Coding agents can act. PX makes them show their work.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Ask a natural language question — auto-routes to the right operation PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Unified AI question surface — auto-routes via QueryLane and emits an AnswerEnvelope JSON. PRD `px-intent-v1` PART 2 Phase 3 PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Route a concrete agent question to the best PX command trail without requiring the agent to memorize the CLI. PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Scan a newly cloned repo and print the exact PX setup path PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Cold-start orientation packet for a fresh agent in an unfamiliar repo PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Aliases for onboarding flows PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Structured agent-facing docs for a px verb PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Report product boundaries that agents might otherwise infer from proof artifacts or historical names PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Show common PX query mistakes and recent miss context PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Recommend next PX commands for an observed outcome PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Explain whether a structural zero result looks real or like an index miss PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Bundle a reproducible bug packet for a suspected PX miss PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Re-render the most recent cached miss packet PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Emit the last cached ask route receipt as a repro packet PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent guesses a tool, runs broad grep or vector search, reads several files, and silently changes strategy when the first result is weak.

Emit the last cached ask --deep execution packet PX frames the question, preserves hard anchors, chooses a typed lane, exposes why it chose it, and emits the next recovery command.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Unified inspector — one call emits every signal an agent prompt needs PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Query the reference index PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Show callers of a symbol — "who calls this?" PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Show callees of a symbol — "what does this call?" PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Impact analysis — "what breaks if I change this symbol?" PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Side-effect facts for a symbol from the `code_effects` table PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Agent context — render markdown for coder/reviewer agents PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Generate a safe phased edit plan for changing a symbol PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Semantic impact of a rename — "is this a mechanical rename or a role change?" PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Estimate refactor LOC for a symbol — rough blast-radius measurement PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Source-coordinate type and scope lookup PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Test coverage evidence for a symbol from indexed test references PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Conservative taint-spine evidence from a seed to indexed sinks PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

SQL reachability evidence lookup PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

Type complexity signature lookup (from code_types) PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent greps a name, follows imports by hand, asks an LSP one language at a time, and spends tokens reconstructing relationships.

fox_file_claim registry status for a path PX returns source coordinates plus indexed relationships, identity diagnostics, and a bounded current-tree cross-check in one evidence surface.

The agent searches prose separately, trusts whichever document sounds current, and mixes design claims with source or runtime truth.

Documentation atlas, Oracle checks, context packs, and Markdown LSP PX binds docs to code subjects, ranks focused semantic sections, reports conflicts and coverage gaps, and preserves provenance.

The agent searches prose separately, trusts whichever document sounds current, and mixes design claims with source or runtime truth.

Inspect and ingest normalized observed tool-run evidence PX binds docs to code subjects, ranks focused semantic sections, reports conflicts and coverage gaps, and preserves provenance.

The agent searches prose separately, trusts whichever document sounds current, and mixes design claims with source or runtime truth.

Parse Markdown docs into canonical doc graph rows. Read-only by default; persist to doc tables with --write-db PX binds docs to code subjects, ranks focused semantic sections, reports conflicts and coverage gaps, and preserves provenance.

The agent searches prose separately, trusts whichever document sounds current, and mixes design claims with source or runtime truth.

FDAG V3 Code Witness receipts and graph packets for code changes PX binds docs to code subjects, ranks focused semantic sections, reports conflicts and coverage gaps, and preserves provenance.

The agent searches prose separately, trusts whichever document sounds current, and mixes design claims with source or runtime truth.

FDAG V3 replay evidence and agent-memory inspection surfaces PX binds docs to code subjects, ranks focused semantic sections, reports conflicts and coverage gaps, and preserves provenance.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Build a paste-ready worker briefing for one or more owned targets PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Understand a target — produce an `Observation` value (canonical symbols + call-graph + impact + cochange + recent_edits + typed caller detail). Phase B1 Wave γ wires the `Understand` trait (α + β) into the CLI surface with optional receipt persistence and `callers_detailed` enrichment. Mirrors the `Commands::Plan` dispatch shape PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Plan an agent-contract edit — produce an `EditPlan` value (ordered sites + steps + rationale + risks + required_doc_updates). Phase B0 Wave γ wires the `Plan` trait (Waves α + β) into the CLI surface so agents can fetch a typed plan + doc-drift survey without touching code. Mirrors the `Commands::Simulate` dispatch shape PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Simulate an agent-contract change — produce a `Simulation` value (affected sites + predicted failing tests + risk score) without touching the code. Phase A Inc 2 Wave E wires the `Simulate` trait (Waves C+D) into the CLI surface PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Verify an edit against the FoxRef index (and optionally a prior plan receipt) — produce a `Verification` value (effect_diff + predicted fails + risks + verdict + fresh receipt). Phase B2 Wave γ wires the `Verify` trait (α + β) into the CLI surface with optional receipt persistence and Option-B `referenced_plan` pre-check against `execution_receipts`. Mirrors the `Commands::Understand` dispatch shape PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Orchestrated contract-chain flows — compose Understand → Plan → Simulate → Verify into a single call with receipt lineage threaded across every step. Phase C0 Wave α opens the consumer layer on top of the substrate proven 4× by Simulate / Plan / Understand / Verify PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Session contracts, worker transition gates, and dispatch preflight PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Main-Claude worker briefing packet for active sessions PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Inspect or record suggested-next transition outcomes PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Read-only advisory mission pack for agent execution control PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Read-only boundary permit report for proposed agent edits PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Read-only operational confidence verdict from explicit agent risk inputs PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent plans from memory, edits first, discovers collateral damage in CI, and writes a retrospective summary with no machine-checkable lineage.

Read-only contradiction check for agent claims against explicit evidence PX turns understanding, planning, simulation, verification, and closeout into a connected evidence trail before and after mutation.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Project health audit — graph-unreferenced signals, hotspots, coupling, and score PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Self-audit px substrate decay signals PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Show index statistics PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Diagnostic status — repo path, binary index state, DB counts for a space PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

LSP diagnostics — unified query over diagnostics from rust-analyzer, tsserver, eslint, etc PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Naming convention analysis — find symbols that break their module's local style PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Architectural pattern detection — match motif clusters against a named pattern library PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Proactive refactoring recommendations — combine all signals into a ranked action list PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Diamond prospecting — rank files, symbols, and clusters by mining fertility PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Compare two projects — find how source capabilities enhance the target PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Scan target repos for source-pack adaptation opportunities PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Verify portable packs for structure, safety, and Codex readiness PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Extract capability manifest — what this project offers PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Analyze path dependency surface — "how much of each dep do we actually use?" PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Generate C4-style architecture diagrams (Mermaid) PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

C4 container membership for a symbol/file PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Interactive code visualization (call graph, directory tree, architecture) PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Temporal risk report for a symbol or file from git history plus px impact PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Read-only supply-chain/dataflow risk cockpit workflow PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Forecast temporal risk for files changed since a git ref PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Diagnose failing logs into a safe repair plan (no apply) PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Lightweight invariant checks over px fact rows PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Reachable dependency inventory from manifests plus import/use evidence PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Compose taint flow and reachable dependency evidence into exposure risk PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Audit phase coverage P1-P6 (type intel / diagnostics / refactor recs) PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

The agent runs disconnected linters and scanners, then overstates heuristic counts as dead code, safety, or architectural fact.

Extractor orphan check (built but not wired to rebuild) PX composes graph, source, dependency, temporal, and diagnostic evidence while labeling what is heuristic, partial, or unproven.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Self-diagnostic — check if px is properly set up and up to date PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Fast substrate refresh — update selected DB feeders without full LSP rebuild PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Adaptive rebuild — incrementally update Stage 1 when a valid prior index exists, then rebuild the Stage 2 binary index PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Take a file fingerprint snapshot for incremental rebuilds PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Diff two snapshots to see what changed PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Build artifacts from Postgres code_references table PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Verify binary-index structural integrity (internal diagnostic) PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Extractors (ported from fox-codegraph). Run with `px extract <tool> --help` PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Index a repository into FoxGres code graph PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

A periodic full scan lags the tree, partial extraction is published as complete, and agents mistake zero indexed rows for absence.

Extract references via LSP (rust-analyzer, typescript-language-server) for "who uses this symbol?" Read-only by default; persist reference rows with --write-db PX tracks index contracts, overlays changed files, classifies misses, validates mappings, and fails closed when replacement evidence is incomplete.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan Rust files for `use` statements and type annotations to create cross-file reference edges. Resolves crate/super/self paths. Read-only by default; persist rows with --write-db PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan Rust Metal host calls (`get_function("kernel")`) and create Rust-to-.metal kernel reference edges. Read-only by default; persist rows with --write-db PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan tests, fixtures, schemas, goldens, migrations, snapshots, hooks, and runtime config files into agent contract reference edges. Read-only by default; persist rows with --write-db PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan source files for function/method call patterns, including Rust same-file edges. Heuristic: matches known symbol names followed by `(` in source code. Read-only by default; persist rows with --write-db using source='call_scan' PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan JS/TS files for ESM/CJS import statements and create cross-file reference edges. Resolves relative imports (./ ../). Read-only by default; persist rows with --write-db PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan JS/TS files for runtime and type-only import truths with resolver-backed module resolution. Read-only by default; persist rows with --write-db or record evidence with --emit-evidence PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan Pascal/Delphi files for `uses` clause imports and call-site references. Discovers cross-file dependencies from unit references and heuristic call matching. Read-only by default; persist rows with --write-db using source='pascal_scan' PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan Java files for import statements and create cross-file reference edges. Resolves package imports to actual files. Read-only by default; persist rows with --write-db PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan SAP BODS ATL files for cross-symbol references via the tree-sitter-atl grammar. Emits `qualified_name` usages (table.column, schema.table, Query.COL) and BODS built-in calls (lookup_ext, sql, decode, …). Read-only by default; persist rows with --write-db using source='atl_scan' PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan SQL DDL files and find heuristic references from source code to SQL objects. Read-only by default; persist reference rows with --write-db. Discovers CREATE TABLE/VIEW/FUNCTION definitions and matches them against string references in .rs/.ts/.tsx/.js/.jsx files PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Extract SQL function definitions and cross-references from PostgreSQL pg_proc. Introspects plpgsql/sql routines in selected schemas (default `public`), dumps definitions to disk, and discovers function-to-function and function-to-table references PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Extract type-aware signatures from every function/method in the repo PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Extract static Rust/JS/TS/Python side-effect facts. Read-only by default; persist to code_effects with --write-db or record evidence with --emit-evidence PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan every indexed file for shell-script-path references (e.g. `scripts/foo.sh`, `.claude/hooks/bar.sh`) and emit path-to-path edges. Read-only by default; persist rows with --write-db using source='script_scan' PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan Rust axum routers + TS fetch/axios callsites and pair them on canonical URL patterns to produce cross-language HTTP-boundary edges. Writes to code_references with source='cross_lang_http_scan', target_kind='cross_lang_http', ref_type='http_call'. Lights up `px callers <axum_handler>` with TS UI-side callers (Phase 2.2.1) PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan Rust `sqlx::query!`/`query_as!` macros + TS `pool.query` / Drizzle `sql\`...\`` template literals and pair them with plpgsql/sql functions in `pg_proc` to produce cross-language SQL-boundary edges. Writes to code_references with source='cross_lang_sql_scan', target_kind='cross_lang_sql', ref_type='sql_call'. Lights up `px callers <plpgsql_fn>` with Rust + TS app-side callers (Phase 2.2.3) PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan Rust `#[napi]`-decorated functions + TS `loadNativeModule()` consumers and pair them on the canonical NAPI export name (snake_case ↔ camelCase normalized). Writes to code_references with source='cross_lang_ffi_scan', target_kind='cross_lang_ffi', ref_type='napi_call'. Lights up `px callers <#[napi]_function>` with TS UI-side callers (Phase 2.2.2) PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

Each language or boundary gets a bespoke grep, and cross-language calls disappear between API strings, SQL, scripts, generated code, and FFI.

Scan `pg_notify('<channel>', ...)` producers (plpgsql / SQL / Rust / TS) + `LISTEN <channel>` / `pool.listen("<channel>")` consumers and pair them by literal channel name. Writes to code_references with source='cross_lang_message_scan', target_kind='cross_lang_message', ref_type='message_publish' or 'message_subscribe'. Closes the Phase 2.2 cross-language series — best-effort coverage; dynamic channel names (variables, format strings) are out-of-scope for V0 (Phase 2.2.4) PX normalizes many extractors into one graph and records exactly which extractor produced each kind of structural evidence.

The agent opens repositories one by one, guesses which local or remote database is authoritative, and manually reconciles versions.

Cross-repo query — search across all spaces via Postgres code_references PX gives the fleet a common space model, cross-repo query lane, host/version inventory, and readiness receipts.

The agent opens repositories one by one, guesses which local or remote database is authoritative, and manually reconciles versions.

Cross-machine PX host and repo readiness packets PX gives the fleet a common space model, cross-repo query lane, host/version inventory, and readiness receipts.

The agent opens repositories one by one, guesses which local or remote database is authoritative, and manually reconciles versions.

List app database truth-source profiles for SQL substrate extraction PX gives the fleet a common space model, cross-repo query lane, host/version inventory, and readiness receipts.

The agent opens repositories one by one, guesses which local or remote database is authoritative, and manually reconciles versions.

Read FoxGres fleet/version index and upgrade planning guidance PX gives the fleet a common space model, cross-repo query lane, host/version inventory, and readiness receipts.

The agent opens repositories one by one, guesses which local or remote database is authoritative, and manually reconciles versions.

Benchmark suites for px quality signals (Wave 2v-2x, 2026-04-20) PX gives the fleet a common space model, cross-repo query lane, host/version inventory, and readiness receipts.

The agent opens repositories one by one, guesses which local or remote database is authoritative, and manually reconciles versions.

Self-calibrate routing recipes from the index's own graph ground truth PX gives the fleet a common space model, cross-repo query lane, host/version inventory, and readiness receipts.

Tool failures vanish into chat transcripts, agents quietly return to grep, and maintainers cannot measure whether fixes changed behavior.

Generate or read repo-specific PX agent briefings PX records outcomes, transitions, misses, adoption, and improvement receipts so agent behavior becomes an observable product metric.

Tool failures vanish into chat transcripts, agents quietly return to grep, and maintainers cannot measure whether fixes changed behavior.

Collect and inspect agent feedback about PX command usefulness PX records outcomes, transitions, misses, adoption, and improvement receipts so agent behavior becomes an observable product metric.

Tool failures vanish into chat transcripts, agents quietly return to grep, and maintainers cannot measure whether fixes changed behavior.

Record and inspect PX tool improvement memory in FDAG V3 PX records outcomes, transitions, misses, adoption, and improvement receipts so agent behavior becomes an observable product metric.

Every agent host wraps the CLI differently, parses prose, and reimplements routing, batching, authentication, and output contracts.

Start HTTP server — expose FoxRef as a JSON API for remote queries PX exposes stable machine-readable surfaces, completion metadata, batch execution, and a service boundary around the same evidence engine.

Every agent host wraps the CLI differently, parses prose, and reimplements routing, batching, authentication, and output contracts.

Emit shell completion script for the given shell (bash, zsh, fish, elvish, powershell) PX exposes stable machine-readable surfaces, completion metadata, batch execution, and a service boundary around the same evidence engine.

Every agent host wraps the CLI differently, parses prose, and reimplements routing, batching, authentication, and output contracts.

Multi-call batching — execute N px verbs in one call, eliminates per-call subprocess overhead PX exposes stable machine-readable surfaces, completion metadata, batch execution, and a service boundary around the same evidence engine.

The agent stitches together shell searches, source reads, and memory without one evidence contract.

Investigate a broad architecture, subsystem, or flow question with composed FoxRef steps The result stays machine-readable and connected to the next evidence step.