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 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