cd /news/ai-agents/agnostic-ai-v0-88-0-inspect-and-chan… · home › topics › ai-agents › article
[ARTICLE · art-157155] src=agnostic-ai.org ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

agnostic-ai v0.88.0: inspect and change agent setup in your browser

Agnostic-ai released version 0.88.0, adding a bundled browser interface that lets users inspect and change agent configuration without a separate frontend installation via the `agnostic-ai view`, `view --global`, and `view --json` commands. The release also adds preview-and-undo for project built-in switches and a `verify compare` command that records command outcome, measured duration, detected CLI identity, configured model, and harness fingerprint to flag regressions between saved verifier runs. Projects without a `targets` list now default to including Pi, Qwen Code, and Mistral Vibe, and users upgrading from v0.87.0 must stop `agnostic-ai watch` processes before installing; no portable spec-format migration is required.

read5 min views3 publishedOct 24, 2026

Upgrading from v0.87.0 #

Stop running agnostic-ai watch processes before changing the installation or project config.

Check implicit target selection before installing. Projects without a targets list now include Pi, Qwen Code, and Mistral Vibe in the defaults. To keep a deliberate selection, write the tools you intend to sync explicitly before updating. Existing explicit lists stay explicit. Import hand-written native configuration before enabling a new target; follow its guide below for trust, model, and output restrictions.

For development checkouts using view authority or view overview, replace those calls with view or view --json.

Install through your existing package manager. For a global npm installation:

npm install -g agnostic-ai@0.88.0

After installing, agnostic-ai upgrade --requires updates requires and the editor schema URL, then syncs. It replaces minimum or range requirements and local overrides too. To preserve a deliberate minimum or range, keep that requirement and update the schema URL to v0.88.0 yourself. See upgrade options.

No portable spec-format migration is required from v0.87.0. Run agnostic-ai sync, review the generated diff, and run agnostic-ai sync --check before restarting watch. OpenCode V2 users should regenerate affected agents so configured tool restrictions are preserved when native V2 fields select that format. Review native edits before importing them over a source spec.

What agnostic-ai changed #

Inspect setup in a bundled browser

Open the project or user configuration without a separate frontend installation:

agnostic-ai view
agnostic-ai view --global
agnostic-ai view --json

Bare view finds the project in the current directory or its parents, then falls back to user configuration outside a project. The browser separates Context, Authority, Tools, Lifecycle, and Settings. Search finds source files, fields, and cached MCP tool results; expandable findings provide details and copyable fix prompts. Light and dark themes and narrow-screen navigation are included.

agnostic-ai serve --open starts the same local interface. Root help now groups commands by task. The browser walkthrough shows how to inspect a warning and find its source.

Review built-in changes before applying them

Project built-in switches preview the source edit and generated changes for every configured target. Apply writes the reviewed files; one shared undo restores the previous files and permissions. Pending source updates can join the preview when files are missing or recorded checksums prove ownership. Hand edits require CLI review, and changed inputs block stale apply or undo requests.

Run an initial sync before using project previews. Undo lasts for the server session. Global switches save source configuration; run agnostic-ai sync --global to apply it. See saving changes and undo for recovery and memory-migration limits.

Compare saved verifier outcomes

With your existing verify.command configured, save an external check before and after changing an instruction:

agnostic-ai verify --target codex --evaluation smoke-v1 --record before.json
agnostic-ai sync
agnostic-ai verify --target codex --evaluation smoke-v1 --record after.json
agnostic-ai verify compare before.json after.json

Records keep command outcome, measured duration, detected CLI identity, configured model, and harness fingerprint. A pass followed by a failure reports a regression. Changed or unknown evaluator identities and different evaluation identifiers are non-comparable. Failed evaluations retain their exit code when saving fails; a saving error after success exits 1. Result files are private and replacing one needs --overwrite. The verification reference defines identity boundaries, JSON, and exit semantics.

Use six more native targets

Pi, Qwen Code, and Vibe join default selection. Grok Build, Kimi Code CLI, and CodeBuddy Code require explicit selection:

Target Shipped configuration and adoption limit
Pi (pi ) Instructions, skills, prompt commands, MCP, and model settings. Native project settings require trust; portable lifecycle hooks and tool restrictions have no native mapping.
Qwen Code (qwen ) Instructions, agents, skills, commands, MCP, hooks, settings, and ignores. Use native exclusions when inherited same-name MCP servers must be blocked.
Mistral Vibe (vibe ) Project and user instructions, subagents, skills, MCP, hooks, and model settings. Host sandbox or agent-socket mode does not load ordinary project .vibe configuration.
Grok Build (grok ) Opt-in native project and user configuration. Keep instructions readable with gitignore.commit: [grok:instructions] ; native trust and hook approval remain required.
Kimi Code CLI (kimi ) Opt-in instructions, agents, skills, MCP, and selected user settings/hooks. Project hooks need a separate explicit data-root export; scoped shell restrictions cannot be represented.
CodeBuddy Code (codebuddy ) Opt-in native configuration. Agent-name and tool-restriction limits apply; root MCP output requires an explicit override. Sync does not grant native trust.

Preserve policy and audit evidence

OpenCode V2 agent fields keep tool restrictions and ordered permission rules through sync and import. Global import retains an existing MCP spec when the native name differs only in case, and project checks skip built-in hooks that do not target the selected tools.

Vendor watch filters known counter-only changes, labels historical additions as unreviewed, and detects edits to existing GitHub release notes. The first fetch against the old release-index format stays changed until normal audit review advances it. The full release notes cover the remaining target and documentation fixes.

What changed in the targets #

Primary sources were checked on October 25, 2026. These upstream changes have their own availability and policy boundaries:

  • Vibe 3.0.2 applies subagent controls and raises concurrency. Released October 22, it fixes hooks inside subagents and applies custom tool policies and skill filters. Defaults rise to 32 alive and 32 concurrent subagents. agnostic-ai preserves native policy throughx-vibe agent fields and acceptsx-vibe.max_subagents andx-vibe.max_concurrent_subagents in settings. These remain native Vibe controls.Vendor release .
  • Codex 0.164.0 includes model-request observers. The October 23 release's source containsPreModelRequest andPostModelRequest , and agnostic-ai now accepts those native event names. Their output is diagnostic; they cannot block or modify a request, and need no matcher.Release ,released implementation .
  • Kimi Code CLI 2.3.1 fixes Windows trust failures and stale session locks. Released October 23, these fixes come from the native CLI update. The newkimi target manages configuration; use the updated Kimi installation for its runtime fixes.Vendor release .
  • OpenCode V2 needs an explicit outside-project policy. This documentation change was discovered October 16 and rechecked October 25. Ordered V2 permission rules use the last match, and outside-project access defaults to allowed. Add a nativeexternal_directory ask rule if you need the earlier prompt behavior. v0.88.0 preserves generated V2 restrictions; it does not silently choose that policy for you.Migration guide ,permissions .
── more in #ai-agents 4 stories · sorted by recency
── more on @agnostic-ai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/agnostic-ai-v0-88-0-…] indexed:0 read:5min 2026-10-24 · —