navori
What's new

What changed, version by version

What you gain by upgrading, in plain language.

  1. v0.12

    First-class Pi, a cheaper gate and retirements with a migration path

    • Pi is a first-class engine and is now offered in the init wizard and in configure engines. It reaches guard and hook parity with Claude Code, and sensitive actions ask for a typed approval (yes, run it). Requires Pi >= 1.1.0.
    • New navori gate <fast|full>: compact output with the verdict on the first line and the full log under .navori/state/gate/. Plus a scoped gate (qualityGate.scoped) that runs only what the diff touched, and a native gate hook (qualityGate.nativeHooks) that avoids running it twice.
    • navori update checks npm, installs the latest CLI after your confirmation and then runs navori refresh, which is the old update (it re-syncs the harness). There is no alias: if your scripts used navori update to refresh the harness, switch them to navori refresh.
    • Breaking retirements: the navori global command and the launchd collector are gone, doctor --json no longer emits globalScope, and the jscpd and semgrep plugins leave the product. Migrate with navori render --apply, which prunes their assets when they are navori's own, and navori doctor, which reports any leftovers.
    • Node >= 22.12 is the minimum. progress/current.md is now capped (the hook blocks agent commits that grow it and doctor warns) and harnessVersion works as an anti-rollback floor.

    Notable patches

    • v0.12.1 — Fixes plan classification, evidence and escalation; adds master plan improvements and Codex approval and profile controls. Codex integration remains under validation.
    • v0.12.2 — A faster dev cycle: the gate reuses results by content identity and runs the destination repo's gate, with a faster test suite and lighter hooks. The scribe can run in parallel (opt-in), Codex dispatch is isolated per session, and doctor warns when the global CLAUDE.md exceeds its context cap or points to obsolete workspace paths. Also fixes the progress cap, completing specs over the limit, Pi plan-gate diagnostics, the CLI shebang and the reviewer's real PR base.
  2. v0.11

    Per-checkout harness state and a guided master plan

    • Ephemeral harness state now lives per checkout and engine-neutral under .navori/state/handoffs/. plan gate, plan check and receipt check look there for workplans and receipts, with the old .claude/progress/ as a read fallback. It is a migration worth noticing.
    • Codex moves toward operational parity, though still partial: default permissions, validated hook output, cross-review permissions and an advisory plan-gate on Codex.
    • New master-plan skill: it guides a staged master plan for large work, backed by its own spec.
    • New library skills: amqplib for RabbitMQ, Mantine patterns for any repo with @mantine/core, and fastapi and pytest for Python repos, with Python dev-dependency detection.
    • The pre-commit, semgrep and jscpd gates no longer run the anchor repo's gate on commits that land in another repository, and init infers the workspace when one already registers the repo. Also, jscpd blocks only new clones, with fixes to check-links, receipts, tailwind-v4 activation, plugin extra variables and init coexistence.

    Notable patches

    • v0.11.1 — Claude first: the harness stops emitting what Claude Code already ships natively, navori audit measures ranges and compares snapshots, and there is a search-routing guard, confirmation when a gate cannot reach a verdict, Pi support and a daily new-version notice.
    • v0.11.2 — Codex parity (spec 0041), verified live on Codex 0.160.0 with a guarantee table: role-guard and spawn guard, plan-gate with a dispatch file for multi-agent V2, publication and general-purpose confirmations by deny, master-plan and context-intake, a tgrep guard and advice on repeated Bash failures. Codex models resolve automatically by family from the local catalog (sol, luna, astra), and doctor checks Codex trust per worktree and warns about a stale global navori. Render fixes: navori:if in Codex agents and orphan pruning.
    • v0.11.3 — Functional deliveries for specs: spec classify and spec check, sdd.deliveries with deliveries, milestones and tasks, one full gate per PR and a master-plan that splits its parts with the same rules. sync replaces file by file (including markerless files), with --accept-new-files and a diff to keep or accept; monorepo.workspaceHarness gains the root mode. navori tools notices new releases of external tools and doctor warns about known-bad versions (engram < 3.0.0). navori audit joins review and receipt outcomes, measures availability, compares cohorts and computes per-task efficiency. Safer hooks: the guard blocks whole-tree discards with uncommitted work, the publisher cannot rewrite history, plan-gate asks for confirmation and blocks say what to do. Codex gets a thinner AGENTS.md and an engram write guard, and no longer downgrades the rendered model.
  3. v0.10

    Tiered planning and a harder-to-break harness

    • Work is planned in tiers: a classifier decides how much planning each task needs before any work is dispatched.
    • The scribe agent can own all of the repo's Markdown, so other agents leave it alone.
    • All AI attribution is removed from commits, PRs and code, and the rule now covers comments too.
    • doctor checks that the binary matches the version pinned in the manifest and that MCP wiring is coherent.
    • New skills: Supabase, authoring your own skills, back-office patterns, and helpers for rebases and worktree cleanup.

    Notable patches

    • v0.10.1 — Nine new library skills, a guard that also blocks absolute paths to managed files, and warnings about missing plugin binaries on init.
  4. v0.9

    More agents and more dependable closes

    • New agents: scribe (writes documentation) and architect (designs the solution before implementation).
    • An advisor suggests which main model fits your session.
    • The gate catches dead links in documentation and caps the word count of CLAUDE.md and managed blocks.
    • New secure-by-design and quality-attributes skills, and the review receipt is now signed and checked with one command.
    • A quality gate killed by a timeout is no longer reported as green.
  5. v0.8

    Indexed search with tgrep

    • New tgrep plugin: indexed text search, with its usage doctrine built into the harness.
    • codegraph now starts loaded with the session instead of deferred.
    • navori audit --arm turns on audit mode in the session already running, without relying on the model.
    • Audit metrics measure better: the version that ran, the real cost of hooks and round-trips per stretch.
    • Closing a cycle leaves the repo parked on the synced base.
  6. v0.7

    More permission modes, foreign harnesses and the global layer

    • navori audit attributes more accurately: hooks are no longer charged to agents that didn't run them, and it shows a histogram per permission mode.
    • The harness recognizes all six Claude Code permission modes, and doctor warns when yours isn't supported.
    • doctor detects foreign harnesses that clash with yours and offers to adopt them.
    • Agents and skills can live in your machine's global scope, and global uninstall removes only what navori wrote.
    • A lighter always-on context and skills updated for Keystone 8, Prisma 7 and Zod 4.
  7. v0.6

    Solution design and ticket triage

    • A solution-design layer reviews the approach with a second pass in a fresh context before implementation.
    • When a ticket arrives, the harness separates the problem from the proposed solution and verifies the real size of the change.
    • render and doctor manage the harness .gitignore and check git hygiene.
    • New navori audit and per-session audit mode, to see what the harness did in your sessions.
    • New babysit-prs skill and five more rules in the review checklist.
    • Library skills are also detected from project files, not only dependencies.

    Notable patches

    • v0.6.5 — navori audit is redesigned: a per-agent card, the harness records its own execution, and activation no longer fails silently.
  8. v0.5

    Security hardening and a better doctor

    • Several attack paths are closed: command injection in generated hooks, workspace path traversal and marker forgery.
    • New global harness: a per-machine base in ~/.claude.
    • New domain knowledge base for the workspace that agents can consult.
    • doctor checks more: codegraph, hook drift, engines and workspaces.
    • New lifecycle hooks (Stop, SubagentStop, PreCompact), and navori now runs on its own harness.
  9. v0.4

    Surgical context with codegraph

    • New codegraph plugin: agents get code context through MCP instead of reading whole files.
    • On session start the harness injects the branch, recent commits and task state.
    • The review receipt is bound to the diff's content, so a review no longer covers later changes.
    • Preset-aware permissions to cut prompts, and the generated CLAUDE.md is more compact and in English.
  10. v0.3

    Full support for Codex, Cursor and Copilot

    • New Codex engine with AGENTS.md, orchestration and per-role sandboxing, plus its own doctor checks.
    • Adapters for Cursor and Copilot: the same harness renders to each tool.
    • render --all applies the harness across all your registered repos at once, and init --full installs everything.
    • Monorepo support, new presets (Bun/Keystone, React Native with Expo) and the auditor agent.
    • CLI messages follow the configured language.
  11. v0.2

    The CLI foundations

    • The init command detects your stack and generates navori.config.json along with the agent harness.
    • render and sync keep managed blocks up to date without touching what you wrote.
    • Presets, plugins and skills installable with add, plus doctor to diagnose the harness.