nav

Appearance

Changelog ​

Unreleased ​

[0.9.1] - 2026-05-21 ​

Added ​

  • /tasks verify [id] — manually trigger the verify+fix loop for a specific done goals-mode task, or for all done standalone goals tasks when no id is given
  • /plans verify <id> — manually trigger the verify+fix loop for all done goals in a plan without re-running implementations

Fixed ​

  • Auto-retry failed goals — /tasks run and /plans run now automatically pick up previously-done tasks that still have failedCriteria and re-run verify+fix for them without manual intervention

[0.9.0] - 2026-05-10 ​

Added ​

  • Goals mode — set planMode: "goals" in nav.config.json for outcome-focused planning; /plan guides defining outcomes and acceptance criteria, /plans split produces goals with required criteria (minimum 2 per goal), and task execution leads with criteria while letting the agent decide implementation
  • Goal verification — after all goals complete, a verification phase checks each acceptance criterion; results stored in task.criteriaResults with pass/fail status and evidence
  • Fix-and-reverify loop — failed criteria trigger automatic rework; the agent receives a fix-focused prompt listing failed criteria, then verification runs again; loop repeats until all pass or taskImplementationMaxAttempts reached (default: 3)

Changed ​

  • Plan mode tools — /plan discussion now allows shell and shell_status for codebase exploration (ls, find, tree); edit/write remain blocked to prevent accidental mutations

Fixed ​

  • Markdown parsing in verification — strips ** formatting from criterion text before matching

[0.8.6] - 2026-05-08 ​

Fixed ​

  • subagent.parallelToolCalls — config setting now works correctly; previously it was ignored and subagents always ran sequentially regardless of the configured value

Changed ​

  • Subagent parallelism — subagents no longer inherit the parent session's parallelToolCalls setting; they default to 1 (sequential) unless subagent.parallelToolCalls is explicitly set in nav.config.json

[0.8.5] - 2026-05-08 ​

Changed ​

  • /plan format — the model now writes the plan as a markdown document with YAML frontmatter (name and description between --- lines) followed by the full plan body; parsed deterministically on confirm, replacing the previous fenced JSON block
  • /plans split format — the agent now produces a markdown task list (## section per task, optional **Files:** and **Criteria:** bullets) instead of a JSON array; tasks are saved with IDs like 1-1, 1-2, etc. (/plans microsplit still uses fenced JSON with codeContext)
  • System prompt — default system prompt is now more compact to reduce token usage per request

Added ​

  • Batched edits (edits[]) — both hashline and searchReplace edit modes now accept an edits array for transactional multi-step edits applied atomically in a single file write

Improved ​

  • Delegated run safety — subagent runs default to parallelToolCalls: 1 (sequential) for determinism and lower blast radius; set subagent.parallelToolCalls explicitly to allow parallel tool execution in delegated sessions; tool behavior hardened for delegated agent runs
  • Planning tool restriction — /plan discussion and /plans split modes limit the LLM to read-only tools (read, skim, filegrep) so planning never accidentally mutates files

[0.8.4] - 2026-04-28 ​

Added ​

  • subagent.allowNestedSubagents — optional true in nav.config.json enables recursive subagent delegation from delegated runs; when enabled and subagent stays in the child allowlist, the child system prompt includes <available_subagents> like the main session

Changed ​

  • Delegated sessions default — omitted or false (default) removes subagent from delegated tool allowlists and omits the subagent catalog in child prompts, even if subagent.tools or inherited allowlists listed subagent

Improved ​

  • Subagents guide — documents nested delegation defaults, allowNestedSubagents, and token-cost guidance

[0.8.3] - 2026-04-28 ​

Added ​

  • Subagent parallel tool limit override — delegated runs now accept subagent.parallelToolCalls so nested sessions can explicitly set their own per-message tool concurrency (1–32)

Improved ​

  • Nested subagent tool labeling — chained delegations now preserve and compose context labels across levels so tool output clearly shows depth (for example, main → subagent → subagent → tool)

[0.8.2] - 2026-04-28 ​

Improved ​

  • Nested subagent assistant output — parallel subagent runs buffer the child model reply and show one compact preview line (with inherited colorSlot and context label) instead of interleaving streamed tokens; TUI uses parallel accent colors; UI server emits status info lines with slot and preview
  • Nordic theme parallel accents — refreshed rotating accent palette for concurrent tool and subagent lines
  • Subagents guide — expanded documentation on defining and using project subagents

[0.8.1] - 2026-04-28 ​

Added ​

  • parallelToolCalls — optional nav.config.json / NAV_PARALLEL_TOOL_CALLS (1–32, default 1) caps how many tool calls from one assistant message run concurrently; batches that include ask_user run sequentially; nested subagent sessions always use sequential tools. TUI and UI server emit optional colorSlot on tool events for clearer interleaved output

Fixed ​

  • Subagent block in nav.config.json — setting only model no longer replaces the parent provider, baseUrl, or contextWindow with values inferred from the child model; unset keys keep the parent’s resolved values

Improved ​

  • Parallel tool colors — Nordic and classic themes define rotating accents for concurrent tool lines, results, and verbose diffs

[0.8.0] - 2026-04-27 ​

Added ​

  • Subagents — project definitions in .nav/subagents/<id>.md (YAML name / description plus body as role prefix), <available_subagents> catalog in the system prompt, and subagent tool for delegation with optional subagent config block (model, contextWindow, handoverThreshold, tools)
  • /create-subagent — interactive flow (optional id and purpose on the command line) for the main agent to draft and write a subagent file; reloads the system prompt when the run finishes
  • Tool allowlists — optional tools array in nav.config.json limits which tools are sent to the LLM and described in the prompt; delegated runs can use subagent.tools

Improved ​

  • Subagent logging — nested tool lines, results, info, and errors show the subagent display name in brackets; WebSocket tool.call may include contextLabel for UIs

[0.7.7] - 2026-04-20 ​

Improved ​

  • Microsplit — enhanced plan microsplit behavior

[0.7.6] - 2026-04-18 ​

Fixed ​

  • ask_user tool registration — the ask_user tool is only offered to the model during /plan discussion mode; normal agent runs no longer expose it

[0.7.5] - 2026-04-18 ​

Fixed ​

  • UI server role persistence — systemPromptPrefix is retained when the system prompt is rebuilt (/clear, /init, and after /plans split / /plans microsplit), including reloading AGENTS.md and skills

Improved ​

  • UI server documentation — WebSocket protocol reference and guide for building custom clients

[0.7.4] - 2026-04-18 ​

Changed ​

  • UI server systemPromptPrefix — when thread.create includes a non-empty prefix, Nav’s default “You are nav…” identity line is omitted so the prefix defines the agent role; operational tool and edit guidance is unchanged

[0.7.3] - 2026-04-18 ​

Fixed ​

  • macOS binary signing — release binaries are now ad-hoc codesigned, preventing SIGKILL on launch

[0.7.2] - 2026-04-17 ​

Improved ​

  • UI server — now supports multiple concurrent threads

[0.7.1] - 2026-03-26 ​

Fixed ​

  • Sandboxing — config file–based sandbox settings now apply correctly

[0.7.0] - 2026-03-23 ​

Added ​

  • Hooks — configurable stop, taskDone, and planDone steps in nav.config.json (shell commands and optional custom-command steps); taskDone / planDone support maxAttempts with feedback to the model on failure; NAV_HOOK_TIMEOUT_MS / hookTimeoutMs for shell step timeouts (default 10 minutes per shell step)
  • taskImplementationMaxAttempts (default 3, env NAV_TASK_IMPLEMENTATION_MAX_ATTEMPTS) — cap full work+verify cycles per task; /tasks run and /plans run stop when exhausted instead of continuing to the next task
  • editMode in nav.config.json — default hashline; set to searchReplace for plain-text reads and literal old_string/new_string edits instead of LINE:HASH anchors
  • Hook command steps: optional args string with ${VAR} substitution (hook env + process.env); result fills custom command {input} placeholders

Improved ​

  • TUI and UI server now show which hook is running (stop, taskDone, planDone) with step index and shell command or custom command label before each step executes
  • Print task name when starting to work on a task

[0.6.0] - 2026-03-17 ​

Added ​

  • ui-server subcommand -- run nav as a local HTTP/WebSocket backend for external UI clients
  • UI server configuration flags/env vars: --ui-host / --ui-port and NAV_UI_HOST / NAV_UI_PORT
  • UI protocol docs in docs/ui-server-protocol.md
  • Core transport abstraction (AgentIO) and WebSocket-oriented agent IO modules to support non-TUI frontends

Improved ​

  • Sandbox re-exec arg forwarding now preserves CLI args correctly in both source and compiled runs
  • Safer slash-command tab completion handling in TUI

[0.5.1] - 2026-03-16 ​

Added ​

  • Agent now knows the current date (injected dynamically into the system prompt)

[0.5.0] - 2026-03-16 ​

Added ​

  • Azure OpenAI provider -- use -p azure with AZURE_OPENAI_API_KEY, AZURE_OPENAI_API_BASE_URL, and AZURE_OPENAI_DEPLOYMENT_NAME environment variables (or config file equivalents)
  • azureDeployment config option -- specify Azure deployment name separately from model name

[0.4.0] - 2026-03-07 ​

Added ​

  • skim and filegrep tools -- lightweight file inspection with hashline output, no shell needed
  • @ file references -- type @path/to/file in the prompt to include file contents inline
  • /plans microsplit command -- generate micro-tasks optimized for small LLMs
  • Arrow key navigation for slash-command autocomplete

Improved ​

  • Tool parsing and output formatting
  • Hashline robustness against duplicate hashes
  • Edit tool simplified and made more robust
  • Ollama: pass num_ctx and num_batch sizes; show prompt_eval_count in verbose output
  • Verbose mode: log full LLM calls when enabled
  • Removed duplicate logging
  • Renamed /plans work and /tasks work to /plans run and /tasks run

[0.3.0] - 2026-02-21 ​

Added ​

  • Plan mode (/plan, /plans, /plans split, /plans run) -- discuss an idea, save it as a named plan, generate ordered implementation tasks, and work through them automatically
  • Task management (/tasks, /tasks add, /tasks run, /tasks rm) -- persistent task list with planned -> in_progress -> done lifecycle
  • config-init command -- creates a .nav/nav.config.json with sensible defaults
  • Task halting -- users can stop a running task mid-execution

Improved ​

  • Edit tool robustness -- returns only affected-line hashes after edits (not the whole file), with detailed error messages and fresh hashes on failure so the model can retry without re-reading
  • Hashline parser -- fixed parse error on certain edge-case lines
  • Plan mode conversation flow and task field layout
  • TUI indentation and visual layout

[0.2.0] - 2026-02-16 ​

Added ​

  • Skills system -- define reusable agent capabilities in SKILL.md files under .nav/skills/, ~/.config/nav/skills/, or .claude/skills/; automatically injected into the system prompt
  • /skills command -- list all available skills
  • /create-skill command -- interactively create a new skill

[0.1.1] - 2026-02-15 ​

Improved ​

  • Seatbelt sandbox profile is now embedded directly into the binary (no external file dependency at runtime)

[0.1.0] - 2026-02-15 ​

Added ​

  • Initial release
  • Hashline-based editing -- each line is prefixed with LINE:HASH; edits reference anchors instead of reproducing content, preventing stale-edit conflicts
  • Multi-provider LLM support: OpenAI, Anthropic (Claude), Google Gemini, Ollama
  • Interactive and one-shot modes
  • Shell command execution with background process support (shell + shell_status tools)
  • ESC to stop agent execution; Ctrl-D to exit
  • Slash commands: /clear, /model, /handover, /help
  • Custom slash commands via .nav/commands/*.md or ~/.config/nav/commands/*.md
  • Tab autocomplete for slash commands with live filtering
  • Auto-handover when approaching context window limit (configurable threshold)
  • macOS Seatbelt sandboxing (-s / NAV_SANDBOX=1)
  • Config files: .nav/nav.config.json (project) and ~/.config/nav/nav.config.json (user)
  • Color themes: nordic (default, 24-bit truecolor) and classic (16-color)
  • AGENTS.md support -- project-specific instructions auto-included in system prompt
  • /init command -- generates an AGENTS.md from project context
  • Session logging to .nav/logs/ as JSONL
  • Pre-built binaries for macOS (arm64/x64), Linux (x64/arm64), and Windows (x64) via GitHub Actions
  • Progress spinner during LLM API calls
  • Verbose mode (-v) -- shows diffs, token counts, timing, and context utilization

Released under the MIT License.