Enterprise context economics / editable scenario
PX is not a bigger search box. An agent asks one hard question and PX composes the investigation: exact symbols, Pascal references, call graph, neural meaning retrieval and reranking, SQL literals, documentation authority, architecture boundaries, impact, and a reusable receipt.
One front door -> 345 registered actions -> one bounded answerA repository map, generated agent surfaces, documentation atlas, evidence coverage, and one bounded first investigation.
Run nextpx ask --space dataEssence --json "Which forms bypass CodeBL and access udmDatabases directly?"
natural_languageanchors.extract()intent.frame()routes.compete()winner = evidenceweaknesses.keep()definitionscallers + refsuses + inheritsforms + unitsembedding.search()rerank by intenttables + literalsqueries + writestests + effectschange forecastindexed truth+ current bytesabsence calibratednext.runproof.savedThe same agent contract can traverse Delphi, Rust, TypeScript, SQL, Java, Python, and cross-language boundaries without teaching the model a new navigation ritual.
PX uses structural evidence where exactness exists, embedding retrieval and reranking where names fail, and never lets neural similarity silently become proof.
Coverage, weaknesses, selected lanes, evidence pointers, and the next bounded command become a receipt another agent can continue.
PX does not stop at code symbols. It ingests the repository's Markdown corpus, binds documentation sections to code and architecture surfaces, and generates agent-ready projections that live outside curated human docs.
The agent asks one question. PX composes exact graph + neural meaning + docs authority automatically.px docs bootstrap --repo /workspace/company --space company --write-db --atlas --all-formats --json
one bounded pass
_docs/px/architecture/atlas.ascii.md_docs/px/architecture/atlas.mermaid.md_docs/px/architecture/atlas.json_docs/px/architecture/atlas.html
px ask --with-docs returns the relevant code, the governing documentation, contradictions, coverage limits, and the next safe command instead of making every model reconstruct the company from raw files.
PX writes machine-generated projections under _docs/px/, preserving curated human documentation and recording provenance instead of mixing both truth classes.
health, coverage, orphans, validate, claims, reconcile, and fs-check expose drift instead of letting old prose quietly govern agents.
px ask routes into docs, graph, neural retrieval, current-tree evidence, or a composed investigation. The receipt tells the agent why and what to run next.
This is not a fake winner-take-all chart. Coding agents, navigation, semantic retrieval, and static analysis are excellent at their advertised jobs. PX sits beside them and turns their evidence into an explicit route, coverage horizon, recovery contract, and reusable receipt.
Industry capability demonstration
Choose a category. The live path shows what the industry surface contributes, where agent certainty can outrun evidence, and what PX adds without replacing the underlying tool.
The final answer can still hide which retrieval path ran, where context ended, and whether a claim exceeded the evidence.
PX makes the route, target identity, coverage, weaknesses, and next recovery step machine-readable.The agent does not choose from a command wall. PX frames the question, arbitrates the lanes, preserves the evidence horizon, and returns one bounded continuation with a receipt another agent can reuse.
Project-aware AI context and semantic codebase indexing.
Retrieved context is optimized for answering, without an independent evidence horizon or typed zero-result meaning.
Expose route choice, coverage, weaknesses, and recovery before context becomes a conclusion.
Repository research, plans, code changes, tests, commits, and pull-request workflows.
Autonomous execution is not the same as a reusable proof of why those files were inspected.
Give any agent the same route receipt, evidence horizon, brief, and verification packet.
Large-scale search and precise cross-repository code navigation; SCIP can provide compiler-derived definitions and references.
Navigation is one evidence lane. Agent tasks also cross docs, dirty bytes, SQL, HTTP, runtime, planning, and post-edit verification.
Compose those lanes and never call a partial index complete.
Queryable program models, data and control flow, security rules, reachability, and vulnerability triage.
Purpose-built analyzers are not a general agent entry, recovery, workflow, or evidence-lineage layer.
Route work to the analyzer, retain its proof class, and connect findings to impact, docs, tests, and closeout.
Question framing, deterministic route arbitration, cross-lane composition, current-tree labels, typed recovery, and receipt-bearing workflows.
PX does not write the patch, replace compilers, or magically prove runtime behavior. It is only as strong as its connected sources and courts.
Make every agent faster to ground, harder to fool with partial evidence, and able to hand the next agent a reusable proof packet.
Competitive honesty: PX combines ideas that exist elsewhere: code graphs, semantic retrieval, static analysis, documentation search, and agent tooling. The differentiator is the product contract around them. Independent head-to-head productivity and accuracy benchmarks remain unknown; this page does not manufacture them.
Captured PX question graph / Delphi evaluation
Fourteen real repository questions stay visible at once. PX connects each question to the route family that understood it, then to the bounded continuation handed to the agent. Click any question to inspect its exact path.
Incidental discovery / enterprise product contract
While answering one question, PX already encounters evidence around it: duplicated capability, graph-unreachable code, contradictory facts, ambiguous targets, and architectural drift. Most tools discard that peripheral evidence. The target contract preserves it without hijacking the task.
$ px ai --space company "who calls prediction_confidence?" --jsonThe agent receives the answer it asked for, its evidence horizon, and the next safe command.
PX does not guess or silently merge them. The proposed capture lane records the collision, confidence, locations, and cleanup recommendation.
Reuse evidence the ordinary query already fetched. No separate audit crawl.
Duplication, unreachable code, contradictory evidence, or architectural drift.
Interrupt only when urgency is high and the finding touches the active task.
Append an evidence-linked observation to the truth layer. Never auto-promote it to truth.
Batch lower-urgency findings into closeout so the next agent inherits the signal.
The enterprise payoff is not another warning feed. It is model-independent organizational learning: every agent can leave behind a compact, attributable observation that future work can test, supersede, or resolve.
{
"finding_type": "ambiguous_target_collision",
"locations": ["router.rs:181", "forecast.rs:64"],
"confidence": 0.94,
"first_observed_by": {
"command": "px ai",
"question": "who calls prediction_confidence?"
},
"urgency": "medium",
"action": "batch_at_closeout",
"recommendation": "Resolve identity before impact analysis",
"mutation_performed": false
}
95.7% measured context reduction across 5 real repository questions. Each line compares broad source-file context with the compact PX evidence payload.
The complete agent surface
Agents start with px ask --json. PX composes from 116 direct command surfaces plus validated continuations for exact graph work, docs authority, planning, verification, index truth, languages, runtime evidence, and fleet operations.
Natural-language router with evidence receipts and one bounded next step.
px ask --space myrepo --json "What breaks if CheckoutService changes?"PX / agentic code intelligence
Start with px ask --json "your question". Give every model the map, meaning, evidence horizon, and next move before it spends a single reasoning token wandering.
PROBAGI / SOURCE PROOF
The generator validated 229 continuation links with zero invalid links.
px-command-catalog-0.1.175
Browser verification confirmed fourteen interactive question selectors, clickable path changes, moving evidence particles, and no page errors.
px-living-question-graph-court-20260728
The scroll court held the terminal at zero lines for four seconds above the fold, then observed 27 progressive lines, two planning turns, two Claude responses, one typed command, and zero page errors after the terminal entered view.
px-progressive-cli-court-20260728
The capture script recorded 28 successful PX 0.1.175 invocations at repository HEAD a920fd9a and redacted local paths.
px-agent-feature-captures-v1
The page explicitly labels the current evidence substrate, held automatic-capture target, read-only observation rule, urgency gate, and no-extra-query boundary.
px-incidental-finding-ultra-20260728
Every economic assumption is visible and editable in the page.
enterprise-context-economics-scenario
Five selectable cases expose exact before and after token counts, reduction, intent, result status, coverage, source files, and measured PX latency.
px-context-economics-benchmark-v1
The page and source identify embedding-backed semantic retrieval, intent reranking, and the rule that similarity never silently becomes proof.
px-neural-meaning-lane-20260728
Current command contracts expose docs bootstrap, atlas, health, coverage, orphans, context, validation, claims, reconciliation, filesystem checks, and generated ASCII, Mermaid, JSON, and HTML outputs under _docs/px/architecture.
px-docs-atlas-command-contract-20260728