: ship-now/tokens-later publishing (npm OIDC day one) + inline image rendering in the TUI by @geaglin in #50

Full Changelog: v0.12.0...v0.12.1

Supply-chain hardening: the npm install is now inert and auditable at a glance.

Changed

• Per-platform binary packages. The prebuilt binary ships as five
  buildwithnexus-<os>-<cpu> packages selected automatically via
  optionalDependencies (the esbuild pattern). The main package is ~7 readable
  files with no install scripts, no network code, no shell-outs (beyond
  spawning the CLI itself), no bundled sources, no eval — supply-chain
  scanners have nothing to flag. Binaries are SHA-256-verified when packaged
  and carry build-provenance attestations.
• Auto-update moved into the binary. The daily npm-registry check and
  silent npm install -g refresh now run inside the CLI (background thread,
  never blocks startup) instead of the npm wrapper. BWN_NO_AUTO_UPDATE=1
  still disables installs; an update notice prints on the next launch.
• Installs with --omit=optional skip the platform binary; point BWN_BIN
  at a self-built binary (documented in the launcher's error message and at
  buildwithnexus.dev/docs/install).
• Added a main entry point (index.js) with a tiny programmatic API
  ({ version, binaryPath, run }) so bundle analyzers stop erroring on the
  bin-only package.

What's Changed

• release 0.12.1: zero supply-chain flags, crates.io publishing (cargo install bwn), CI action bumps (supersedes #47) by @geaglin in #49
• release: ship-now/tokens-later publishing (npm OIDC day one) + inline image rendering in the TUI by @geaglin in #50

Full Changelog: v0.12.0...v0.12.1

v0.12.0

The Ferrari release: a full UI/UX overhaul — instant, multimodal, and clean.

Added

• Live slash-command autocomplete. Typing / (or @, or a sub-argument)
  opens a popup above the composer with one-line descriptions for all built-in
  commands; ↑/↓ navigate, Tab/Enter accept, Esc dismisses. Removed the ghost
  commands (/effort, /plugin, /marketplace) that autocomplete offered but
  no handler implemented.
• True multimodal input. Ctrl+V pastes clipboard images (Wayland/X11/
  macOS/WSL) as attachments; @clip.mp4 (and other containers) is parsed with
  ffmpeg/ffprobe into up to 8 evenly-sampled frames plus a metadata block.
  Both are gated on a per-model vision-capability check — text-only models get
  an explicit notice instead of silently dropped images.
• Clickable files and links (OSC 8). Markdown links and ⏺ edit/write
  headers are terminal hyperlinks: click a path to open the file in the OS
  default app. The ANSI scanners learned OSC strings so links cost zero
  display columns.
• npm auto-update. A detached background process checks the registry at
  most once a day and silently installs newer versions; the next launch prints
  a one-line notice. Opt out with BWN_NO_AUTO_UPDATE=1. Startup never waits
  on the network.
• Esc interrupts the agent (with queued prompts auto-sending next turn),
  double-click word / triple-click line selection with a soft theme highlight
  and a "⎘ copied" footer flash, prefix-filtered ↑ history that preserves the
  in-progress draft, Ctrl/Alt+←→ word jumps, and mode-aware cursor shapes
  (accent bar / vim block / visual underline) hidden while the agent works.

Changed

• GitHub-grade diffs. One renderer for edit previews, write previews, and
  applied changes: dual line-number gutter, background-tinted rows, word-level
  change emphasis on replacement pairs, hunk elision, and a 40-row cap.
  NO_COLOR keeps signed - / + text.
• Bordered composer box (opencode-style) with the spinner inside showing
  elapsed seconds and an interrupt hint; minimal banner (gradient wordmark +
  aligned model/cwd/mode rows) replacing the emoji-heavy boxed header.
• Grouped, auto-aligned /help, rendered markdown for /verify, rule
  violations, /agents, and /memory (no more raw ##/**), human-readable
  verification statuses, and a consistent ✓ / ✗ / ⚠ / ⟳ message vocabulary.
• Removed fabricated telemetry: the hardcoded est. cost figure and the
  tok/s badge derived from character counts are gone; the context meter shows
  only real used/total tokens.

Fixed

• Tool denials rendered dim-grey (now red); failed background workflows were
  announced in success-green; a raw eprintln! corrupted the alt-screen
  during /resume; HTTP retries were silent for up to 10s (now a visible
  ⟳ retrying line); tool previews and the permission prompt could smear a
  multi-line command across the screen (capped at 80 cols, prompt legend on
  its own line).

Performance

• ~215× faster streaming renders. The renderer re-wrapped the entire
  transcript on every streamed chunk/keystroke/scroll; now each line wraps
  once (incremental cache), repaints are coalesced to ~60fps and applied as
  atomic frames (DEC 2026), and multi-line blocks (diffs, code, command
  output) paint in one repaint instead of one per row.
• Instant startup: dependency probes moved off the critical path; the screen
  paints chrome immediately so there is never a black frame.

What's Changed

• release: 0.12.0 — UI/UX overhaul, 215× faster renders, true multimodal, npm auto-update by @geaglin in #48

Full Changelog: v0.11.4...v0.12.0

The Ferrari release: a full UI/UX overhaul — instant, multimodal, and clean.

Added

• Live slash-command autocomplete. Typing / (or @, or a sub-argument)
  opens a popup above the composer with one-line descriptions for all built-in
  commands; ↑/↓ navigate, Tab/Enter accept, Esc dismisses. Removed the ghost
  commands (/effort, /plugin, /marketplace) that autocomplete offered but
  no handler implemented.
• True multimodal input. Ctrl+V pastes clipboard images (Wayland/X11/
  macOS/WSL) as attachments; @clip.mp4 (and other containers) is parsed with
  ffmpeg/ffprobe into up to 8 evenly-sampled frames plus a metadata block.
  Both are gated on a per-model vision-capability check — text-only models get
  an explicit notice instead of silently dropped images.
• Clickable files and links (OSC 8). Markdown links and ⏺ edit/write
  headers are terminal hyperlinks: click a path to open the file in the OS
  default app. The ANSI scanners learned OSC strings so links cost zero
  display columns.
• npm auto-update. A detached background process checks the registry at
  most once a day and silently installs newer versions; the next launch prints
  a one-line notice. Opt out with BWN_NO_AUTO_UPDATE=1. Startup never waits
  on the network.
• Esc interrupts the agent (with queued prompts auto-sending next turn),
  double-click word / triple-click line selection with a soft theme highlight
  and a "⎘ copied" footer flash, prefix-filtered ↑ history that preserves the
  in-progress draft, Ctrl/Alt+←→ word jumps, and mode-aware cursor shapes
  (accent bar / vim block / visual underline) hidden while the agent works.

Changed

• GitHub-grade diffs. One renderer for edit previews, write previews, and
  applied changes: dual line-number gutter, background-tinted rows, word-level
  change emphasis on replacement pairs, hunk elision, and a 40-row cap.
  NO_COLOR keeps signed - / + text.
• Bordered composer box (opencode-style) with the spinner inside showing
  elapsed seconds and an interrupt hint; minimal banner (gradient wordmark +
  aligned model/cwd/mode rows) replacing the emoji-heavy boxed header.
• Grouped, auto-aligned /help, rendered markdown for /verify, rule
  violations, /agents, and /memory (no more raw ##/**), human-readable
  verification statuses, and a consistent ✓ / ✗ / ⚠ / ⟳ message vocabulary.
• Removed fabricated telemetry: the hardcoded est. cost figure and the
  tok/s badge derived from character counts are gone; the context meter shows
  only real used/total tokens.

Fixed

• Tool denials rendered dim-grey (now red); failed background workflows were
  announced in success-green; a raw eprintln! corrupted the alt-screen
  during /resume; HTTP retries were silent for up to 10s (now a visible
  ⟳ retrying line); tool previews and the permission prompt could smear a
  multi-line command across the screen (capped at 80 cols, prompt legend on
  its own line).

Performance

• ~215× faster streaming renders. The renderer re-wrapped the entire
  transcript on every streamed chunk/keystroke/scroll; now each line wraps
  once (incremental cache), repaints are coalesced to ~60fps and applied as
  atomic frames (DEC 2026), and multi-line blocks (diffs, code, command
  output) paint in one repaint instead of one per row.
• Instant startup: dependency probes moved off the critical path; the screen
  paints chrome immediately so there is never a black frame.

What's Changed

• release: 0.12.0 — UI/UX overhaul, 215× faster renders, true multimodal, npm auto-update by @geaglin in #48

Full Changelog: v0.11.4...v0.12.0

The Ferrari release: a full UI/UX overhaul — instant, multimodal, and clean.

Added

• Live slash-command autocomplete. Typing / (or @, or a sub-argument)
  opens a popup above the composer with one-line descriptions for all built-in
  commands; ↑/↓ navigate, Tab/Enter accept, Esc dismisses. Removed the ghost
  commands (/effort, /plugin, /marketplace) that autocomplete offered but
  no handler implemented.
• True multimodal input. Ctrl+V pastes clipboard images (Wayland/X11/
  macOS/WSL) as attachments; @clip.mp4 (and other containers) is parsed with
  ffmpeg/ffprobe into up to 8 evenly-sampled frames plus a metadata block.
  Both are gated on a per-model vision-capability check — text-only models get
  an explicit notice instead of silently dropped images.
• Clickable files and links (OSC 8). Markdown links and ⏺ edit/write
  headers are terminal hyperlinks: click a path to open the file in the OS
  default app. The ANSI scanners learned OSC strings so links cost zero
  display columns.
• npm auto-update. A detached background process checks the registry at
  most once a day and silently installs newer versions; the next launch prints
  a one-line notice. Opt out with BWN_NO_AUTO_UPDATE=1. Startup never waits
  on the network.
• Esc interrupts the agent (with queued prompts auto-sending next turn),
  double-click word / triple-click line selection with a soft theme highlight
  and a "⎘ copied" footer flash, prefix-filtered ↑ history that preserves the
  in-progress draft, Ctrl/Alt+←→ word jumps, and mode-aware cursor shapes
  (accent bar / vim block / visual underline) hidden while the agent works.

Changed

• GitHub-grade diffs. One renderer for edit previews, write previews, and
  applied changes: dual line-number gutter, background-tinted rows, word-level
  change emphasis on replacement pairs, hunk elision, and a 40-row cap.
  NO_COLOR keeps signed - / + text.
• Bordered composer box (opencode-style) with the spinner inside showing
  elapsed seconds and an interrupt hint; minimal banner (gradient wordmark +
  aligned model/cwd/mode rows) replacing the emoji-heavy boxed header.
• Grouped, auto-aligned /help, rendered markdown for /verify, rule
  violations, /agents, an...

v0.11.4

Gemma local-model support, from a real gemma-2-2b-it session on llama.cpp.

Added

• Parse Gemma's tool_code tool-call format. Gemma emits tool calls as a
  ```tool_code fenced Python call — write_file("/p", """…"""),
  sometimes wrapped in print(...). The harness only understood JSON/<tools>,
  so it treated the call as prose and (eventually) published an empty artifact.
  The recovery parser now handles the Python-call syntax: it unwraps print(),
  splits arguments while skipping triple/single/double-quoted strings, and maps
  keyword and positional args through each tool's signature. Verified: Gemma's
  write_file now executes on the first attempt.

Fixed

• Gemma multi-turn no longer crashes on the chat template. Gemma's llama.cpp
  template rejects the system/tool roles and requires strictly alternating
  user/assistant turns, so any turn carrying a tool result 400'd and the agentic
  loop died after one step. On such a template error the OpenAI-compatible path
  now retries once with a flattened, strictly-alternating message body. (qwen's
  template accepts the standard roles, so its path is unchanged.)

What's Changed

• release: 0.11.4 — Gemma local-model support (tool_code parsing + template-role flatten) by @geaglin in #46

Full Changelog: v0.11.3...v0.11.4

Gemma local-model support, from a real gemma-2-2b-it session on llama.cpp.

Added

• Parse Gemma's tool_code tool-call format. Gemma emits tool calls as a
  ```tool_code fenced Python call — write_file("/p", """…"""),
  sometimes wrapped in print(...). The harness only understood JSON/<tools>,
  so it treated the call as prose and (eventually) published an empty artifact.
  The recovery parser now handles the Python-call syntax: it unwraps print(),
  splits arguments while skipping triple/single/double-quoted strings, and maps
  keyword and positional args through each tool's signature. Verified: Gemma's
  write_file now executes on the first attempt.

Fixed

• Gemma multi-turn no longer crashes on the chat template. Gemma's llama.cpp
  template rejects the system/tool roles and requires strictly alternating
  user/assistant turns, so any turn carrying a tool result 400'd and the agentic
  loop died after one step. On such a template error the OpenAI-compatible path
  now retries once with a flattened, strictly-alternating message body. (qwen's
  template accepts the standard roles, so its path is unchanged.)

What's Changed

• release: 0.11.4 — Gemma local-model support (tool_code parsing + template-role flatten) by @geaglin in #46

Full Changelog: v0.11.3...v0.11.4

Gemma local-model support, from a real gemma-2-2b-it session on llama.cpp.

Added

• Parse Gemma's tool_code tool-call format. Gemma emits tool calls as a
  ```tool_code fenced Python call — write_file("/p", """…"""),
  sometimes wrapped in print(...). The harness only understood JSON/<tools>,
  so it treated the call as prose and (eventually) published an empty artifact.
  The recovery parser now handles the Python-call syntax: it unwraps print(),
  splits arguments while skipping triple/single/double-quoted strings, and maps
  keyword and positional args through each tool's signature. Verified: Gemma's
  write_file now executes on the first attempt.

Fixed

• Gemma multi-turn no longer crashes on the chat template. Gemma's llama.cpp
  template rejects the system/tool roles and requires strictly alternating
  user/assistant turns, so any turn carrying a tool result 400'd and the agentic
  loop died after one step. On such a template error the OpenAI-compatible path
  now retries once with a flattened, strictly-alternating message body. (qwen's
  template accepts the standard roles, so its path is unchanged.)

What's Changed

• release: 0.11.4 — Gemma local-model support (tool_code parsing + template-role flatten) by @geaglin in #46

Full Changelog: v0.11.3...v0.11.4

Gemma local-model support, from a real gemma-2-2b-it session on llama.cpp.

Added

• Parse Gemma's tool_code tool-call format. Gemma emits tool calls as a
  ```tool_code fenced Python call — write_file("/p", """…"""),
  sometimes wrapped in print(...). The harness only understood JSON/<tools>,
  so it treated the call as prose and (eventually) published an empty artifact.
  The recovery parser now handles the Python-call syntax: it unwraps print(),
  splits arguments while skipping triple/single/double-quoted strings, and maps
  keyword and positional args through each tool's signature. Verified: Gemma's
  write_file now executes on the first attempt.

Fixed

• Gemma multi-turn no longer crashes on the chat template. Gemma's llama.cpp
  template rejects the system/tool roles and requires strictly alternating
  user/assistant turns, so any turn carrying a tool result 400'd and the agentic
  loop died after one step. On such a template error the OpenAI-compatible path
  now retries once with a flattened, strictly-alternating message body. (qwen's
  template accepts the standard roles, so its path is unchanged.)

What's Changed

• release: 0.11.4 — Gemma local-model support (tool_code parsing + template-role flatten) by @geaglin in #46

Full Changelog: v0.11.3...v0.11.4

v0.11.3

Small-model BUILD reliability, from a real qwen2.5-coder-1.5b session that
looped instead of building a page.

Fixed

• Small models no longer stall on the clarifying-question tool. The
  question tool is dropped from the compact (local/small-model) tool set — a
  1.5B model was re-asking "use a framework?" endlessly instead of building. It
  now acts on sensible defaults; larger models and PLAN mode keep the tool.
• Question answers now echo what you type. The answer prompt was a
  multi-line string, which mis-positioned the alt-screen composer cursor and
  hid typed input. The question prints on its own line and the answer is read
  with a single-line prompt.
• HTML artifacts that link a local stylesheet are rejected with an
  actionable message (name the file, inline the CSS), mirroring the existing
  local-<script src> check; the too-small message now explicitly demands a
  single self-contained file with no external links.

What's Changed

• release: 0.11.3 — small-model BUILD reliability (question stalling, answer echo, external CSS) by @geaglin in #45

Full Changelog: v0.11.2...v0.11.3

Small-model BUILD reliability, from a real qwen2.5-coder-1.5b session that
looped instead of building a page.

Fixed

• Small models no longer stall on the clarifying-question tool. The
  question tool is dropped from the compact (local/small-model) tool set — a
  1.5B model was re-asking "use a framework?" endlessly instead of building. It
  now acts on sensible defaults; larger models and PLAN mode keep the tool.
• Question answers now echo what you type. The answer prompt was a
  multi-line string, which mis-positioned the alt-screen composer cursor and
  hid typed input. The question prints on its own line and the answer is read
  with a single-line prompt.
• HTML artifacts that link a local stylesheet are rejected with an
  actionable message (name the file, inline the CSS), mirroring the existing
  local-<script src> check; the too-small message now explicitly demands a
  single self-contained file with no external links.

What's Changed

• release: 0.11.3 — small-model BUILD reliability (question stalling, answer echo, external CSS) by @geaglin in #45

Full Changelog: v0.11.2...v0.11.3

Small-model BUILD reliability, from a real qwen2.5-coder-1.5b session that
looped instead of building a page.

Fixed

• Small models no longer stall on the clarifying-question tool. The
  question tool is dropped from the compact (local/small-model) tool set — a
  1.5B model was re-asking "use a framework?" endlessly instead of building. It
  now acts on sensible defaults; larger models and PLAN mode keep the tool.
• Question answers now echo what you type. The answer prompt was a
  multi-line string, which mis-positioned the alt-screen composer cursor and
  hid typed input. The question prints on its own line and the answer is read
  with a single-line prompt.
• HTML artifacts that link a local stylesheet are rejected with an
  actionable message (name the file, inline the CSS), mirroring the existing
  local-<script src> check; the too-small message now explicitly demands a
  single self-contained file with no external links.

What's Changed

• release: 0.11.3 — small-model BUILD reliability (question stalling, answer echo, external CSS) by @geaglin in #45

Full Changelog: v0.11.2...v0.11.3

Small-model BUILD reliability, from a real qwen2.5-coder-1.5b session that
looped instead of building a page.

Fixed

• Small models no longer stall on the clarifying-question tool. The
  question tool is dropped from the compact (local/small-model) tool set — a
  1.5B model was re-asking "use a framework?" endlessly instead of building. It
  now acts on sensible defaults; larger models and PLAN mode keep the tool.
• Question answers now echo what you type. The answer prompt was a
  multi-line string, which mis-positioned the alt-screen composer cursor and
  hid typed input. The question prints on its own line and the answer is read
  with a single-line prompt.
• HTML artifacts that link a local stylesheet are rejected with an
  actionable message (name the file, inline the CSS), mirroring the existing
  local-<script src> check; the too-small message now explicitly demands a
  single self-contained file with no external links.

What's Changed

• release: 0.11.3 — small-model BUILD reliability (question stalling, answer echo, external CSS) by @geaglin in #45

Full Changelog: v0.11.2...v0.11.3

v0.11.2

Document generation, web-content quality, TUI, and small-model streaming
refinements — all additive fixes, no breaking changes.

Added

• Rich Word document generation. create_docx now renders inline markdown
  (**bold**, *italic*, `code`) as real Word runs instead of literal
  asterisks, and converts markdown tables (| col | col |) into bordered Word
  tables with a bold header row. Emphasis is conservative — arithmetic like
  5 * 3, glob patterns, and snake_case are left untouched.
• Structured web search. web_search returns numbered title / url / snippet results (recovering the real target URL from DuckDuckGo's redirect
  wrapper) instead of the raw page run through an HTML stripper — far easier for
  small models to read. Falls back to the stripped text if the markup changes.

Fixed

• Numeric HTML entities in fetched content. strip_html and the search
  parser now decode decimal/hex character references (’, ’,
  —), so fetched pages and snippets no longer show garbled curly quotes
  and dashes.
• TUI markdown emphasis. An unbalanced `, **, or single * (e.g.
  5 * 3) no longer styles the rest of the line — the inline renderer only
  opens a style when a matching closer exists ahead and the marker flanks
  non-space text.
• <think> reasoning leakage. Reasoning models (DeepSeek-R1 distills, Qwen
  thinking variants) that emit <think>…</think> inline in content no longer
  leak that into the final answer or written files: it's stripped on the
  non-streaming parse path, and a <think>/</think> tag split across streaming
  chunks (routine with token-by-token local streaming) is now reassembled
  instead of leaking a partial marker.

What's Changed

• release: 0.11.2 — doc generation, web quality, TUI, small-model streaming by @geaglin in #44

Full Changelog: v0.11.1...v0.11.2

Document generation, web-content quality, TUI, and small-model streaming
refinements — all additive fixes, no breaking changes.

Added

• Rich Word document generation. create_docx now renders inline markdown
  (**bold**, *italic*, `code`) as real Word runs instead of literal
  asterisks, and converts markdown tables (| col | col |) into bordered Word
  tables with a bold header row. Emphasis is conservative — arithmetic like
  5 * 3, glob patterns, and snake_case are left untouched.
• Structured web search. web_search returns numbered title / url / snippet results (recovering the real target URL from DuckDuckGo's redirect
  wrapper) instead of the raw page run through an HTML stripper — far easier for
  small models to read. Falls back to the stripped text if the markup changes.

Fixed

• Numeric HTML entities in fetched content. strip_html and the search
  parser now decode decimal/hex character references (’, ’,
  —), so fetched pages and snippets no longer show garbled curly quotes
  and dashes.
• TUI markdown emphasis. An unbalanced `, **, or single * (e.g.
  5 * 3) no longer styles the rest of the line — the inline renderer only
  opens a style when a matching closer exists ahead and the marker flanks
  non-space text.
• <think> reasoning leakage. Reasoning models (DeepSeek-R1 distills, Qwen
  thinking variants) that emit <think>…</think> inline in content no longer
  leak that into the final answer or written files: it's stripped on the
  non-streaming parse path, and a <think>/</think> tag split across streaming
  chunks (routine with token-by-token local streaming) is now reassembled
  instead of leaking a partial marker.

What's Changed

• release: 0.11.2 — doc generation, web quality, TUI, small-model streaming by @geaglin in #44

Full Changelog: v0.11.1...v0.11.2

Document generation, web-content quality, TUI, and small-model streaming
refinements — all additive fixes, no breaking changes.

Added

• Rich Word document generation. create_docx now renders inline markdown
  (**bold**, *italic*, `code`) as real Word runs instead of literal
  asterisks, and converts markdown tables (| col | col |) into bordered Word
  tables with a bold header row. Emphasis is conservative — arithmetic like
  5 * 3, glob patterns, and snake_case are left untouched.
• Structured web search. web_search returns numbered title / url / snippet results (recovering the real target URL from DuckDuckGo's redirect
  wrapper) instead of the raw page run through an HTML stripper — far easier for
  small models to read. Falls back to the stripped text if the markup changes.

Fixed

• Numeric HTML entities in fetched content. strip_html and the search
  parser now decode decimal/hex character references (’, ’,
  —), so fetched pages and snippets no longer show garbled curly quotes
  and dashes.
• TUI markdown emphasis. An unbalanced `, **, or single * (e.g.
  5 * 3) no longer styles the rest of the line — the inline renderer only
  opens a style when a matching closer exists ahead and the marker flanks
  non-space text.
• <think> reasoning leakage. Reasoning models (DeepSeek-R1 distills, Qwen
  thinking variants) that emit <think>…</think> inline in content no longer
  leak that into the final answer or written files: it's stripped on the
  non-streaming parse path, and a <think>/</think> tag split across streaming
  chunks (routine with token-by-token local streaming) is now reassembled
  instead of leaking a partial marker.

What's Changed

• release: 0.11.2 — doc generation, web quality, TUI, small-model streaming by @geaglin in #44

Full Changelog: v0.11.1...v0.11.2

v0.11.1

Fixed

• Qwen2.5-Coder tool calls now work on llama.cpp/Ollama. Small local coder
  models emit tool calls as <tools>{…}</tools> / <tool_call>{…}</tool_call>
  text in the message content rather than as native tool_calls. The text
  recovery parser only understood bare or ```json-fenced JSON, so these calls
  were treated as prose and the model never acted (an end-to-end run against
  qwen2.5-coder-1.5b produced no file). The parser now strips the XML tool tags
  and extracts a string-aware balanced JSON object, so a content field full of
  CSS `{ }` braces no longer truncates the parse. Also recovers a bare JSON
  object embedded after a leading sentence.

What's Changed

• fix(agent): parse Qwen-Coder tool calls from small local models (0.11.1) by @geaglin in #43

Full Changelog: v0.11.0...v0.11.1

Fixed

• Qwen2.5-Coder tool calls now work on llama.cpp/Ollama. Small local coder
  models emit tool calls as <tools>{…}</tools> / <tool_call>{…}</tool_call>
  text in the message content rather than as native tool_calls. The text
  recovery parser only understood bare or ```json-fenced JSON, so these calls
  were treated as prose and the model never acted (an end-to-end run against
  qwen2.5-coder-1.5b produced no file). The parser now strips the XML tool tags
  and extracts a string-aware balanced JSON object, so a content field full of
  CSS `{ }` braces no longer truncates the parse. Also recovers a bare JSON
  object embedded after a leading sentence.

What's Changed

• fix(agent): parse Qwen-Coder tool calls from small local models (0.11.1) by @geaglin in #43

Full Changelog: v0.11.0...v0.11.1

Fixed

• Qwen2.5-Coder tool calls now work on llama.cpp/Ollama. Small local coder
  models emit tool calls as <tools>{…}</tools> / <tool_call>{…}</tool_call>
  text in the message content rather than as native tool_calls. The text
  recovery parser only understood bare or ```json-fenced JSON, so these calls
  were treated as prose and the model never acted (an end-to-end run against
  qwen2.5-coder-1.5b produced no file). The parser now strips the XML tool tags
  and extracts a string-aware balanced JSON object, so a content field full of
  CSS `{ }` braces no longer truncates the parse. Also recovers a bare JSON
  object embedded after a leading sentence.

What's Changed

• fix(agent): parse Qwen-Coder tool calls from small local models (0.11.1) by @geaglin in #43

Full Changelog: v0.11.0...v0.11.1

Fixed

• Qwen2.5-Coder tool calls now work on llama.cpp/Ollama. Small local coder
  models emit tool calls as <tools>{…}</tools> / <tool_call>{…}</tool_call>
  text in the message content rather than as native tool_calls. The text
  recovery parser only understood bare or ```json-fenced JSON, so these calls
  were treated as prose and the model never acted (an end-to-end run against
  qwen2.5-coder-1.5b produced no file). The parser now strips the XML tool tags
  and extracts a string-aware balanced JSON object, so a content field full of
  CSS `{ }` braces no longer truncates the parse. Also recovers a bare JSON
  object embedded after a leading sentence.

What's Changed

• fix(agent): parse Qwen-Coder tool calls from small local models (0.11.1) by @geaglin in #43

Full Changelog: v0.11.0...v0.11.1

v0.11.0

Reliability, small-open-weight-model support, and TUI UX overhaul. Grounded in a
full-codebase review plus field research on what opencode / vtcode / aider /
cline ship for weak models.

Added

• Native Ollama protocol (/api/chat): detects the model's real context
  window via /api/show and sets num_ctx, eliminating Ollama's silent
  front-of-prompt truncation — the single biggest measured quality loss for
  local models. Sends repeat_penalty: 1.0 and a low default temperature.
  …/v1 base URLs stay on the OpenAI-compatible path for backward compatibility.
• Live colored diffs rendered in the TUI as files are written and edited
  (+/- hunks with add/remove counts), replacing opaque tool-call lines.
• Markdown rendering on every output path, including streaming — headings,
  bold/italic, inline code, lists, and fenced code blocks now render instead
  of showing raw markdown.
• Auto mode-switching: a build/plan request made in BRAINSTORM escalates to
  the appropriate mode instead of only chatting about it.
• "Act, don't explain" nudge: an imperative task answered with prose-only
  instructions is pushed to actually use its tools.
• Tiered lenient edit apply: whitespace/indentation-tolerant matching
  auto-rescues near-miss edits from weak models; similarity matches remain
  diagnostic-only.
• Configurable temperature, max_tokens, and context_tokens settings.
• CI now runs cargo fmt --check and cargo audit.

Changed

• Reply carries a normalized stop_reason across both wire protocols;
  max_tokens truncation now triggers a bounded continuation instead of
  processing a truncated response.
• Retry policy is status-code-based and covers HTTP 529, honors Retry-After.
• Loop guard counts only repeated errors/no-ops (threshold 3), nudges once, then
  stops honestly — legitimate re-reads no longer end sessions.
• Compaction can no longer sever tool_use/tool_result pairs, always pins the
  original task, and preserves the recent tail; context-overflow errors
  force-compact and retry.
• Verifier is wired to real tool-call and changed-file data and feeds violations
  back to the model.
• System prompt restructured: role/mode contract first, deduplicated,
  game/artifact guidance gated on task type, standing rules dump removed.
• Compact tool set now applies to all local-context models.
• Command output truncation keeps head and tail so failing-test summaries and
  exit codes survive; run_command/python_tool gain a 120s timeout.
• Ask-mode auto-allow list trimmed from 44 commands to 15 read-only ones.
• Rewrote SECURITY.md (it described the removed Python/NEXUS package) and
  corrected README.md drift.

Fixed

• Truncated streamed tool-call JSON no longer executes with empty input — it
  surfaces as an invalid-arguments error fed back to the model.
• Removed the auto-repair hijacks that overwrote correct model output with
  literal task substrings and rewrote shell commands into file writes.
• Removed the hardcoded "fallback canvas game" that shipped unrelated code as
  a successful result; artifact rejections now report the exact reason and retry.
• Artifact validator no longer rejects valid apps over .../todo
  substrings; rejections quote the offending snippet and the exact rule.
• Queued-message TUI deadlock on edit/remove/consume during streaming.
• /undo no longer truncates binary or oversized files.
• Display-width-aware wrapping (emoji/CJK), atomic session saves, hook-execution
  watchdog with distinct failure codes, and edit-mismatch diagnostics that point
  at the closest near-match.
• Read-only command classifier hardened against ;/|/redirection smuggling,
  sed -i, find -delete, and git clean; mutating file tools are confined to
  the working directory.

What's Changed

• build(deps): bump github/codeql-action from 4.36.2 to 4.36.3 in the actions group by @dependabot[bot] in #40
• feat(harness): reliability + small-model + TUI overhaul (0.11.0) by @geaglin in #42

Full Changelog: v0.10.7...v0.11.0

Reliability, small-open-weight-model support, and TUI UX overhaul. Grounded in a
full-codebase review plus field research on what opencode / vtcode / aider /
cline ship for weak models.

Added

• Native Ollama protocol (/api/chat): detects the model's real context
  window via /api/show and sets num_ctx, eliminating Ollama's silent
  front-of-prompt truncation — the single biggest measured quality loss for
  local models. Sends repeat_penalty: 1.0 and a low default temperature.
  …/v1 base URLs stay on the OpenAI-compatible path for backward compatibility.
• Live colored diffs rendered in the TUI as files are written and edited
  (+/- hunks with add/remove counts), replacing opaque tool-call lines.
• Markdown rendering on every output path, including streaming — headings,
  bold/italic, inline code, lists, and fenced code blocks now render instead
  of showing raw markdown.
• Auto mode-switching: a build/plan request made in BRAINSTORM escalates to
  the appropriate mode instead of only chatting about it.
• "Act, don't explain" nudge: an imperative task answered with prose-only
  instructions is pushed to actually use its tools.
• Tiered lenient edit apply: whitespace/indentation-tolerant matching
  auto-rescues near-miss edits from weak models; similarity matches remain
  diagnostic-only.
• Configurable temperature, max_tokens, and context_tokens settings.
• CI now runs cargo fmt --check and cargo audit.

Changed

• Reply carries a normalized stop_reason across both wire protocols;
  max_tokens truncation now triggers a bounded continuation instead of
  processing a truncated response.
• Retry policy is status-code-based and covers HTTP 529, honors Retry-After.
• Loop guard counts only repeated errors/no-ops (threshold 3), nudges once, then
  stops honestly — legitimate re-reads no longer end sessions.
• Compaction can no longer sever tool_use/tool_result pairs, always pins the
  original task, and preserves the recent tail; context-overflow errors
  force-compact and retry.
• Verifier is wired to real tool-call and changed-file data and feeds violations
  back to the model.
• System prompt restructured: role/mode contract first, deduplicated,
  game/artifact guidance gated on task type, standing rules dump removed.
• Compact tool set now applies to all local-context models.
• Command output truncation keeps head and tail so failing-test summaries and
  exit codes survive; run_command/python_tool gain a 120s timeout.
• Ask-mode auto-allow list trimmed from 44 commands to 15 read-only ones.
• Rewrote SECURITY.md (it described the removed Python/NEXUS package) and
  corrected README.md drift.

Fixed

• Truncated streamed tool-call JSON no longer executes with empty input — it
  surfaces as an invalid-arguments error fed back to the model.
• Removed the auto-repair hijacks that overwrote correct model output with
  literal task substrings and rewrote shell commands into file writes.
• Removed the hardcoded "fallback canvas game" that shipped unrelated code as
  a successful result; artifact rejections now report the exact reason and retry.
• Artifact validator no longer rejects valid apps over .../todo
  substrings; rejections quote the offending snippet and the exact rule.
• Queued-message TUI deadlock on edit/remove/consume during streaming.
• /undo no longer truncates binary or oversized files.
• Display-width-aware wrapping (emoji/CJK), atomic session saves, hook-execution
  watchdog with distinct failure codes, and edit-mismatch diagnostics that point
  at the closest near-match.
• Read-only command classifier hardened against ;/|/redirection smuggling,
  sed -i, find -delete, and git clean; mutating file tools are confined to
  the working directory.

What's Changed

• build(deps): bump github/codeql-action from 4.36.2 to 4.36.3 in the actions group by @dependabot[bot] in #40
• feat(harness): reliability + small-model + TUI overhaul (0.11.0) by @geaglin in #42

Full Changelog: v0.10.7...v0.11.0

Reliability, small-open-weight-model support, and TUI UX overhaul. Grounded in a
full-codebase review plus field research on what opencode / vtcode / aider /
cline ship for weak models.

Added

• Native Ollama protocol (/api/chat): detects the model's real context
  window via /api/show and sets num_ctx, eliminating Ollama's silent
  front-of-prompt truncation — the single biggest measured quality loss for
  local models. Sends repeat_penalty: 1.0 and a low default temperature.
  …/v1 base URLs stay on the OpenAI-compatible path for backward compatibility.
• Live colored diffs rendered in the TUI as files are written and edited
  (+/- hunks with add/remove counts), replacing opaque tool-call lines.
• Markdown rendering on every output path, including streaming — headings,
  bold/italic, inline code, lists, and fenced code blocks now render instead
  of showing raw markdown.
• Auto mode-switching: a build/plan request made in BRAINSTORM escalates to
  the appropriate mode instead of only chatting about it.
• "Act, don't explain" nudge: an imperative task answered with prose-only
  instructions is pushed to actually use its tools.
• Tiered lenient edit apply: whitespace/indentation-tolerant matching
  auto-rescues near-miss edits from weak m...

v0.10.7

chore(release): bump version to 0.10.7

v0.10.6

fix(npm): resolve workspace target path in postinstall and binary res…