CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[0.1.95] - 2026-03-18

Changed

• Bumped tiny-agent-os from 1.2.9 to 1.2.11.
• Added the v0.1.94 release run artifact to capture the full publication workflow.

Fixed

• Skipped the live alchemy integration test when the binding is unavailable.
• Applied small isinstance(..., X | Y) cleanups required by the current Ruff rules.

[0.1.94] - 2026-03-17

Changed

• Enforced strict config source validation and updated quality harness documentation.
• Added git safety practices documentation and updated agent guidance.
• Removed stale worktree metadata and unused state machine infrastructure.
• Cleaned up defensive logic and tightened TUI cold-start paths.

Fixed

• Fixed remaining typing issues and tool argument validation.
• Corrected tinyagent imports and split agent configuration for better modularity.
• Fixed CodeRabbit agent feedback issues.
• Stabilized headless tinyagent serialization.

[0.1.93] - 2026-03-16

Changed

• Reduced TUI cold-start overhead by tightening command-registry and package import loading paths.
• Cleaned up stale project artifacts by removing outdated .claude/ and .tickets/ repository files.

[0.1.92] - 2026-03-14

Changed

• Added clipboard copy support for selected UI text and extracted agent-text streaming into a dedicated UI module.

Fixed

• Relaxed clipboard copy verification so successful copies no longer fail when verification is unavailable.
• Stabilized tmux system tool coverage by making discover stats and skill-load confirmations explicit and by deflaking tmux wait logic around persistent evidence.

[0.1.91] - 2026-03-13

Changed

• Hardcoded AGENTS.md as the guide file path and removed guide-file configurability.
• Cleaned up obsolete docs and removed the unused memory-bank/ directory.

Fixed

• Retried /update installs against the active Python interpreter when uv tool upgrade cannot locate the current install.
• Added regression coverage for the uv tool update fallback flow.

[0.1.90] - 2026-03-09

Changed

• Smoothed the model picker flow and hardened model selection state handling in the TUI.

Fixed

• Stabilized tmux system startup by waiting for the editor prompt before sending end-to-end tool commands.
• Added explicit coverage for the loaded-skills panel title so the skills UI contract is tested outside tmux.

[0.1.89] - 2026-03-07

Changed

• Removed the bottom status bar from the TUI and flattened the editor border to match the panel design.

Fixed

• Removed tmux system test dependence on the status bar so the end-to-end tool coverage stays stable after the UI cleanup.

[0.1.88] - 2026-03-06

Added

• Added tmux system coverage that loads a local skill and verifies the agent uses its referenced file end-to-end.

Changed

• Consolidated skill lookup, loading, and selected-skill summary resolution onto a single registry-backed path.

Fixed

• Preserved missing selected skills in UI summary surfaces while keeping prompt-building fail-loud for unresolved skill loads.
• Kept skill catalog and summary views limited to skill names and descriptions while direct selected-skill loads still inject full SKILL.md content plus absolute file paths.

[0.1.87] - 2026-03-06

Changed

• Increased the default shell tool timeout to two minutes to reduce premature command termination in normal workflows.

Fixed

• Removed a flaky headless CLI system test from the release train.
• Stabilized tmux discover-tool system coverage by waiting for rendered scan stats before asserting output.

[0.1.86] - 2026-03-06

Added

• Introduced prompt context injection for actively loaded skills so selected skill guidance is surfaced to the agent runtime.

Changed

• Refined skills UI styling and load behavior to improve local/global skill discovery ergonomics.

Fixed

• Resolved selected-skill path handling to use discovered absolute paths reliably across local and global skill locations.
• Included absolute skill paths in prompt context to prevent path-resolution drift when rendering skill metadata.

[0.1.85] - 2026-03-06

Fixed

• Made /skills autocomplete prefer the most relevant skill match instead of fuzzy-reordering short prefixes.
• Isolated slash-command autocomplete to command names so /skills navigation and Enter selection no longer get hijacked by hidden command suggestions.

[0.1.84] - 2026-03-06

Fixed

• Improved skills loading UX and restored slash-command submit behavior.

[0.1.83] - 2026-03-06

Added

• Introduced a first-class skills subsystem with discovery, loading, session persistence, prompt rendering, and cache-backed registry support.
• Added /skills catalog, search, clear, and loaded-skill management commands in the TUI.

Changed

• Exposed selected skills in the session inspector and included available/selected skill context in agent prompt construction.

Fixed

• Restored compatibility with legacy markdown-only SKILL.md files that do not start with YAML frontmatter.
• Fixed local-over-global precedence for case-insensitive skill name collisions.
• Prevented invalid discovered skill summaries from crashing agent creation for unrelated sessions.

[0.1.82] - 2026-03-05

Changed

• Bumped the packaged TunaCode version to 0.1.82 to cut a new PyPI release.

[0.1.81] - 2026-03-04

Changed

• Migrated the live alchemy usage-contract integration test to the typed chutes flow for stricter runtime contract coverage.
• Refined tinyagent typing and compaction boundaries to tighten type-safety and reduce ambiguity in session compaction behavior.
• Removed residual __all__ shim exports to satisfy Gate 0 and enforce direct exports only.
• Refreshed README screenshots to match the current UI.

Fixed

• Declared pydantic as a direct project dependency to satisfy dependency analysis and prevent missing-direct-dependency failures.

[0.1.80] - 2026-03-03

Fixed

• Aligned the bash tool timeout contract to seconds across prompt docs and runtime validation to prevent repeated millisecond-based tool failures.
• Hardened bash timeout bounds handling by rejecting timeout=0 and added regression tests for timeout validation and prompt contract drift.

[0.1.79] - 2026-03-03

Added

• Added a full prompt-versioning pipeline with computed hashes, mtime-aware caching, and agent-level observability hooks.

Fixed

• Fixed original_query being reset on every request instead of preserved across multi-turn sessions.
• Hid internal tool validation errors from user-facing output.
• Fixed typing issues in the thinking panel widget and constants module.

[0.1.78] - 2026-02-26

Changed

• Removed dead NeXTSTEP UI tool renderer modules for retired tools (glob, grep, list_dir) and cleaned renderer exports/tests to match the tinyagent-only tool surface.

[0.1.77] - 2026-02-25

Changed

• Tightened the session inspector presentation and removed a redundant compaction field to reduce TUI chrome noise.

[0.1.75] - 2026-02-23

Added

• Tool panel CSS flow with status-based classes (running/completed/failed)
• Compaction awareness indicator in context panel
• CSS tint styling for file tool states (read/update)
• Parallel tool lifecycle coverage: capped-concurrency execution test, RequestOrchestrator parallel batch tests, and status-bar callback sequencing tests

Changed

• Enhanced /compact command with error handling and user feedback
• Updated read_file renderer to support new hashline format (1:ab|content)
• Improved tool panel rendering with semantic CSS classes
• Bumped tiny-agent-os dependency to >=1.2.5 and refreshed lockfile resolution
• Enforced max 3 in-flight tool executions via shared tool semaphore wrapping
• Hardened RequestOrchestrator tool-start arg normalization and made batch-mode duration reporting explicit (suppressed for multi-tool batches)
• Updated status bar running-state behavior to stay coherent while multiple tools are active

[0.1.74] - 2026-02-21

Changed

• Rewrite system prompt to describe tools by purpose, not signature
• Clean up discover.py (remove section comments and inline comments)

Fixed

• Normalize @mentions to absolute paths in UI

[0.1.73] - 2026-02-20

Changed

• Extracted thinking state and lifecycle management from app.py into dedicated classes
• System prompt tooling refresh
• Temporarily disabled live text streaming; default thoughts ON
• Removed rich_log alias and enforced full cutover
• Trimmed app.py docstrings to stay under 600-line hook limit

Fixed

• Eliminated blank space between chat content and streaming output

Removed

• Dead code: unused openai_response_validation.py
• Added orphan module detection to CI and removed dead code from agent_helpers

[0.1.72] - 2026-02-20

Fixed

• /update command now works for global installs (pipx, uv tool). Version check uses PyPI JSON API instead of shelling out to pip. Upgrade detects tool-managed venvs and runs the correct upgrade command.

[0.1.71] - 2026-02-20

Added

• discover tool: unified code discovery replacing manual glob/grep/read chains with a single natural-language query

Improved

• Replaced the synthetic discover benchmark with a real-repo harness in tests/benchmarks/bench_discover.py comparing single-call discover against the legacy list_dir -> glob -> grep -> read_file chain
• Benchmark output now reports end-to-end metrics that matter for agent workflows: cold/warm latency (including p50/p95), tool-call count, output token footprint, file/symbol recall, and actionability
• Current baseline run on TunaCode shows discover at 1.0 average tool calls vs 6.0 for legacy (6x fewer round trips), ~4.2x fewer output tokens (~3.1k vs ~12.9k), ~9% lower latency (cold: 2587ms vs 2811ms, warm: 2597ms vs 2832ms), and higher retrieval quality/actionability

Removed

• glob tool (replaced by discover)
• grep tool and grep_components/ module (replaced by discover)
• list_dir tool (replaced by discover + bash)

[0.1.70] - 2026-02-19

Improved

• API key UX: environment variable fallback, inline entry screen, and better error surfacing

[0.1.69] - 2026-02-19

Added

• Inline API key entry screen for model-selection flow when provider credentials are missing
• Unit coverage for API key entry behavior and authentication error rendering

Changed

• Model command now reuses a named provider/model delimiter constant for provider parsing
• MiniMax provider endpoints normalized to /v1 contract paths in integration/unit coverage

Fixed

• API key save flow now raises user-facing errors for filesystem write failures without swallowing unexpected exceptions
• Headless-mode background task handling now shields request task execution and avoids timeout-driven cancellation races

[0.1.67] - 2026-02-19

Added

• First-class MiniMax alchemy routing for coding-plan and default execution paths
• MiniMax provider contract coverage in integration and unit tests

Changed

• Updated provider defaults and models registry to expose MiniMax API-key contract entries
• Pinned tiny-agent-os dependency to >=1.2.1 to support MiniMax provider routing

[0.1.66] - 2026-02-17

Added

• Thinking content streaming for extended thinking models (Claude 3.5+ with thinking enabled)
• /thoughts command to toggle thinking panel visibility
• Thinking panel renderer with truncation and collapsible display
• Tests for thinking stream routing, response extraction, and panel rendering

Changed

• Updated bundled models registry from models.dev (92 providers)

[0.1.65] - 2026-02-16

Added

• Slopgotchi pet widget in context inspector with click interaction and ASCII art cycling
• Context panel with model/token/cost summary and edit tracking
• Codebase structure tree documentation
• /cancel command for request cancellation
• /exit command and refresh UI command docs/tests
• Dead-code CI checks (unimport, vulture) in lint workflow
• Unit test for context panel summary

Changed

• Renamed tamagochi module to slopgotchi
• Side-by-side diff view uses explicit Before/After captions and visible change lane
• Context rail border title set to "Session Inspector" with theme-colored accent
• Reduced core agents main module to 600 lines via helpers extraction
• Updated README interface section and clarified tool execution

Removed

• Dead tools parsing module (src/tunacode/tools/parsing/)

Fixed

• Pre-commit file-length hook performance
• Chat container tuple write issue

[0.1.64] - 2026-02-12

Fixed

• Allow manual compaction below threshold
• Keep streaming updates and remove tint crash path

Changed

• Updated theme architecture handoff and CSS architecture documentation

[0.1.63] - 2026-02-12

Added

• Live integration test for tinyagent alchemy usage contract (tests/integration/core/test_tinyagent_alchemy_usage_contract_live.py)
• Dedicated debug usage/resource lifecycle tracing module (src/tunacode/core/debug/usage_trace.py)
• Unit coverage for strict session usage schema loading and resource bar session cost propagation
• Research map documenting TunaCode Rust-only tinyagent migration path (.claude/metadata/research/2026-02-12-tinyagent-rust-only-migration-map.md)

Changed

• Migrated TunaCode agent runtime stream path to tinyagent Rust alchemy stream (stream_alchemy_openai_completions)
• Migrated compaction summary generation stream path to tinyagent Rust alchemy stream
• Updated usage model to strict canonical tinyagent contract (input/output/cache_read/cache_write/total_tokens/cost)
• Bumped tiny-agent-os to 1.1.5

Fixed

• Enforced fail-loud behavior when assistant usage payload is missing or violates canonical contract
• Resource bar session cost now reads canonical session_total_usage.cost.total

[0.1.62] - 2026-02-11

Added

• Core Agent: Migrated from pydantic-ai to tinyagent as the core agent loop
• Text Selection: Added Rich-renderable mouse selection support in chat panels
• Visual Styling: SelectableRichVisual for text selection in Rich renderables
• CSS-Based Theming: Textual CSS styling system with 5 stylesheet files (panels, theme-nextstep, layout, widgets, modals)
• NeXTSTEP Theme: 3D bevel borders with light top/left and dark bottom/right for raised effect
• Context Management: Context compaction system with overflow retry capabilities
• Token Tracking: OpenRouter token usage tracking for streaming responses
• Models Registry: Expanded models_registry.json with full provider catalog and API URL routing
• Provider Routing: Native TinyAgent OpenRouterModel(base_url) for provider API URL routing

Changed

• BREAKING: Agent session persistence uses dict messages only - existing sessions may not load correctly
• BREAKING: Tool execution is now sequential (was parallel via custom orchestrator)
• Replaced Rich Panel wrappers with CSS-based PanelMeta pattern
• Agent creation now constructs tinyagent.Agent with AgentOptions
• Revamped theme architecture with CSS variable-based theming system
• Updated layout, panels, and widget styles for improved visual consistency
• Bumped tiny-agent-os to v1.1.3 (includes Rust bindings)

Removed

• Orchestrator and tool dispatcher components (tinyagent handles tool execution)
• Streaming components (pydantic-ai specific)
• Tool executor (tinyagent owns tool execution)
• pydantic-ai dependency (replaced with tiny-agent-os)

Fixed

• Restored status bar content row with top bevel

[0.1.61] - 2026-02-06

Added

• End-to-end tests for mtime-driven caches
• Typed cache accessors for agents, context, and ignore manager
• Strict cache infrastructure with strategies

Changed

• Migrated remaining lru_cache caches into CacheManager layer
• Refactored agent_config to use typed cache accessors
• Cache accessor now used for ignore manager; removed global ignore cache
• Reduced McCabe complexity to ≤10 for 14 functions
• Reduced cognitive complexity of 13 functions to under 10
• Re-enabled Ruff mccabe complexity check (max 10)

Fixed

• Satisfied pre-commit dead code checks