dev

.gitignore

@@ -4,8 +4,11 @@ __pycache__/
 *.pyc
 .env*
 .nanocode
+.codegraph/
 .venv/
 .mypy_cache/
 .ruff_cache/
 uv.lock
 .python-version
+.code-symbol-index/
+.code-workflow-probe.json

CHANGELOG.md

@@ -1,15 +1,74 @@
 # Changelog
 
+## 0.4.8 - 2026-05-23
+
+### Changed
+- Renamed the `EditFile` tool to `Edit` across the codebase and tests.
+
+## 0.4.5 - 2026-05-21
+
+### Changed
+- Updated the built-in code index integration for `code-symbol-index` 0.1.7.
+- Added indexed symbol filters for kind, path, and exact matching.
+- Added file-local symbol outlines and bounded pending-index details in `/status`.
+
+## 0.4.4 - 2026-05-20
+
+### Added
+- Added built-in indexed code navigation backed by project data and `/index` for manual init/sync.
+
+### Changed
+- Replaced the external code-navigation CLI integration with the bundled code index API.
+- Hid code navigation tools until an index exists, while lightly updating existing indexes at startup.
+- Updated status/docs to describe code index availability without exposing dependency-install wording.
+
+## 0.4.3 - 2026-05-20
+
+### Changed
+- Removed stable knowledge state while keeping current-task known facts.
+- Extracted shared numbered-content and line-range helpers for tool output/range handling.
+- Trimmed thin helper wrappers in List and indexed code-inspection tools.
+
+## 0.4.2 - 2026-05-19
+
+### Added
+- Added indexed code inspection tools for symbol lookup, symbol investigation, and file outlines when the local index is available.
+- Added queued user feedback during long-running turns.
+- Added `PatchFile` for multi-location file edits.
+
+### Changed
+- Moved model calls to the OpenAI SDK and function-tool protocol.
+- Reworked task-shape prompts for chat, one-shot tasks, and tracked tasks.
+- Prioritized indexed code inspection for structural lookup while keeping Search/Read for exact literals and edit ranges.
+- Improved terminal UX with persistent status, queued-input handling, Bash live preview, and terminal-friendly assistant output rules.
+- Renamed `ListDir` to `List`.
+- Improved `Read`, `Edit`, `ReplaceRange`, `PatchFile`, `Bash`, and `Git` tool guidance.
+- Simplified gate behavior so only deterministic, correctable model errors are refused.
+
+### Fixed
+- Fixed duplicate final replies for goal-only text answers.
+- Fixed repeated recall loops and several format/tool-name compatibility issues.
+- Fixed PatchFile diagnostics and empty-hunk handling.
+- Fixed queued feedback delivery, Ctrl-C/Ctrl-D handling, and Bash interrupt reporting.
+
 ## 0.3.35 - 2026-05-16
 
+### Added
+- Added batched `ReplaceRange` edits for multiple independent ranges in the same file.
+- Added a design document covering agent state, context construction, tool-result storage, observe policy, and verification.
+
 ### Changed
 - Aligned tool-result context layout with the design document.
+- Refined tool-result context reduction around unreduced raw results, retained results, and checkpoint-based pruning.
 - Compressed ACT and OBSERVE system prompts.
 - Reduced routine OBSERVE triggers by raising the pending-result threshold and keeping ordinary tool failures in ACT for repair.
+- Simplified agent gate and feedback handling, including single active plan item normalization.
 - Added soft feedback for state-update-only ACT turns so models continue with frontier tools, verification, or completion.
+- Highlighted recognized slash commands and reported unknown slash commands directly.
 
 ### Fixed
 - Accepted harmless model output variants including trailing progress text, action type casing, and `message` action aliases.
+- Ignored pending verification requests instead of treating them as blocking model output.
 
 ## 0.3.34 - 2026-05-16
 
@@ -115,7 +174,7 @@
 
 ### Changed
 - Tightened completion gates, verification blockers, and compact state update grouping.
-- Simplified Search argument parsing and removed legacy `/knowledge update` behavior.
+- Simplified Search argument parsing and removed legacy knowledge-update behavior.
 - Made provider reasoning payload shape configurable.
 
 ## 0.3.20 - 2026-05-15

README.md

@@ -2,7 +2,7 @@
 
 A lightweight terminal-based AI coding assistant.
 
-nanocode is used to help building itself, including features such as `@file` path completion.
+nanocode is used to help building itself.
 
 Pre-1.0 note: nanocode is still evolving quickly. Functionality, commands, configuration, and behavior may change incompatibly before a 1.0 release.
 
@@ -14,9 +14,9 @@ Pre-1.0 note: nanocode is still evolving quickly. Functionality, commands, confi
 
 ## Features
 
-- **Constrained Output**: Force model replies into auditable action frames.
+- **Function Tools**: Route model decisions through auditable tools.
 - **Verified Edits**: Reject stale range edits before they touch files.
-- **Autonomous Loop**: Chain reading, editing, running, and verification.
+- **Autonomous Loop**: Chain reading, editing, running, and checks.
 - **Live Telemetry**: Stream tool intent, token use, and status.
 
 ## Install
@@ -61,7 +61,6 @@ Ask a source-aware question about nanocode itself:
 CLI arguments:
 
 - `--yolo`: Skip tool execution confirmations.
-- `--plan`: Plan changes without editing files or running commands.
 - `--debug`: Write request prompts to the current session directory under `~/.nanocode/sessions/`.
 - `--config <path>`: Path to config file (default: `~/.nanocode/config.toml`).
 - `--init-config`: Create a default config file.
@@ -75,34 +74,42 @@ If you do not fully trust the model, tools, prompts, or workspace, run nanocode
 
 USE AT YOUR OWN RISK.
 
+nanocode currently targets macOS and Linux. Windows is not supported.
+
 ## Tools
 
-- File: `Read`, `LineCount`, `ListDir`, `Search`.
-- Edit: `Edit`, `ReplaceRange`.
+- File: `Read`, `LineCount`, `List`, `Search`.
+- Code navigation: `InspectCode` after `/index` builds the project index.
+- Edit: `CreateFile`, `EditFile`.
 - Shell: `Bash`, `Git`.
 - Memory: `Recall` reads stored tool results by key.
 
+`Search`, `Read`, and `InspectCode` mode=inspect return 0-based `line:hash|code` lines that can be used as edit anchors. For broad mechanical text replacement, shell text pipelines are acceptable when followed by a focused diff or test.
+
 ## Commands
 
-- Info: `/help [question]`, `/status`, `/rules`, `/knowledge`, `/compact`.
-- Config: `/config`, `/set <key> <value>`, `/model [model_name]`, `/reason`, `/provider [name]`, `/plan [on|off|question]`, `/yolo`.
-- Maintenance: `/clean`.
+- Info: `/help [question]`, `/status`, `/rules`, `/compact`.
+- Config: `/config`, `/set <key> <value>`, `/api [auto|chat|responses]`, `/model [model_name]`, `/reason`, `/reason-payload [value]`, `/provider [name]`, `/yolo`.
+- Maintenance: `/index [force]`, `/clean`.
 - Exit: `/exit`, `/quit`.
 
-Selectors support `j`/`k`, arrows, `/keyword`, Enter, and Esc. `/model` lists configured models before discovered ones, then prompts for reasoning; `/model <name>` and `/reason` are direct shortcuts.
+Selectors support `j`/`k`, arrows, `/keyword`, Enter, and Esc. `/api responses` switches the current provider to Responses format. `/reason` sets `provider.reasoning` to `off` or an effort value; `/reason-payload` controls the Chat-only reasoning payload shape. `/model` lists configured models before discovered ones, then prompts for reasoning.
+During a slow model request, press `Ctrl-G` to cancel that request and resend the same prompt.
 
 ## Configuration
 
 Run `nanocode --init-config` to create `~/.nanocode/config.toml`.
 
-- Provider config: `[provider] active = "<name>"` plus `[provider.<name>]` url, key, model, `available_models`, and model options. `reasoning_payload` controls whether effort is sent as `reasoning`, `reasoning_effort`, or not sent.
+- Provider config: `[provider] active = "<name>"` plus `[provider.<name>]` url, key, model, `available_models`, and model options. `api` selects `chat`, `responses`, or `auto`; auto uses exact-host profiles. Responses uses standard `reasoning.effort`; Chat reasoning is mapped by provider/model profile when known.
+- Provider auto-detection covers common providers: OpenAI/OpenRouter prefer Responses API; DeepSeek, selected OpenCode models, and DashScope models use their matching Chat reasoning payload shapes.
 - Path config: `[paths] data_dir = "~/.nanocode"`.
 - Runtime config: `[runtime]`.
+- `/context [low|medium|high]` shows or switches tool-result context budgets; lower budgets reduce token usage and observe overhead.
 - Session data: debug prompts and tool-result logs are stored under `~/.nanocode/sessions/<session_id>/`.
-- Tool-result logs from inactive sessions are auto-cleaned after `runtime.auto_clean_recent` (default `3d`; use `off` to disable). `/clean` removes inactive session logs immediately.
-- Project data: user rules are stored under `~/.nanocode/projects/<project_key>/`.
+- Old inactive session directories are auto-cleaned after `runtime.auto_clean_recent` (default `1d`; use `off` to disable). `/clean` removes inactive sessions immediately.
+- Project data: user rules and code indexes are stored under `~/.nanocode/projects/<project_key>/`.
 
 ## Status
 
-- Status bar: active model, reasoning, active yolo/plan modes, conversation context, current-turn tool calls, tokens, elapsed time, and active model-call time.
-- `/status`: active provider, model state, session id, runtime state, conversation/tool counters, per-model calls/tokens, task, goal, and verification.
+- Status bar: active model, reasoning, active yolo mode, conversation context, current-turn tool calls, tokens, elapsed time, and active model-call time.
+- `/status`: active provider, model state, session id, runtime state, conversation/tool counters, per-model calls/tokens, goal, and checks.

design.md

@@ -17,6 +17,18 @@ The agent has a work path and a cleanup path:
 
 Conversation compaction is a background maintenance path. It summarizes old conversation history when the conversation list grows too large.
 
+## Model Output Protocol
+
+Model decisions use function tools:
+
+- state tools update goal, plan, hypotheses, known facts, verification, and result retention
+- repository tools read, search, edit, run commands, and recall stored results
+- compaction uses a dedicated `compact` function tool
+
+Assistant text is optional user-facing text. It must not replace the next useful
+function tool. Completing work still requires a `goal` function tool call with
+`complete=true`.
+
 ## Task State
 
 The main task state lives in the blackboard:
@@ -27,55 +39,61 @@ The main task state lives in the blackboard:
 - plan
 - hypotheses
 - known facts: settled facts for the current task
-- stable knowledge: rare reusable codebase facts
 - verification state
 - recent edits
 
 New user input keeps the previous task state available for follow-ups like "continue".
 
-Old task state is cleared only when the model explicitly starts a different goal. When that happens, transient investigation state such as hypotheses and selected tool-result context is reset, while durable knowledge is kept.
+Old task state is cleared only when the model explicitly sets a different goal. When that happens, transient investigation state such as hypotheses and selected tool-result context is reset, while durable knowledge is kept.
 
 ## New Goal Handling
 
 New user input does not immediately clear the previous task. This keeps short
 follow-ups such as "continue" usable.
 
-When the model outputs `start` with a different goal:
+When the model outputs `goal` with a different current-task goal:
 
 - goal and plan are replaced
 - hypotheses are cleared
 - verification is reset
 - kept tool results are cleared
 - visible raw tool results are compacted into summaries
 - full tool logs remain available through `Recall tr.N`
-- known and stable knowledge remain available
+- known facts remain available
 
 ## Context Construction
 
 ACT mode receives a working context:
 
 - goal, plan, hypotheses, verification
+- environment, including whether local symbol inspection is available
 - Tool Result Index
 - Kept Tool Results
 - Unreduced Tool Results
 - Latest Tool Results
 - errors
 - recent edits
-- known and stable knowledge
+- known facts
 - conversation history
 - latest user request
 
 OBSERVE receives a smaller cleanup context:
 
 - latest user request
 - goal, plan, hypotheses
-- known and stable knowledge
+- known facts
 - kept tool results
 - observe errors
 - unreduced raw tool results selected from recent/latest storage
 
 OBSERVE reduces tool-result noise before ACT continues.
 
+The code navigation tool is environment-gated. `InspectCode` is shown only when
+the built-in code index is available. It supports `find`, `inspect`, and
+`outline` modes for symbol queries or file paths, not natural-language
+questions. The index is created explicitly with `/index`, rebuilt with
+`/index force`, and lightly updated at startup when it already exists.
+
 Context layout:
 
 Layout rules:
@@ -93,7 +111,6 @@ ACT user prompt, top -> bottom
 +--------------------------------------------------+------------------------------+
 | Background                                       | compact_at                   |
 |   - Environment                                  |                              |
-|   - Stable Knowledge                             |                              |
 |   - User Rules                                   |                              |
 |   - Conversation History                         |                              |
 +--------------------------------------------------+------------------------------+
@@ -113,7 +130,7 @@ ACT user prompt, top -> bottom
 | Current Decision                                 | section-local limits         |
 |   - Recent Edits                                 |                              |
 |   - Known                                        |                              |
-|   - Task Code / Work Mode                        |                              |
+|   - Current Phase / Work Mode                    |                              |
 |   - Goal / Plan / Hypotheses / Verify            |                              |
 |   - Errors                                       |                              |
 |   - Latest User Request                          |                              |