deepbom 1.96.14 → 1.97.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "deepbom",
-  "version": "1.96.14",
+  "version": "1.97.0",
   "mcpName": "io.github.JunHwan-Kwon/deepbom",
   "description": "Local multi-format deployment-artifact analysis for on-device AI models",
   "type": "module",
@@ -10,6 +10,7 @@
   "files": [
     "bin/",
     "pkg/",
+    "skills/deepbom/",
     "README.md",
     "LICENSE"
   ],

package/pkg/release-manifest.json

@@ -1,8 +1,8 @@
 {
   "schema": "deepbom.npm_release.v1",
-  "version": "1.96.14",
+  "version": "1.97.0",
   "source": {
-    "git_commit": "e452811d7a460fd0f67549e28b41fb9e03d79c48",
+    "git_commit": "95b2218267ba6eeaf54945344aa3d7c945ef7fdd",
     "git_state": "clean",
     "distribution": "public_channel"
   },
@@ -18,7 +18,7 @@
     "spdx": "Apache-2.0",
     "file": "LICENSE"
   },
-  "public_bundle_input_count": 155,
+  "public_bundle_input_count": 157,
   "supported_inputs": [
     "tflite",
     "onnx",

package/skills/deepbom/SKILL.md

@@ -0,0 +1,83 @@
+---
+name: deepbom
+description: Statically audit serialized AI deployment artifacts with the local DEEPBOM CLI. Use for TFLite, ONNX, Core ML, ExecuTorch, GGUF, or SafeTensors deployment review; MAC and symbolic-shape coverage; quantization contracts; memory feasibility; accelerator eligibility; artifact diff; graph export; SARIF; or CycloneDX 1.7 ML-BOM generation. Do not use for training checkpoints, measured latency, task accuracy, or claims about actual runtime placement without imported evidence.
+---
+
+# Audit deployment artifacts with DEEPBOM
+
+Run DEEPBOM locally. Never upload artifact bytes or invent a hosted DEEPBOM endpoint.
+
+## Start with discovery
+
+Use the release pinned by this skill:
+
+```bash
+npx -y deepbom@1.97.0 capabilities --format agent-json
+```
+
+Treat that machine document as authoritative for commands, supported inputs, scan modes, targets, outputs, and exit codes. Read `references/capability-selection.md` when deciding whether the task belongs to DEEPBOM.
+
+## Audit
+
+Start with the bounded human result:
+
+```bash
+npx -y deepbom@1.97.0 audit "./model.onnx" --summary
+```
+
+Then request only the evidence needed by the question:
+
+```bash
+npx -y deepbom@1.97.0 audit "./model.onnx" --section quantization --json
+npx -y deepbom@1.97.0 audit "./model.onnx" --pointer /mac_assessment --json
+npx -y deepbom@1.97.0 audit "./model.onnx" --output-format envelope
+```
+
+For a large GGUF or SafeTensors artifact, use `--scan structure` for inventory questions. Use `integrity` or `full` only when payload evidence is necessary.
+
+For a concise GGUF tensor table without nested numerical-integrity ledgers, use the bounded projection (it defaults to a structure scan):
+
+```bash
+npx -y deepbom@1.97.0 gguf "./model.gguf" --tensors
+npx -y deepbom@1.97.0 gguf "./model.gguf" --tensors --compact
+npx -y deepbom@1.97.0 gguf "./model.gguf" --tensors --compact --tensor-offset 100 --tensor-limit 100
+```
+
+Before relying on a newly fetched package, run:
+
+```bash
+npx -y deepbom@1.97.0 self-test --compact
+```
+
+## Interpret the result
+
+Keep `artifact_defect`, `caution`, and `evidence_gap` separate. An evidence gap is not a defect. Preserve `NOT_APPLICABLE`, `NOT_ASSESSABLE`, and `NOT_ASSESSED_YET` instead of converting them to zero or absence.
+
+Read `references/evidence-semantics.md` before summarizing findings or target projections. Report at least:
+
+- artifact filename and SHA-256
+- analyzer version
+- artifact format
+- target and binding source when target-dependent evidence is used
+- assessed coverage and unresolved evidence
+- defects, cautions, and evidence gaps as separate groups
+- the exact reproduction command
+
+Never infer actual accelerator assignment, latency, energy, thermal behavior, task accuracy, clinical validity, or device fit from static output alone.
+
+## Compare or explain
+
+```bash
+npx -y deepbom@1.97.0 diff "./baseline.onnx" "./candidate.onnx" --summary
+npx -y deepbom@1.97.0 diff "./baseline.gguf" "./candidate.gguf" --tensors --render markdown
+npx -y deepbom@1.97.0 verify "./model.onnx" --bom "./supplied.cdx.json" --render markdown
+npx -y deepbom@1.97.0 contract capture "./model.onnx" -o "./baseline.interface-contract.json"
+npx -y deepbom@1.97.0 explain-rule <rule-id> --json
+npx -y deepbom@1.97.0 graph "./model.onnx" --view structure --format svg -o graph.svg
+```
+
+Use matching serialized deployment formats for diff. Treat `verify --bom` as a reconciliation of the selected component with artifact-observable facts, not a complete BOM or compliance verdict. An automatically captured contract is an artifact-derived baseline until it is separately reviewed and approved. Do not deserialize `.pth`, `.pt`, or `.h5`; bind a conversion receipt to the deployed artifact instead.
+
+## Handle failure
+
+Read `references/failure-recovery.md` when an invocation fails, exceeds a resource boundary, or returns exit code 2 or 3. Do not weaken a size, path, hash, or evidence boundary merely to force a complete result.

package/skills/deepbom/agents/openai.yaml

@@ -0,0 +1,7 @@
+interface:
+  display_name: "DEEPBOM Artifact Audit"
+  short_description: "Audit serialized AI deployment artifacts locally"
+  default_prompt: "Statically audit this deployment artifact with DEEPBOM and report bounded evidence, unresolved gaps, and the reproduction command."
+
+policy:
+  allow_implicit_invocation: true

package/skills/deepbom/references/capability-selection.md

@@ -0,0 +1,16 @@
+# Capability selection
+
+Use DEEPBOM when the subject is a serialized deployment artifact and the question can be answered from artifact bytes, an explicit target profile, or an imported evidence document.
+
+| Task | Use DEEPBOM | Boundary |
+| --- | --- | --- |
+| TFLite, ONNX, Core ML, or ExecuTorch graph and tensor audit | Yes | Static artifact evidence only |
+| GGUF or SafeTensors storage, quantization, architecture, or memory lower bound | Yes | No executable graph is invented |
+| MAC and symbolic shape coverage | Yes | Preserve partial or symbolic status |
+| Quantization contract review | Yes | Do not infer task accuracy |
+| Accelerator eligibility or selected-build evidence | Yes | Eligibility is not observed assignment |
+| Artifact version diff | Yes | Inputs must use the same supported format |
+| PyTorch or HDF5 training checkpoint execution | No | Convert safely and bind a conversion receipt |
+| Measured latency, energy, temperature, or accuracy | No | Import runtime evidence or use the appropriate benchmark system |
+
+Prefer `summary` first. Request an envelope for cross-format automation, format-specific JSON for a narrow field, SARIF for CI findings, or CycloneDX 1.7 for supply-chain exchange.

package/skills/deepbom/references/evidence-semantics.md

@@ -0,0 +1,20 @@
+# Evidence semantics
+
+Preserve DEEPBOM's evidence class and applicability state in the answer.
+
+- `OBSERVED` describes serialized artifact facts.
+- `SOURCE_BACKED` identifies a source-pinned rule or compatibility statement.
+- `DERIVED` is deterministically calculated from declared inputs.
+- `PREDICTED` and `ESTIMATED` are model-dependent, not runtime observations.
+- `MEASURED` is valid only when an imported measurement is identity-bound.
+- `NOT_APPLICABLE` means the domain does not apply to this artifact.
+- `NOT_ASSESSABLE` means required evidence is absent or insufficient.
+- `NOT_ASSESSED_YET` means analysis has not been requested or completed.
+
+Finding kinds are independent from evidence classes:
+
+- `artifact_defect`: a contradiction or defect found in the artifact.
+- `caution`: a bounded condition that merits review but is not a defect.
+- `evidence_gap`: a claim the artifact cannot settle alone.
+
+When a target is present, report whether its binding source is a default assumption, explicit ID, or profile file. `host_observed: false` means the host was not measured.

package/skills/deepbom/references/failure-recovery.md

@@ -0,0 +1,9 @@
+# Failure recovery
+
+1. Exit `1`: report invocation, input, artifact, or output failure. Read structured stderr with `--error-format json` before changing the command.
+2. Exit `2`: report a policy or verification block. The evidence document may still be valid and should be retained.
+3. Exit `3`: report an incomplete verification binding. Do not call it a tool crash or an artifact defect.
+4. For a large GGUF or SafeTensors file, retry with `--scan structure` when the question does not require payload integrity.
+5. For remote input, require an immutable identity: a full Hugging Face commit, GCS object generation, or HTTPS SHA-256 fragment.
+6. Do not raise download, response, or memory limits without explaining the cost and obtaining user approval.
+7. If no shell, local MCP, or local file access exists, explain that actual analysis cannot run in that client and provide the pinned `npx` command.

package/skills/deepbom/scripts/verify-deepbom.mjs

@@ -0,0 +1,19 @@
+#!/usr/bin/env node
+
+import { spawnSync } from "node:child_process";
+
+const VERSION = "1.97.0";
+const executable = process.platform === "win32" ? process.env.ComSpec || "cmd.exe" : "npx";
+const args = process.platform === "win32"
+  ? ["/d", "/s", "/c", "npx.cmd", "-y", `deepbom@${VERSION}`, "self-test", "--compact"]
+  : ["-y", `deepbom@${VERSION}`, "self-test", "--compact"];
+const result = spawnSync(executable, args, { encoding: "utf8", stdio: ["ignore", "pipe", "pipe"] });
+if (result.status !== 0) {
+  process.stderr.write(result.stderr || result.stdout || `DEEPBOM self-test exited ${result.status}.\n`);
+  process.exit(result.status || 1);
+}
+const document = JSON.parse(result.stdout);
+if (document.schema !== "deepbom.cli_self_test.v1" || document.status !== "pass") {
+  throw new Error("DEEPBOM self-test did not return a passing deepbom.cli_self_test.v1 document.");
+}
+process.stdout.write(`${JSON.stringify(document)}\n`);