pi-rea

REA (Reverse Engineer Anything) as a Pi-native extension: REA's tool interface, its session and Evidence kernel, and a reverse-engineering workflow skill, reimplemented with Pi's own mechanisms.

pi-rea is an independent implementation. It contains no REA code, depends on no REA package, and never calls the REA CLI or MCP server. Tool names, parameter sets and observed result shapes are kept for compatibility; every description, schema and line of code here is this project's own (see Independence).

This package advertises the rea-agents 6.8.0 agent-facing catalog (229 tools). 66 of them are implemented handlers (the session/Evidence kernel, composed workflows and absorbed static analyzers). The other 163 are registered fail-closed provider or capability refuses that match measured REA 6.8.0 missing-provider envelopes. binary_session reports that split, and schemas/tool-contracts.json records it as counts.implemented: 66, counts.unavailable: 163. Catalog size is not behavioral parity.

Measured DiffCase parity lives under parity/pi-rea/: 257 frozen cases are exact vs npx rea-agents 6.8.0 envelopes on the Darwin host that ran the harness, covering all 229 catalog tools with at least one DiffCase. This includes binary_session, open_binary, close_binary, all session kernel tools, inspect_pcap_client, inspect_pcap_capture, batch_decompile, and the PE import budget and boundary paths. That smoke includes fixture-scoped semantic_graph goldens for known roots and harness-aligned refuse prose for missing providers. Measured residual envelope divergences are listed in parity/pi-rea/parity-ledger.md.

What works today

An unimplemented call answers with the canonical error envelope. It never returns a substitute result, so a coverage gap can never be mistaken for a finding about a target. parity/pi-rea/ records the baseline and decisions.

Install

pi install npm:@justjoshn/pi-rea

Or, from a clone of this repository:

pi install ./extensions/pi-rea

During development, load the entry directly:

pi -e ./extensions/pi-rea/src/index.ts

The extension registers its 66 implemented handlers with direct exposure and the 163 refuse handlers with deferred, so tool_search carries the catalog instead of filling the prompt with tools that cannot run. A refuse is a coverage gap, never a finding about the target.

Remove it with pi remove ./extensions/pi-rea.

Regenerating the contract

cd extensions/pi-rea
bun run generate:contracts

scripts/generate-contracts.ts rebuilds schemas/tool-contracts.json from parity/pi-rea/rea-6.8.0-tools-list.json and marks kernel tools from schemas/kernel-manifest.json, which test/contract.test.ts checks against the registered handler table in src/registry.ts. Every parameter the oracle publishes survives regeneration; the script only adds documented field detail where the oracle schema is a bare object. Do not hand-edit the generated catalog.

Host requirements

pi-rea targets Node 22.19 or newer (engines). The tools below need more on the host. A missing tool answers with a refuse, never a substitute result.

inspect_pe_*, inspect_macho and the other binary-format inspectors parse in process and behave the same on macOS and Linux.

The REA_* variables below are read from the launch environment. The first group tunes implemented behavior.

• REA_MCP_MAX_RESPONSE_BYTES, REA_ASAR_MAX_HEADER_BYTES, REA_REGEX_MATCHING_DEADLINE_MS, REA_REGEX_WORKER_HEAP_REQUEST_MIB and REA_CAPTURE_DECODER_OLD_GENERATION_MIB set the response budget and the analysis limits that inspect_analysis_settings and update_analysis_settings report.
• REA_GHIDRA_STARTUP_TIMEOUT_MS, REA_HOPPER_STARTUP_TIMEOUT_MS, REA_GHIDRA_DECOMPILE_TIMEOUT_SECONDS and REA_JADX_OPERATION_TIMEOUT_MS are reported by the same settings tools. They have no effect until a provider is bridged.

The provider variables REA_IDA_MCP_CONFIG, REA_KUNA_COMMAND, REA_ADB_PATH, REA_JEB_MCP_URL, REA_RIZIN_COMMAND, REA_APKTOOL_COMMAND, REA_BINWALK_COMMAND, REA_UNBLOB_COMMAND, REA_WAKARU_COMMAND, REA_JADX_MCP_JAR, REA_WABT_BIN_DIRECTORY, REA_SIGROK_COMMAND, REA_TSHARK_PRLIMIT_COMMAND and REA_SIGROK_PRLIMIT_COMMAND appear only in refuse remediation text. REA_IDA_MCP_CONFIG and REA_KUNA_COMMAND also change the refuse diagnostics. Setting any of them does not enable a tool, because no provider bridge exists.

Use

Ask for an investigation that depends on a shipped artifact rather than source:

Open /usr/local/bin/thing, tell me its format, architecture and digest, and note
what you could not establish about it.

Typical kernel session:

1. binary_session: what is available, what is open, what provider exists.
2. open_binary: bind a file and read its identity (format, kind, architecture, SHA-256).
3. record_unknown: keep every open question explicit, with the authority and confidence a resolution would need.
4. get_evidence_bundle / export_evidence_bundle: read or publish the session as canonical JSON. close_binary clears the session, so export first.
5. close_binary: drain, clear and finish.

/rea prints the current status: open target, availability counts and retained Evidence.

Independence

• No REA dependency: no npm package, no vendored code, no CLI, no MCP server, no network. The only child processes are host tools such as objdump, tshark and swift-demangle.
• No REA bytes: the shipped contract transcribes interface facts only (tool names, parameter names, types, bounds, required sets, flags, effect metadata, observed result shapes). The reference's prose, titles, examples, defaults, per-mode remediation text and output schemas are excluded, and every description is written here.
• The workflow skill is this project's own text.

extensions/pi-rea/test/independence.test.ts enforces this over the shipped src, schemas and skills directories and over package.json. It does not scan README.md or THIRD_PARTY_NOTICES.md, which carry no code.

Verification

cd extensions/pi-rea
bun run test
bun run typecheck

bun run typecheck exits 0. bun run test passes 1510 tests in 44 files on Ubuntu with Node 24 (one SQLite test is skipped when the sqlite3 CLI is absent, so the count can differ by host). bun run test:coverage meets its configured 80% thresholds: the measured suite covers 96.84% of statements, 90.41% of branches, 95.75% of functions and 97.76% of lines. Coverage instrumentation slows test/pi-loading.test.ts past its 5 s default timeout on a cold run, so rerun with --testTimeout=60000 if only that test fails. Behavior tests drive the absorbed analyzers through their public functions with synthetic bytes and real fixtures; the successful paths are also exercised through the DiffCase envelopes in parity/pi-rea/.

Oracle captures, design notes and decisions live under parity/pi-rea/.

State and lifetime

The Evidence ledger and the unknown ledger follow the active Pi branch. Each Evidence-producing tool mirrors its record into the tool result's details and appends a rea-evidence entry with pi.appendEntry; record_unknown and update_unknown append rea-unknown entries alongside their mutation Evidence. session_start rebuilds both ledgers from ctx.sessionManager.getBranch(), so a /fork or /resume sees exactly the records its branch holds, and an abandoned branch cannot leak into the live one. close_binary still clears both ledgers in the running session. export_evidence_bundle publishes them as canonical JSON, and import_evidence_bundle restores them; identifiers stay stable across the round trip.

Limits

• Nothing here starts an analysis engine: the deep-analysis surface works only once a provider bridge exists.
• Pi owns the agent loop, tool lifecycle, sessions and permissions; pi-rea only registers tools and session state through the public extension API.
• Evidence identifiers are content-addressed with this project's projection, so a bundle exported by REA itself is refused by the identity check rather than silently re-identified (KERN-044).
• Provider-gated and engine-free-gap tools are declared but refuse closed with measured unavailable envelopes; never treat those refuses as target findings.
• An oversized result is truncated, not rejected. The 10 MiB budget stays as the truncation threshold, and an oversized payload is replaced by {truncated, response_budget_bytes, required_bytes, guidance}. This is a deliberate divergence from REA, whose transport rejects the frame: Pi's extension docs require truncating large model-facing results and telling the model where to read the rest, and Pi has no frame limit that needs the refusal. Evidence results keep the full record in the tool result's details and the guidance points at export_evidence_bundle; other payloads are not retained, so their guidance asks for a narrower request. REA_MCP_MAX_RESPONSE_BYTES raises the threshold. It must be a decimal integer of at least 10485760.

License

MIT. See THIRD_PARTY_NOTICES.md.