feat(llm,tests): live-Ollama tests + chat UX fixes#65
Merged
Conversation
…bug reports
Adds a new `zig build ollama` step that exercises the LLM module
against a real `OLLAMA_HOST` / `LLM_API_BASE` endpoint instead of
mocking HTTP. Each test starts with a reachability probe and returns
`SkipZigTest` when the endpoint isn't responding — so CI without
Ollama sees "N skipped" instead of a failure.
Three tests cover the protocol shape (not model content):
- single-mode round-trip returns a non-empty command
- dialog-mode round-trip parses as a valid envelope
- destructive-task dialog parses (skips on model drift)
Run with: `OLLAMA_HOST=http://localhost:11434 ATTY_TEST_MODEL=qwen3-coder:latest zig build ollama`
Also fixes four real-session bugs from screenshots:
1. **Exec command echoed into chat input row instead of shell prompt.**
`paintInlineChat` was leaving the real cursor inside the chat
panel's input row, so bash's echo of the injected command landed
there. Now hide the real cursor (`?25l`) + DECSC on open, DECRC
on close — the visible block-cursor glyph stays in the input row
for the user, but bash's echo lands at the SHELL prompt above.
2. **Raw JSON envelope shown in chat scrollback.**
`assistant_exec` turns held the raw `{"action":"exec",...}` for
model context. New `renderTurnContent` helper at paint time parses
the envelope and shows: description + cyan command for `.exec`,
italic question for `.question`, mauve ✓ + reason for `.done`.
Falls back to raw when parse fails (single-mode, drift).
3. **Question/done from LLM never appeared as turns in chat.**
The `.question` branch now also pushes the assistant envelope so
it lands in scrollback alongside the existing hint surface. (`.done`
skipped — `dialogReset` wipes the turn ring; the conclusion banner
scrolls into shell history instead.)
4. **Retry message was generic.**
`requestParseRetry` now detects three specific drift patterns —
trailing `{"open_chat": true}` second-object, markdown fences,
prose preamble — and includes a concrete fix example in the
corrective user turn ("emit ONE object with `open_chat` as a
FIELD, not a separate object").
Plus: leading newlines in the conclusion banner went from `\n\n` to
`\r\n` — fixes the "banner indented" rendering shown in image #7
where the banner inherited the prompt's cursor column instead of
starting at column 1.
`zig build test` (471 pass) + `zig build itest` + `zig build ollama`
(3 pass on local qwen3-coder) + `zig fmt --check` clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
fentas
added a commit
that referenced
this pull request
May 17, 2026
docs: reflect recently-shipped features `docs/llm.md`: - New **Chat surfaces** section explaining Alt+C inline panel vs. Alt+Shift+C full overlay (same conversation ring, mutually exclusive cursor focus). - New **Keybindings** section listing every shipped LLM binding + Alt+H cheat-sheet pointer. - Configuration reference split into sections (Core / Chat surfaces / Persistence / Visual signals / Buffer sizes / Model struct) so the table doesn't overwhelm — and adds rows for `models`, `inline_chat_rows`, `overlay_open_policy`, `chat_persist_enabled`, `chat_persist_path`, `chat_persist_max_bytes`, `history_turns_max`, `max_turn_bytes`, `dialog_system_prompt`, `dialog_parse_retry_max`. - Documents the `Model` struct with the per-model `history_turns_max` trim knob. `docs/modules.md`: - Adds `default_bindings`, `onResize`, `isInlineChatActive`, `extraReserveRows` to the hook list. - New **default_bindings** section walks the module-owned keymap pattern (dispatcher concat + user-overrides-win precedence). - New **extraReserveRows** section documents the inline-panel reservation contract used by the LLM module's Alt+C panel.
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds
zig build ollamafor real-endpoint integration tests (skipped when unreachable) AND fixes four UX issues spotted in real chat sessions after PR #64 merged.Tests
Three tests covering protocol SHAPE (not content):
All three pass locally against
qwen3-coder:latest. With noOLLAMA_HOSTset: all three skip.UX fixes
Exec command echoed into chat input, not shell prompt —
paintInlineChatnow hides the real cursor + DECRC on every paint so bash's echo of an injected command lands at the SHELL prompt above, while the chat panel's block-cursor glyph stays in the input row as the visible marker.Raw JSON envelope shown in chat scrollback — new
renderTurnContentparsesassistant_execturns at paint time and renders:description → cyan commandfor exec, italic question for question, mauve ✓ + reason for done. Falls back to raw on parse failure.Question/done from LLM never appeared as turns in chat —
.questionbranch now pushes the assistant envelope so it lands in scrollback alongside the hint. (.doneskipped:dialogResetwipes the ring; the conclusion banner scrolls into shell history.)Retry was generic ("wasn't valid JSON") —
requestParseRetrynow detects three specific drift patterns (trailing{"open_chat":true}second object, markdown fences, prose preamble) and includes a concrete fix example.Plus: conclusion banner went from
\n\nto\r\n— fixes the "banner indented to prompt column" rendering in image #7.Test plan
zig build test— 471/471 passzig build itest— 6/6 passzig build ollama— 3/3 pass withOLLAMA_HOSTset; 3/3 skip withoutzig fmt --check— cleantest→ Enter → command echo lands at shell prompt (not in panel input)🤖 Generated with Claude Code