Start here

A little clarity
goes a long way.

Get from “why is my agent ignoring this?” to a source file, a loading rule and a next step.

Install Sightline

You’ll need Node.js 22 or later on macOS, Windows or Linux. Open a terminal in your project folder and run this command to inspect your repository in the browser.

npx --yes https://github.com/iamzwxyy/agent-sightline/releases/download/v0.3.47/agent-sightline-0.3.47.tgz ui

The inspector reads project and user configuration locally. Keep the terminal running while you inspect, then press Ctrl+C to stop.

For repeated use, install the v0.3.47 CLI package.

Download CLI package ↓

Runtime dependencies are bundled. After downloading the package, you can install offline with npm install -g --offline ./agent-sightline-0.3.47.tgz, or install directly from the public package URL.

npm install -g https://github.com/iamzwxyy/agent-sightline/releases/download/v0.3.47/agent-sightline-0.3.47.tgz

Open the bundled sample in your browser. It includes intentional configuration problems and excludes personal configuration.

agent-sightline demo --ui

You can also open the sample with a single command.

npx --yes https://github.com/iamzwxyy/agent-sightline/releases/download/v0.3.47/agent-sightline-0.3.47.tgz demo --ui

The package is distributed through public GitHub Releases and website downloads. Public npm-registry publication is pending.

Your first scan

Open a terminal in any repository subdirectory and run the CLI. Sightline finds the nearest Git root and includes applicable ancestor configuration. It discovers documented configuration sources and reports deterministic problems.

Relative target paths resolve from your current directory. Use --repo to select an explicit inspection boundary and resolve targets from that directory. Diagnostics follow the selected target and client. Discovered files from other projects retain their source status. Select their target or project boundary to diagnose them.

agent-sightline

For the browser inspector, run the UI command. It opens a local page with an overview of your clients and actionable findings.

agent-sightline ui

The first terminal summary lists loaded instruction sources, effective bytes, conditional sources and truncation. In the browser overview, choose Inspect conditional loading to open a source and read its trust, approval or session requirements.

Choose a target file in the top bar. Search by filename, path or fuzzy query, then use the arrow keys and Enter to select a result. You can also inspect a typed path. Press ⌘K on macOS or Ctrl+K on Windows and Linux to focus the target picker.

Follow an instruction through

Select a client from the sidebar. Sources appear in loading order. Choose a file to read its source preview, inspect its loading decision and see the exact byte contribution. Repeated rows for the same file keep their own loading decision and import origin. In the explanation panel, open an import origin to inspect its declaring line. Reload and browser history retain the selected occurrence in local and portable reports.

“All discovered” includes ignored and shadowed files, so you can find a configuration that exists on disk but doesn’t reach this client. Press / to search source paths and import origins. Combine terms such as shared beta.md:3 to find a specific import. Matching ignores case and slash direction. If In scope hides matching sources, choose Show all discovered to reveal and focus them. Clear search restores results and focuses the search input. Tab enters the list once. Use arrow keys or Home/End to move and Enter to select. A source jump focuses the filename heading with its loading status. Shift+Tab returns to the source list. Long lists reveal the selected row after source jumps, reload and browser history. Desktop search and filters remain above the scrolling list. Source origins and instruction coverage links clear the source search so the destination appears in the list. Browser Back/Forward clears a search that excludes its destination.

Claude on-demand instructions use the captured model and context settings to determine their character budget. A truncated source shows the first omitted original line and column. Unresolved session model or context inputs retain conditional limits. Refresh after changing the model to inspect its new loaded prefix. Exported reports retain the captured inspection.

Switch to Read effective text to read the exact accumulated instructions. Each source in the loading chain links back to its definition. Open Settings to search resolved values, inspect the winning definition and follow the layers it replaces.

You can get the same explanation in the terminal.

agent-sightline explain src/payments/refund.ts --client codex --trust trusted

The target’s directory models the starting directory of a fresh agent session. The trust selector changes the modeled assumptions. It doesn’t modify your client’s saved trust settings.

Inspect a Codex profile

When you launch Codex with --profile review, inspect the same profile in Sightline. It loads $CODEX_HOME/review.config.toml after user defaults and before project settings. Select a setting to follow its winning and replaced definitions.

agent-sightline explain src/payments/refund.ts --client codex --profile review --trust trusted

The local inspector’s Codex profile field selects the same file. Leave it empty and apply to return to base configuration. The choice survives reloads and appears in report snapshots. Profile files require user configuration, so --no-user excludes them.

Compare client perspectives

Open Compare clients to find identical instruction paragraphs across different source files. Filter shared content or different recipients, then follow each occurrence to its client and original source lines. Repeated paragraphs retain every occurrence. The exact loaded text also appears side by side, followed by a source-by-client coverage matrix. Select the clients you use and turn on Differences only to focus on loading differences. Click any status to inspect the source and its reason for that client.

Coverage includes structured imports and each client’s configuration role. A status with multiple contributions opens the individual occurrences and their source lines.

Extension comparison shows commands and definitions for MCP, skills, hooks and permissions. Every active and conditional candidate has its own action and loading status. Select a definition to inspect its configuration and sources.

agent-sightline diff src/payments/refund.ts --clients codex,claude,copilot

Your client selection, content filter, Differences only choice and expanded paragraphs survive reload and browser history. Refresh scan retains the comparison, and Export report saves it for offline viewing and re-sharing.

Line-by-line comparison aligns two selected clients’ redacted effective text. Choose the left and right client, show unchanged lines for context and use Inspect sources to open a client’s loading chain. Line numbers refer to the displayed effective text. The view marks additions and removals in the alignment and highlights changed words and characters inside replacements. Emoji and combining accents stay together. Larger replacements keep complete line-level highlighting. Longer comparisons can expand beyond the first 200 rows. The pair, unchanged context and expanded lines survive reload, refresh and offline sharing.

Share what you found

Export report creates a standalone HTML snapshot. Opening the file restores the selected view and source location, including its client, import occurrence and highlighted line. Read effective text also reopens in that mode. Re-export an offline report to share its new location. A browser link with a view or source fragment opens that destination. It includes the interactive source explorer, client comparison and findings. Related sources link to the owning client and source line. Source jumps bring the selected line into view. Reload and browser Back/Forward retain the line in local and portable inspections. Reload also reveals the linked line after keyboard navigation to the source controls. Refresh scan retains the selected line in the local inspector. After refreshing, the fixed toolbar summarizes current errors, warnings and discovery coverage. Use Review findings or Review scan scope to continue. Navigation and export stay accessible while reading long sources. The local inspector also keeps target controls and scan errors in view. Copy location gives you the path and position for a filesystem finding. Open it in a browser or send it to a teammate.

agent-sightline report --format html --output report.html

The snapshot replaces captured repository/home paths, verified configuration aliases and custom agent roots. It redacts environment/header values and recognized secrets. Unrecognized paths in prose or shell snippets can remain. Review personal information before sharing.

v0.3.47 conceals captured private values split at instruction truncation boundaries. Regenerate and replace affected earlier reports before sharing them. Existing exports retain their captured content.

v0.3.13 corrects saved Copilot trust matching for directory aliases, symlink escapes and malformed or conflicting declarations. Rerun affected v0.3.11 inspections to refresh workspace MCP and project-setting states.

v0.3.11 protects personal-state previews when the inspection boundary contains the active Copilot configuration directory. v0.3.10 could include internal config.json state in JSON and HTML previews under that condition. Review affected older reports before sharing.

v0.3.10 fixes copied-value redaction after malformed TOML. In v0.3.9, an environment value from that file could appear in a report when repeated in loaded instructions. Review affected older reports before sharing.

Explore an example report →

Catch configuration errors in CI

Use repository-only discovery for a reproducible CI scan. Set the severity that should fail your build.

agent-sightline ci --no-user --fail-on warn --json

Exit code 0 means the inspection completed without reaching your threshold. Code 1 means findings reached the threshold. Code 2 identifies an input or inspection error. Rule IDs and JSON schema version 1.0 support scripted integrations.

Know the inspection scope

Codex 0.167.2 base/project scalar and MCP values and their winning files are compared with its real configuration API in twenty-three isolated scenarios. Twenty-one scripted native requests also compare exact AGENTS.md bodies, byte limits, override selection, root markers, trust and selected profile budgets/fallback filenames. Other profile fields and source line numbers use pinned source/schema fixtures. Codex Skill names, canonical files, enabled states and user/project scope also have native inventory comparisons, including directory links and disabled paths. Ninety-eight isolated native Claude Code 2.1.322 scenarios compare instruction order, exact loaded text, model/context character budgets, rules-file imports, AGENTS modes, path-scoped loading and saved external-import approval. Scoped rules expand imports at startup while their own body waits for a matching file. Project imports outside the session directory use separate saved approval, with pending and rejected states retained. Linked worktrees use a shared Git project identity. Standard .git directories select the main checkout, and separate Git directories retain their own key. Windows launch aliases that differ from the canonical path retain their own entry. Windows approval keys use forward slashes. MCP transports, approval boundaries and stdio directory behavior are also compared with the native CLI. Eleven native local-scope registrations compare saved project keys across directory aliases, nested/non-Git sessions, wider inspection boundaries, linked worktrees and separate Git directories. Local servers replace same-name user and project candidates. Invalid local declarations retain eligible lower-scope definitions. Nineteen isolated permission checks use a loopback scripted API to inspect native tool execution, deny/ask/allow priority and workspace grant admission. Local grants retain conditional Git/session checks until workspace trust is explicitly modeled. Copilot CLI 1.0.110 settings merges, personal aliases, file loading, workspace MCP admission and shared Claude-rule discovery are compared with its native configuration API and CLI inventories. Personal setting schemas have partial coverage. Other Claude capabilities use official documentation contracts with executable fixtures. Gemini tree and flat context imports have thirty-one executable comparisons with its published 0.68.0 processor. Imported text retains source segments and native memory boundaries. Gemini, Cursor and OpenCode V2 have experimental adapters with partial scope.

Open Scan scope & rule versions in the inspector for official source links and each adapter’s limits. Command-line overrides, running-session toggles, installed plugins and remote managed policy may contribute additional configuration.

Discovery defaults to 50,000 repository entries and 4,000 matching files per user configuration root. Use --max-entries and --max-user-files to adjust these budgets. Incomplete discovery stays visible in the inspector and exported reports, including when findings are filtered.

Scans run locally, with no telemetry, model calls or uploads. The default scan reads your repository and documented user configuration locations. Add --no-user for repository-only discovery. Use --system to inspect Claude managed settings, policy drop-ins and exclusive managed-mcp.json. Blocked MCP entries link to the policy that excludes them. External private trees are skipped.

By default, extension diagnostics check discovery and configuration. The explicit doctor --probe command starts local stdio MCP processes, initializes them and lists tools. Starting server code can have side effects. Hook commands and MCP tools are never invoked by a scan.

Copy this command

Copy the selected command with ⌘C or Ctrl+C. On a phone, press and hold the text and choose Copy.