pi-lens Usage Guide
September 10, 2026 · View on GitHub
This page holds the detailed usage/reference material that does not need to live in the repository front page.
Lifecycle overview
pi-lens hooks into the pi agent lifecycle:
session_startresets runtime state, hydrates project caches, preinstalls likely tools, warms LSPs, and starts background project scans.tool_callrecords read coverage and prepares read-guard/autopatch state before a write/edit lands.tool_resultrecords file mutations, runs format/autofix/LSP/runners, and stores diagnostics on the runtime coordinator.turn_endmerges blockers/advisories, refreshes project-diagnostic caches, runs selected project-level checks, and injects findings for the next turn.
On-write pipeline
For each write/edit, pi-lens runs a language-aware pipeline. Format and safe autofix (steps 1–2) run on different schedules depending on the tool:
write(new file, full overwrite, bash-authored write): autofix runs immediately, in the same tool result, carrying the fixed file's full content back to the agent when it fits under the per-file cap.edit: autofix defers toagent_end, joining the same per-file queue as deferred formatting; autofix drains before format so the final state is formatter-stable. Awritefollowed by anediton the same file in the same turn demotes the write's autofix to deferred too.
- Format queue / immediate formatting when configured (deferred to
agent_endfor anediteven under--immediate-format, so it lands after autofix reformats the file). - Safe autofix from tools with deterministic fix support — immediate for
write, deferred toagent_endforedit(see above). - LSP file sync and diagnostic wait.
- Parallel dispatch runners: LSP, ast-grep, tree-sitter, fact rules, and
language-specific linters/security scanners. For a deferred
edit, these run against the not-yet-autofixed disk state, so a lint finding autofix would have cleared may appear here and resolve itself atagent_end. - Cascade diagnostics for likely affected neighbors.
- Deduplication and routing to blockers, actionable warnings, or code-quality history.
See docs/agent-guide.md
for the consumer-facing version of this routing.
Agent tools
pi-lens exposes these high-value tools to agents:
lens_diagnostics—source=session(default) reports cached findings; an empty cache is not proof of a clean file. If changed files have no cached diagnostics or stale findings, usesource=lsp scope=pathsfor targeted active verification (orscope=workspaceto sweep a whole directory/project).lsp_navigation— IDE-style code navigation (definition/references/etc.).ast_grep_search/ast_grep_replace— AST-aware structural search/replace.module_report/read_symbol— navigable outline and targeted symbol-body reads; prefer these before broad full-file reads.
Project config
Project-level .pi-lens.json can configure mutation policy, ignore patterns,
and selected rule thresholds. Global config lives under
~/.pi-lens/config.json.
Typical project config:
{
"format": { "enabled": false },
"autofix": { "enabled": false },
"actionableWarnings": {
"autoFix": { "enabled": false }
},
"ignore": ["generated/**", "fixtures/**"],
"rules": {
"high-complexity": { "threshold": 20 },
"high-fan-out": { "threshold": 25 }
},
"trivy": {
"enabled": true,
"minSeverity": "HIGH"
}
}
The three mutation controls disable auto-formatting, deterministic pipeline autofix, and actionable-warning quickfixes respectively. Diagnostics continue to run. See Global and Project Config for precedence and the complete project schema.
Runtime flags
# Standard mode (LSP enabled by default)
pi
# Optional switches
pi --no-lens # Start pi-lens disabled for this session; /lens-toggle can re-enable
pi --no-lens-context # Disable automatic context injection only (tools/LSP/read-guard/format stay on); /lens-context-toggle
pi --no-lsp # Disable unified LSP diagnostics
pi --no-autoformat # Skip auto-formatting entirely
pi --immediate-format # Format immediately after each edit instead of deferring to agent_end
pi --no-autofix # Skip auto-fix (Biome, Ruff, ESLint, stylelint, sqlfluff, RuboCop)
pi --no-tests # Skip test runner
pi --no-delta # Disable delta mode (show all diagnostics, not just new ones)
pi --lens-guard # Block git commit/push when unresolved blockers exist (experimental)
pi --lens-checkout-guard # Decline git commands that rewrite the working tree while another live pi-lens session shares this dirty checkout (experimental)
pi --no-opengrep # Disable the Opengrep security scanner (default-on auxiliary LSP)
pi --no-read-guard # Disable the read-before-edit behavior monitor
pi --lens-turn-summary # Persist a per-turn summary of diagnostics, autofixes, and autoformats
pi --lens-compact-tool-line # Render tool results as one compact, theme-aware line (closes #1327)
pi --no-lazy-tools # Keep every pi-lens tool active instead of activating the situational ones on demand
pi --lens-turn-end-madge # Run the madge circular-dependency check at every turn end, not just at session start
pi --no-knip # Disable the knip session-start analyzer
pi --no-jscpd # Disable the jscpd session-start analyzer
pi --no-madge # Disable the madge session-start analyzer
pi --no-gitleaks # Disable the gitleaks session-start analyzer
pi --no-govulncheck # Disable the govulncheck session-start analyzer
pi --no-dead-code # Disable the dead-code session-start analyzer
pi --no-complexity # Disable the complexity session-start analyzer
# Actionable warnings (all default off)
pi --lens-actionable-warnings # Report fixable warnings at turn end
pi --lens-actionable-warning-actions # Enrich the report with LSP code-action titles
pi --lens-actionable-warning-autofix # Apply conservative LSP quickfixes at agent_end
pi --lens-actionable-warning-all # Report every warning, not just this turn's
Every one of these has a ~/.pi-lens/config.json equivalent, so you can set it
once instead of per session. See
Global and Project Config for the
flag-to-key table and the precedence order.
Rules
Tree-sitter rules
Tree-sitter rules live under rules/tree-sitter-queries/<language>/ and are
query-based. Use them when you need precise tree relationships or post-filters.
See docs/custom-rules.md and the
pi-lens-write-tree-sitter-rule skill.
ast-grep rules
ast-grep rules live under rules/ast-grep-rules/rules/. Every shipped rule must
have a fixture in rules/ast-grep-rules/rule-tests/. Use the
pi-lens-write-ast-grep-rule skill for schema and runner gotchas.
The shipped baseline combines native pi-lens rules with vendored CodeRabbit
security rules under rules/ast-grep-rules/coderabbit/rules/.
Security and dependency scans
Session-level scanners run in the background and surface at turn end:
gitleaks— committed secrets.govulncheck— reachable Go vulnerabilities.trivy— dependency CVEs, hardcoded secrets, license risk, and IaC config scans when explicitly enabled.
Per-edit IaC config scanning currently covers Dockerfiles and Kubernetes-style
YAML when trivy.enabled is true.
MCP mirror
pi-lens also ships an MCP server for Claude Code or other MCP clients. It is a
second host adapter that calls the same clients/lens-engine.ts seam as the pi
extension. Use npm run build:dist after MCP/engine changes so the user-scoped
server loads fresh compiled code.
Two Claude Code hooks give the MCP mirror the same automatic cadence the pi
extension has. PostToolUse runs the per-edit pass, Stop runs the per-turn one:
{ "hooks": {
"PostToolUse": [
{ "matcher": "Edit|Write",
"hooks": [ { "type": "command", "command": "pi-lens-analyze --hook" } ] } ],
"Stop": [
{ "hooks": [ { "type": "command", "command": "pi-lens-analyze --turn-end", "timeout": 60 } ] } ]
} }
The Stop hook only works against a running MCP server because only that
process owns the session state and pending turn work. With no warm server the
turn-end is skipped, not faked: one line on stderr and nothing on stdout.
Troubleshooting
- Run
npm run buildbefore tests after editing TypeScript; tests import generated.jsartifacts. - Use
lens_diagnostics mode=allto surface stale blockers from the current session. - Check
~/.pi-lens/sessionstart.log,~/.pi-lens/latency.log, and~/.pi-lens/cascade.logfor lifecycle/performance/debug traces. - For live tool validation, use
node scripts/smoke-tools.mjswith the relevant--lsp,--format, or--autofixlayer.