rea/docs/verification/native-function-summary-2026-10-23.md

Native function summary verification — 2026-10-23

Behavior and source

MCP analyze_function defaults to a compact derived summary. The complete function Evidence stays in the session and can be selected by the returned parent_evidence_id through inspect_analysis_view. detail: "complete" keeps the historical full reply. Delivery detail stays outside provider inputs and parent Evidence identity. This MCP default is a breaking change.

CLI rea function still defaults to complete output. Opting into summary requires an exclusive --evidence_output bundle destination. The authenticated bundle can be paged by an absolute evidence-file source and explicit record ID after the provider closes. Plain records also honor an explicit ID.

The implementation reuses the existing JSON authentication/headroom owner and one bundle-authentication function for import and file views. It reads one file once per request. Ambiguous-bundle guidance retains complete IDs and bounds JSON-encoded reference previews to 2 KiB, omitting oversized operation labels. Native summaries inherit function locations and relationship coverage; omitted or unavailable facts stay unknown rather than becoming zero.

Measured source:

• Before: fb6c88d518ab9b84ed396cf82dcb1c28a608c501.
• After: 4e3d930ada93014efa43610265b58307de49a687.
• Harness SHA-256: d67b23db21024361b34f796add1ff91bf3fbcda1991ea9e369ef8bcf64c71ec4.
• Linux x86-64; pinned Node 24.22.1 / npm 11.16.0; bring-your-own Ghidra 12.1.4 and JDK 21.0.12.1+1. No target execution.
• Target SHA-256: 9c49d2eef9c2083f6f66f38de4ed0ee1bac17ca5b73dc8f3a90d9ba62a8d1d5d. Its 3,000-branch C source is the timeout verifier fixture, SHA-256 f99074876faae8e18e45265dab007b6fcc396fe909d57efab72adc8925a960c7. Known entries 0x1010f9 and 0x128ed6 apply only to this exact artifact.

Completed public MCP comparison

One fresh production stdio MCP client discovers all 221 tools, opens the target, then obtains two answers: the small function's identity, availability/counts and complete pseudocode; and the large function's identity, availability, assembly/ block counts and first 20 assembly rows. It repeats both reads in the same session, performs one matched complete-small compatibility read and closes.

Before, each default function reply carries the full dossier. After, each answer uses the default summary and one retained-Evidence page. Both lanes use a one-second native decompile limit, an explicit 600-second SDK tool deadline and matching 128 MiB server/client receive budgets. Selected facts match between implementations and rounds. The large function has 45,007 assembly rows and 9,001 blocks with pseudocode unavailable; the small pseudocode is 64 bytes, SHA-256 c947ff22cab1ffef9b962b12d82e12d0e5f5e2a6c5634a2ed5ea3c353e1d5807. Each run retained the same observed JVM start identity through its tasks and completed cleanup with an unchanged target digest.

Whole benchmark	Before	After
Estimated exchange tokens	44,217,047	1,033,322
Elapsed seconds	237.715	128.600
Waited process-family CPU seconds	211.656	123.083
Sampled family RSS MiB	3951.7	2827.8
Catalog/tool exchanges	8	12

Estimated context falls 97.66% for this completed task, including the larger catalog and four extra page calls. The catalog alone grows from 982,457 to 985,662 estimated tokens. The complete-small compatibility read remains included in both totals.

MCP phase elapsed seconds	Before	After
First small function, including deferred JVM startup	80.3747	69.0745
First large function, with the JVM already warm	57.2394	19.8205
Repeated small function	0.0534	0.0601
Repeated large function	56.4148	19.7744

The repeated small task adds a page call and takes about 6.7 ms longer in this pair. Large partial dossiers remain live and retry native analysis; healthy small-function snapshot reuse remains available. This change reduces agent output and its serialization/transport work while still producing and retaining the complete parent dossier.

Tokens are tokenx 1.3.0 estimates over the actual catalog, arguments and whole SDK results, including both text and structured projections. Protocol framing, handshake, prompts, client rendering/replay and actual model usage are unmeasured. The whole benchmark includes token estimation, hashing, resource sampling and exchange-file writes. Summed measured MCP phases exclude the subsequent counting/writing overhead: 217.327 → 127.538 seconds. The whole-benchmark timings must not be presented as pure provider analysis time.

Node heaps were 2 GiB, Ghidra's heap 1 GiB, affinity CPUs 0/1 and Java active processor count two. The observed-family early-stop guard allowed 6 GiB RSS, 600 seconds CPU and 900 seconds elapsed; neither completed run hit a guard. RSS was sampled every 50 ms through recursive /proc observations. Shared pages can be double-counted and brief/unobserved child transitions can be missed; this is not a kernel cgroup peak or hard family memory limit. Waited CPU includes reaped descendants. Baseline ran first, OS caches were not flushed and the host was shared with other work. This is one paired investigation, with cold/result-cache distinctions above, rather than a universal speed claim.

An earlier baseline attempt used a 240-second SDK deadline and began with the large function. It timed out, then closed successfully; the whole attempt took 280.628 seconds and 219.562 seconds waited CPU, with no resource-guard stop. Its failed receipt/exchanges remain separate. It is excluded from completed-task comparisons. The final pair uses the same longer deadline and task order on both implementations.

Public CLI and retained files

The real default CLI command returned a complete small-function dossier with the expected pseudocode (70.627 s elapsed, 70.034 s CPU). The first verification harness assumed bare Evidence and failed to unwrap the successful public ok/data/meta envelope. After correcting the harness, its completed receipt and output were verified and reused without another provider analysis.

The summary CLI command analyzed the large function and wrote a 29,141,012-byte bundle. It returned an 8,499-byte summary envelope in 78.163 seconds, with 94.196 seconds CPU and 2,381.2 MiB sampled family RSS. Two fresh CLI processes, with the Ghidra installation path unavailable, authenticated that bundle and returned its first and final 20 assembly rows. Coverage reports 45,007 total rows, correct continuation and exhaustion. The target digest remained unchanged.

Those independent file-page commands took 17.218 / 18.175 seconds, 15.150 / 17.015 seconds CPU and about 666 / 669 MiB RSS. Each fresh command reads and authenticates the complete bundle; this is not sublinear disk access or a cross-process cache. Complete evidence remains available at the caller's file.

Validation and limits

• Full npm run check passed after integration with main a062529bb2aa064be37e82df39c59c4fa798fe50 at source b7c3f0cd79c4a02b135da56fff06de08dd456ff7. docs:check also passed.
• The latest focused contract/progress/snapshot/local-path run passed 139 tests; the measured integration's selected run passed 136 and an earlier broader CLI-selector/contract run passed 169. Overlaps are not summed. The earlier domain/Evidence run had 87 passes and one legacy progress-fixture failure. The explicit-complete fixture and new summary's increasing notifications subsequently passed in a separate eight-case MCP run.
• Actual SDK/schema checks passed all 15 cases after regenerating ignored test catalog metadata. The initial run had 14 passes and one stale-catalog equality failure; its receipt is retained.
• The production CLI/MCP observations above use real Linux Ghidra. Hopper, IDA, Kuna, macOS and Windows behavior is covered only to the extent of the selected source/boundary tests, not new real-provider verification.
• Earlier source integration fixes and later main changes were not included in every historical test/performance pin. The measured pair remains tied to the explicit source commits above.

Raw exchanges, protocol observations, control failures, scripts and resource receipts are retained under /tmp/rea-maintenance-20261023/5660/. Provider projects, generated binaries, build outputs and local planning files are not committed.