# Flattened devicetree inspection `inspect_devicetree` / `rea inspect-devicetree` reads a local DTB without an engine, target execution or external tools. It supports FDT versions 16 and 17; other versions return `unsupported_target`. Corrupt recognized blobs return `invalid_request`, and exhausted budgets return `resource_constraint` without partial success. [DTSpec ยง5.4.1](https://devicetree-specification.readthedocs.io/en/stable/flattened-format.html#lexical-structure) requires zero padding after structure node names and property values. For example, [libfdt's node rename](https://github.com/dgibson/dtc/blob/c954f9dd835764ec69d3d9afb467c4541b47b66d/libfdt/fdt_rw.c#L253-L273) can leave bytes from the old name after the new terminator. REA reports the count and first file offset of nonzero padding while retaining the tree's observations when all fields fit. ```bash rea inspect-devicetree /artifacts/board.dtb --json rea inspect-devicetree /artifacts/board.dtb --detail summary --evidence_output /artifacts/board-evidence.json --json ``` The complete Evidence includes the artifact SHA-256, validated header fields, 64-bit memory reservations as exact hexadecimal strings, and a flat node tree with explicit parent paths. Each property retains its file offset, byte length, and up to 256 raw bytes as hex. `truncated` qualifies the preview only; parsing coverage includes the entire declared tree. Version 16 has no declared structure size, so `header.size_dt_struct` is null. Derived fields carry `confidence: derived`: standard ASCII string lists, scalar phandles and cell counts, and `reg` tuples using the parent's `#address-cells` and `#size-cells` (defaults 2 and 1). Zero-cell components are null (omitted), distinct from an encoded zero. Cell counts above four, invalid counts and incomplete tuples remain undecoded with a limitation. Addresses are bus-local; no `ranges` translation is attempted. Invalid or ambiguous strings remain raw, rather than being guessed. The summary is labeled derived and includes root compatible/model, chosen bootargs/stdout-path and aliases. An undecodable alias retains its name with a null path; it is not silently treated as absent. Single-cell `interrupt-parent`, `remote-endpoint` and `msi-parent` references resolve through phandle declarations. Missing or duplicate targets have null target paths. A node whose `phandle` and `linux,phandle` disagree adds a limitation, and references to either value stay unresolved. A present `phandle` that is not one cell also adds a limitation; `linux,phandle` cannot resolve that node when the canonical declaration is malformed. Binding-specific reference arrays and overlays are not resolved; an empty reference list does not prove absence of hardware links. MCP `detail: summary` retains complete Evidence in the session. Follow its `parent_evidence_id` with `inspect_analysis_view`: ```json { "source": { "kind": "retained-evidence", "evidence_id": "ev_" }, "view": { "kind": "devicetree", "collection": "properties", "path": "/serial@1000", "offset": 0, "limit": 32 } } ``` Collections are `nodes` (identity and property counts), `properties` (requires an exact node path), `reservations`, and `references`. Pages use retained parsed facts and do not reopen the file. Each page carries fixed-size context (artifact, header, coverage and counts) rather than the summary strings, so a large `bootargs` value does not repeat on every page. Source indices are scoped to the selected collection and parent Evidence; property indices are scoped to their node. | Input budget | Default | Ceiling | | ---------------- | ------: | ------: | | `max_file_bytes` | 4 MiB | 16 MiB | | `max_nodes` | 8,192 | 65,536 | | `max_properties` | 32,768 | 262,144 | | `max_depth` | 64 | 256 | Decoded names, paths, values and row overhead share a fixed 32 MiB representation allowance, charged before expansion. It has no runtime override. A parser budget refusal names the allowance in `details.reported_limits.boundary` and its effective bound in `details.reported_limits.maximum`. For example, `--max-nodes 1` reports `{ "boundary": "max_nodes", "maximum": 1 }` if the blob contains more than one node. The reservation count also uses `max_properties`. CLI flags use kebab case, for example `--max-nodes 16384`. View pages contain at most 256 rows. ## Verification Synthetic tests cover v16/v17, malformed headers/blocks/names/tokens, resource limits, derived values, and CLI/MCP delivery including retained pages after removing the input. The optional real lane requires **dtc 1.8.1**: ```bash npm run build:cached node scripts/verify/firmware/devicetree.mjs ``` It compiles a DTS with reserved-memory, chosen, aliases and phandles, compares REA's node paths against `dtc -I dtb -O dts`, and checks the derived report. `REA_DTC_PATH` selects the verification executable; it is not runtime setup. The dtc 1.8.1 lane ran on macOS arm64 with Node 26.5.0. Its seven node paths matched the decompiled tree. The report included `model: "REA DTB verification"`, `bootargs: "console=ttyS0"`, alias `serial0 -> /serial@1000`, reservation `0x80000000 + 0x1000`, and a resolved reference from `/consumer` to `/serial@1000`. The deliberately minimal remote-endpoint fixture produces dtc graph-binding warnings; this lane verifies FDT structure and selected property derivations, not hardware-binding validity. A repeat run of that complete compile/decompile/CLI verification took 2.13 s elapsed, 1.52 s user CPU and 0.27 s system CPU under `/usr/bin/time -l`. This is a small fixture with warm filesystem caches, not a cold-start or large-board benchmark. Peak process-family memory and whole-workflow token usage were not measured; the sandbox denied the memory-reporting system query. No pre-feature equivalent REA workflow exists for comparison. The pinned Node 24.22.1/npm 11.16.0 toolchain was unavailable; dependency installation used the local cache with Node 26.5.0/npm 12.2.0 instead.