Skip to content

Releases

Yamori 0.20.0 — project guides and test walkthroughs

Choose a tag to compare

@cowana-ai cowana-ai released this 16 Nov 12:20

Yamori 0.20.0 adds a project guide for reading a codebase in context.

  • Local project maps with folder navigation, reading order, definitions, imports and references, using bundled tree-sitter grammars.
  • AI project notes and glossary, plus read-only project lookups for Codex and Claude Code when explaining mapped files.
  • Cross-file explanation links, editor hovers and CodeLens, reading tours and optional one-file-ahead preparation.
  • Python unit-test walkthroughs showing actual inputs, executed steps, outputs and exceptions. Real runs require a trusted folder and explicit user approval; AI simulations are labeled separately.
  • Fast-model settings for quick jobs, compact step links, clearer controls and visible generation stages.

Validation: all 97 automated tests and syntax checks pass. Generated Ajv validators match their schemas. An extracted VSIX loads its engine, MCP server and WASM parser without development dependencies. The package includes all grammar assets and the Python runner.

Yamori 0.19.4 — compiled schema validation

Choose a tag to compare

@cowana-ai cowana-ai released this 15 Nov 20:59

Ajv now validates generated explanations, difficulty selections and evaluator feedback using shared JSON Schemas. Invalid fields report precise paths; source-range and candidate-block reference checks remain contextual.

Validators are generated at build time, bundled locally and add about 19 KB. No runtime Ajv installation or schema compilation is needed. Existing cached explanations remain readable. Invalid reviewer output keeps the existing bounded repair and explicitly unreviewed draft fallback.

Validation: all 61 automated tests pass, including strict types without coercion/defaults, field paths, cache compatibility and isolated loading without Ajv or runtime compilation. Syntax checks and an extracted VSIX loading check pass.

Yamori 0.19.3 — Recover from evaluator feedback errors

Choose a tag to compare

@cowana-ai cowana-ai released this 15 Nov 20:46

Fix evaluator-reference failures without losing the generated explanation.

  • Restrict evaluator feedback to the views and candidate-local block indices actually being reviewed. Clearly distinguish local indices from source line numbers and original document indices.
  • Accept a current-detail view alias only when its mapping to the main view is unambiguous. Never guess block numbering, and preserve critical fidelity objections.
  • Retry only malformed evaluator feedback once with the same candidate and source. The other valid review is not repeated.
  • If review still cannot finish, retain and cache the validated generated draft with an explicit Review incomplete notice. It is not labeled as having passed both reviewers.
  • Retry review reuses the cached draft for unchanged code rather than generating it again.
  • Keep known fidelity failures as hard failures, including a known failed review accompanied by a malformed sibling response. Cancellation remains cancellation.
  • Remove blank activity entries and replace the first-use welcome screen on errors with an actionable retry control.

Validation: 58 tests passed, including forced malformed-feedback repair/failure, cache round trips, resuming without a generation request, cancellation and fidelity gating. A live Codex run accepted the stricter schema and completed review. An isolated native VS Code run reproduced persistently malformed review output, retained both explanation blocks, and reopened the draft from cache with no extra inference.

Reload VS Code once after updating to 0.19.3, then open the file again. A previous failed generation did not save a complete result; the first retry may still need a new generation.

Yamori 0.19.2 — Read while reviews run

Choose a tag to compare

@cowana-ai cowana-ai released this 14 Nov 16:49

Read the first draft while Yamori reviews the difficult blocks.

  • Restore immediate streaming even when refinement is enabled. A complete draft stays readable with a logic map while reviews run; final source/edit actions activate after validation.
  • Stop early when both evaluators report no actionable changes with low rubric scores, or when a revision fails to improve an already faithful candidate. N remains a maximum.
  • Use compact source/task context for selection and review instead of repeatedly sending the full writing guide and its illustrative examples. Preserve actual source, line numbers, supporting definitions, explicit user proposals and the uncertainty/reading-cost objective.
  • Keep final caching, semantic-fidelity gating, selected-block scope, cancellation and Restricted Mode support.

Validation: 55 tests passed. The structured provider fixture used 7 requests instead of 10 after a useful first revision; a satisfied initial candidate needs 4. In one small live Codex run, text began streaming at 4 seconds, the complete draft was ready at 26 seconds, and review finished at 44 seconds with no writer pass needed. These timings are one example, not a general speed guarantee.

For one-request generation, choose AI menu → Refine → Off · Fast draft. Reload VS Code once after updating.

Yamori 0.19.1 — Ready from the first install

Choose a tag to compare

@cowana-ai cowana-ai released this 14 Nov 16:02

Yamori is now available from the first install across local script files, including TypeScript and Restricted Mode.

  • Keep the gecko editor action and Yamori status-bar button visible in untrusted folders. No workspace-trust step is needed to explain source.
  • Use a theme-aware gecko toolbar glyph and a one-time Open current file welcome action.
  • Automatically choose an installed Codex or Claude CLI on first open if the user has not already chosen a provider.
  • Keep Restricted Mode enabled. Ignore workspace-controlled CLI paths and usage settings; skip project language-service definition lookups.
  • Use isolated, tool-disabled provider sessions outside the project, and reject relative executable paths. Authentication and installation still use the local provider CLI.

Validation: 52 tests passed. A fresh native VS Code profile activated Yamori and explained a TypeScript file while isTrusted remained false, with a mock provider. Untrusted project CLI and usage overrides were ignored. No workspace trust settings were changed.

Existing users: restart or reload VS Code once after updating. Fresh installations activate automatically.

Yamori 0.19.0 — Refine the difficult parts

Choose a tag to compare

@cowana-ai cowana-ai released this 14 Nov 15:00

Yamori now refines difficult explanations before showing the final result.

  • A selector picks up to six blocks with non-obvious reasoning, subtle behavior or missing conceptual bridges. Straightforward functions remain untouched; the selector may skip refinement entirely.
  • Two independent evaluator sessions run in parallel: one checks remaining uncertainty, the other reading cost. Each also checks semantic fidelity.
  • A writer revises only selected blocks for up to N steps. Each candidate is reviewed again; Yamori keeps the lowest rubric objective that passes both fidelity checks and preserves easy sections exactly.
  • Choose 0–5 steps in AI menu → Refine or Yamori: Choose Refinement Steps. Default: 2. All sessions use the selected provider/model/effort with tools disabled.
  • The working score is uncertainty + 0.5 × reading cost. These are LLM rubric judgments, not measured entropy or a proof of correctness.
  • New explanations use 2 model requests when no hard blocks are selected, or up to 3N+4 with refinement. Setting 0 uses one request. Cache reads never rerun evaluation.
  • Completed text and review metrics stay cached locally. Previous caches remain readable. After an accepted revision, other detail levels are derived from the final explanation when requested.
  • Stop cancels both reviewers. Failed later passes retain an already reviewed candidate with a notice; unresolved fidelity errors withhold the new result.

Validation: 49 automated tests passed, including mocked Codex/Claude pipelines, concurrent review, fidelity gating, exact preservation of easy blocks, cancellation, invalid feedback and cache policy separation. A live Codex smoke test selected an agreement-count identity while leaving a simple seed getter alone, completed one refinement step and retained the improved candidate. Live Claude refinement has not been verified.

After updating, run Developer: Reload Window once. Use Refresh all to apply refinement to an older cached explanation.

Yamori 0.18.0 — Local cache and source-aware refresh

Choose a tag to compare

@cowana-ai cowana-ai released this 14 Nov 13:23

Yamori now keeps explanations and editing drafts on your device and helps you update them when source code changes.

  • Persist completed detail views, approved revisions, source snapshots, dates and hashes locally, with no automatic 100-entry eviction.
  • Autosave block drafts, instructions, cursor position and undo history across reloads.
  • Detect editor, Git and external filesystem changes; show dated warnings and affected blocks without automatically spending model tokens.
  • Refresh one block or the whole cached explanation in context. Patch-only refreshes preserve untouched blocks exactly; new/shared/removed code can rebuild the logic map.
  • Rebase source links locally and disable links only for stale blocks. Cancel late results when source changes again.
  • Surface disk-save failures and retain the previous cached version if a revision cannot be saved.

Validation: 42 offline tests passed, plus a native VS Code host check for file watching, persisted freshness metadata, independent file tabs and cache reuse. Browser fixture checks covered warning controls and draft restoration; those UI checks made no provider inference requests.

After updating, run Developer: Reload Window once. Existing matching legacy caches are indexed when opened; historical unindexed caches cannot identify an unknown earlier source version.

Yamori 0.17.0

Yamori 0.17.0 Pre-release
Pre-release

Choose a tag to compare

@cowana-ai cowana-ai released this 14 Nov 11:05

Yamori now starts with a compact logic map instead of a wall of blocks. Choose an idea, read one block, and use Previous/Next or return to the map. Navigation starts no model request.

Editing: CodeMirror adds line numbers, Tab/Shift+Tab indentation, continued indentation on Enter, find, undo/redo, readable emphasis and clickable typeset equations. Drafts, selections and history survive map navigation and theme/status updates. Ordinary deletion does not expose hidden formatting markers. Use Cmd/Ctrl+Enter to update the selected explanation with your provider.

Interface: a minimal monochrome pixel style, a tiny pixel gecko samurai, squared controls, restrained loading/navigation motion, compact model settings, light/dark themes and reduced-motion support.

Validation: 35 offline tests and syntax checks pass. Browser checks cover 360/480 px panels, normal typing, indentation, undo/redo, formatting-aware deletion, equation editing, draft/history retention and failed-update recovery. A native VS Code extension-host check passed for startup activation, two independent file tabs and cache reuse. UI checks use fixture data and do not run provider inference.

Reload the editor window after updating. Existing caches and palette preferences are retained. Source files are unchanged by explanation edits.

Yamori 0.16.1

Yamori 0.16.1 Pre-release
Pre-release

Choose a tag to compare

@cowana-ai cowana-ai released this 14 Nov 10:40

Yamori’s one-click controls now use its own identity: the gecko samurai logo in the script toolbar and a compact gecko mark beside Yamori in the status bar. Click either to open the active script’s explanation. Browsing files starts no AI request; existing per-file tabs and cached views are reused.

The custom status icon inherits the editor text color and adds a single 780-byte font asset. The toolbar reuses the existing mascot artwork.

Validation: 34 offline tests and syntax checks pass; the font was rendered at 16 px and 24 px on light and dark surfaces. Native VS Code and Cursor extension-host checks for the underlying one-click workflow passed, including startup activation, two independent tabs, and reopening without another request.

Reload the editor window after updating.

Yamori 0.16.0

Yamori 0.16.0 Pre-release
Pre-release

Choose a tag to compare

@cowana-ai cowana-ai released this 14 Nov 10:32

Open a script, then click the book icon at the top right or Yamori in the bottom status bar. Both open the active file in one click; existing explanation tabs and cached views are reused.

The bottom button appears automatically after editor startup, names its target file on hover, and follows script changes. Browsing files starts no AI request. The keyboard shortcut and Command Palette remain available.

Validation: 34 offline tests and syntax checks pass. A real Cursor extension-host check verifies startup activation, two independent explanation tabs in one reader group, and reopening without another request. No new project files were added.