Android manifest components verification, 2026-10-24¶

Scope and revisions¶

PR #5025 adds a static component inventory to inspect_android_package, using the namespace-aware decoded manifest already retained by the JADX adapter. Activities, aliases, services, receivers and providers retain declarations, qualified class names where established, export declarations/defaults, permission provenance, intent actions/categories, service types, authorities and alias targets. These facts do not establish runtime registration, class existence, merged-manifest acceptance or APK authenticity.

The maintenance comparison pins baseline e3802c16491b772118ccdfa195c6b51c3a337bdd (contributor head c822b00c4755e584ee081eb3e7587735b630e249 integrated with main) and candidate fde65b1d47c00c189e003c54ad2a2c1fc93b443a. Both contain the feature; this comparison checks the maintenance changes rather than estimating the cost of answering the same question without the feature. Later main integration is outside these measurements.

Corrections and automated coverage¶

Alias target matching indexes valid spellings once for preceding activities. It now adds a bare abbreviation only for a class directly inside the manifest package. sub.Target is already a fully qualified name and cannot match com.example.sub.Target; .sub.Target and the complete qualified name can. The contributor had already removed repeated package expansion in c822b00c; that performance correction is not attributed to this follow-up. The existing 900 KiB package / 511 alias regression remains covered.

Older Evidence without this inventory now receives a fresh unavailable inventory per parse. Mutating one parsed default cannot affect another parse. Before the corrections, the new identity/default cases had 6 passes and 3 failures (two false alias matches and one shared default). Afterward all 50 component cases passed. The broader selected component/JADX/SDK-summary set passed all 83 cases, and npm run check:fast passed on the measured candidate. After integrating main 21a20306b4887ff8e6b614b6ae34a387cfbdf1d8 in cdc2829131045d85692cf4bbf0acb12f36f65af7, the same 83 cases passed again. check:fast, docs:check and test discovery (2,333 files, each once) also passed.

The parser bounds declarations and retained name/permission/authority copies before expansion. Truncation, unknown export rules, unresolved names and unsupported permission overrides remain explicit. The existing Android summary includes the structured inventory and retains the complete manifest for paging; no second inventory parser or new view implementation was added.

Real engine and public APK¶

Linux x86-64, Node 24.22.1 / npm 11.16.0, existing full OpenJDK 17.0.20.1, JADX headless MCP 0.7.1, one analysis worker, 512 MiB JVM heap and ActiveProcessorCount=2. The task process family was restricted to CPUs 0/1. The APK was read as an inert artifact and was never installed or executed.

• Appium ApiDemos 6.0.18, source revision 3e716ee04ef3390284bfab761e28da8c15c08143, Apache-2.0: ApiDemos-debug.apk, 6,521,217 bytes, SHA-256 a9eecf37b26cd084855c530db81c2bb1b91f4c1b095a04f47aa7c20e2791f686.
• JADX headless MCP 0.7.1, source revision 5844800a486d2248fa949b74c9896285f5b54de2, JAR SHA-256 6e5eacf500b64292bfb73c49797c1958f6ee44646e43e868039ae7feb573ff75.

The engine reported package io.appium.android.apis, target SDK 33, 2,876 classes and 869 resources. The complete 152,541-code-unit manifest produced 355 components: 331 activities, 11 services, 7 receivers, 4 providers and 2 aliases. Component coverage was complete, with no projection limitations. APK signature verification remained explicitly not_performed.

An independent Python ElementTree projection checked every component field, declaration order, namespace attribute, class-name qualification, permission source, export/default, intent value, authority and preceding alias target against the retained XML on all six measured CLI tasks. It also rehashed the APK. This oracle checks REA's projection of JADX-decoded XML; it does not independently validate binary XML decoding or Android runtime behavior. Interpretation was checked against Android's activity, alias, service, receiver and provider documentation.

An actual stdio MCP client launched the same engine, validated the advertised input/output schemas with Ajv2020, checked every structured summary fact, reconstructed the exact manifest in three 70,000-code-unit pages, made a warm complete call, and re-read the retained summary. All normalized facts and provenance matched CLI. Raw engine snapshots and Evidence IDs were not equated: they truthfully retain session-specific private APK paths, Java map key order, load duration and the commands actually run. The warm complete request does not repeat the earlier load command. This client explicitly allowed a 64 MiB receive buffer; default client-buffer compatibility was not established.

Completed-task costs¶

Three alternating before/after pairs ran rea inspect-android-package <same APK> --json. Each task launched a fresh CLI and JVM. OS/JAR caches were already warm; cold disk and first bridge compilation were not isolated. Full normalized results, target identity and provenance were equal; raw per-run engine records were retained without rewriting them.

The elapsed and CPU ranges overlap; these runs establish no speed improvement. The candidate's median sampled memory was slightly higher. Tokens use tokenx 1.3.0 estimates, not model billing: all complete CLI stdout plus the invocation (135,926 + 32 tokens). Discovery, prompts, client parsing and envelopes are outside this known CLI task.

The actual MCP verification task took 132.448 seconds, 40.767 CPU seconds and 1,820.8 MiB sampled peak family RSS, including handshake, full discovery, schema validation, summary, three manifest pages, complete call, retained summary and shutdown. An earlier verification attempt used an incorrect cross-session raw-Evidence equality assertion and failed after 36.131 seconds; that diagnostic attempt is not counted as a successful workflow.

For the successful MCP task, compact JSON payloads estimate 47,583 tokens for the summary versus 135,759 for the warm complete Evidence. These are individual payload estimates, not a measured whole-agent saving; the 222-tool discovery, envelopes, repeated text/structured representations and earlier attempt are not included in that payload comparison. CLI and MCP transports have different costs and are not interchangeable baselines.

CPU comes from waited child-process accounting. RSS samples descendants every 50 ms, may count shared pages more than once and may miss short-lived children; it is not a kernel-enforced process-family maximum. No Android device, split APK, signature-verification, native-library or runtime coverage is claimed. Issue #4448's broader enabled/process, metadata, URI-grant and location requests remain open.

SDK omission and XML namespace follow-up¶

The subsequent review fixes are pinned in dfdba0fac79e07949b12299c85a1a2f3e9c2f115. Android's uses-sdk contract defaults an omitted target to the minimum SDK and an omitted minimum to 1. The component XML owner now combines the bridge's declaration values with attribute-presence facts from the already parsed document. A present-empty attribute or an unrecognized nonempty value stays unknown. Declaration-level target_sdk and min_sdk remain unchanged. Root child nodes are reused, and the effective numeric SDK/provider default is computed once; the redundant per-provider parsing helper was removed.

One unnamespaced-element predicate now guards the manifest root, uses-sdk, application, components, filters, actions and categories. Foreign lookalikes cannot contribute component facts or satisfy an alias's preceding target. Namespace URIs on Android attributes continue to govern attribute selection.

Eleven new negative cases fail on the published baseline, with three positive SDK cases already passing. Three further present-empty SDK cases fail with a simple string fallback and pass with the owned presence facts. Unknown-SDK fixtures now supply an explicit unrecognized value rather than using omission to mean unknown. All 100 selected component/JADX/protocol/SDK-summary cases pass, as do check:fast, format:check, docs:check and discovery of all 2,333 files. Complete inventory coverage does not establish every SDK or export state; nullable facts and limitations retain that uncertainty.

Another three alternating public-APK pairs compare published b50ee503fdff0912c12f1aaa76f043235d187a77 with the corrected revision above. Every normalized fact/provenance field and all 355 component rows match the independent oracle again. The exact-raw-record exclusions and fresh-JVM, warm-OS-cache conditions described above apply.

The ranges overlap and sampled memory median is higher; no speed or memory improvement is established by this comparison. The actual stdio MCP workflow also passes again, including all summary facts, the three manifest pages, warm complete inspection and retained summary. That completed verification took 26.991 seconds, 35.482 CPU seconds and 1,740.9 MiB sampled peak family RSS. Its schema/client-buffer and comparison qualifications above remain in force. These legacy-SDK edge cases are production-boundary/golden checks; the real APK has an explicit SDK 33 target, so a legacy APK without uses-sdk was not independently exercised through binary XML decoding or on a device.

DEX names and hexadecimal SDK follow-up¶

Commit 27b7dc1656b66a71796d6f0d1119d79be7ab586a corrects two additional interpretation boundaries. Class identity expansion accepts the documented DEX SimpleName character ranges, including leading digits, hyphens and supplemental Unicode, instead of Java source identifier categories. The same predicate supplies activity and alias reference matching. Package declaration validation remains separate. Spaces are admitted by the character union introduced in DEX 040; this decoded manifest projection does not establish a DEX version, class existence or runtime loadability.

SDK interpretation now accepts decimal or hexadecimal positive signed-32-bit resource integers once per inventory. JADX's typed value decoder prints TYPE_INT_HEX with a hexadecimal prefix. The unchanged Java bridge retains those declaration strings; only the local numeric interpretation used for default rules changes. Unrecognized previews, malformed hex, exponent syntax, nonpositive values and signed-int overflow remain unknown. Omission and present-empty handling from the prior correction remains intact.

Nine new controls fail before the correction: four legal DEX name/alias combinations, three hexadecimal target levels, hexadecimal minimum fallback, and decimal signed-int overflow. The former malformed-name fixture incorrectly rejected a digit-led class; it now uses forbidden descriptor separators instead. The component/MCP selection passes 77 checks. check:fast, complete formatting and documentation pass. The first combined 117-case selection has 116 passes and one 30-second MCP timeout; the unchanged MCP file then passes all three cases in isolation, with its complete retained-manifest workflow finishing in 16.957 seconds. That workflow supplies a 1 MiB non-XML string; the changed class and SDK interpreters are not reached for that input. No deadline was raised. Further combined verification is recorded below.

Six alternating completed CLI tasks compare the immediately preceding head 58b71c19b35204ebb37c94d9510025bcd18fc4ee with this correction using the unchanged ApiDemos APK. Every projected field for all 355 components matches the independent decoded-XML oracle, and all normalized facts and provenance agree. Per-session raw engine observations and their Evidence IDs are excluded for the same documented reason as the previous comparison. Estimated full CLI stdout plus invocation tokens remain 135,958 (tokenx 1.3.0).

Elapsed medians are 19.619 -> 17.893 s (ranges 12.708–23.907 / 13.658–30.088), waited CPU 30.279 -> 29.507 s (19.684–32.572 / 24.001–37.071), and sampled process-family RSS 1094.6 -> 1064.9 MiB (985.0–1106.2 / 1038.3–1108.5). Every range overlaps; the environment varies substantially and these runs establish no speed or memory improvement. CLI and JVM processes are fresh, with warm OS/JAR caches; cold-disk and fresh bridge-compilation costs remain unmeasured. The actual stdio SDK workflow also passes summary, three manifest pages, warm complete output, retained summary and advertised-schema validation again: 72.614 s elapsed, 78.638 s CPU and 1693.3 MiB sampled family RSS. Its full schema/catalog/transport token cost is not measured. That workflow validates the existing APK's explicit SDK 33 values; no binary-XML hex fixture or numeric-led DEX class was executed or qualified on a device. The new edge cases are boundary controls supported by the format/decoder sources, not claims of real-device coverage.

The subsequent unchanged full selection passes all 117 cases in 67.186 s elapsed, 52.269 s waited CPU and 2028.3 MiB sampled family RSS. This rerun does not erase the earlier timeout recorded above.

Android package identity correction¶

The final pre-merge review found that the package qualifier still used Java identifier categories. Five new negative controls fail before correction: com._hidden, com.$hidden, com.Δelta, an ordinary single-segment package, and a name longer than Android's 223-character parser limit. Existing negatives also cover digit-led and empty segments. Package qualification now has its own Android application-ID grammar: ASCII letter-led segments, ASCII letters/digits/ underscores, at least one separator, and the platform android exception. Unsupported recorded package/name strings remain source evidence; relative component identities and alias references remain unresolved with partial coverage. DEX class-name rules stay separate and unchanged. This does not claim that a syntactically supported package or component is installable or loadable.

Authorities are the Android application-ID documentation, and Android 16's pinned package parser and name validator. The documentation requires nonempty segments; the pinned parser is permissive about empty segments, which this projection leaves unsupported rather than claiming package identity from them. No Package Manager/device execution was performed for these cases.

Four existing amplification/copy-budget controls used invalid huge or Unicode packages. Three now use a valid package and huge ASCII/multibyte component names, preserving aggregate exhaustion, first-row admission and UTF-8 expansion coverage at the correct producing boundary. The fourth preserves 511 unresolved aliases under a huge unsupported package without expanding it. The obsolete Java package predicate is removed; package interpretation is shared by component and alias qualification.

Earlier real-ApiDemos measurements remain pinned to their stated source commits; they are not measurements of this package correction. Response schemas and valid ApiDemos facts are unchanged by source reasoning; no fresh speedup or whole-agent cost claim is made for this correctness change.

Final executed source pin 7b86379e63448391973098572b884b692fbf5d8f incorporates main 98ec00f4068bab894639d0d1d83ac0bfc428bcbf. All 129 selected component/JADX/protocol/MCP checks pass: 47.290 s elapsed, 41.923 CPU s, 1,486.0 MiB sampled peak process-family RSS. The retained-manifest MCP case passes with its existing 30-second deadline. Type/lint, full formatting and generated-document checks also pass. The preceding attempt had 127 passes, one obsolete huge-package identity expectation and one 30-second MCP timeout; these attempted costs are not excluded from the narrative. The fixture expectation is corrected as described above. No deterministic product failure was established from the timeout and no deadline was increased. This completed selection uses primed OS/file caches and a fresh runner; it is a validation cost, not a before/after performance comparison. Family-memory sampling limitations above remain applicable.

Exact DEX class spelling follow-up¶

A later review identified another identity/display conflation: DEX format 040 permits U+0020 in SimpleName, including its boundaries, but component and alias handling still trimmed it. Five source-boundary regressions fail before the correction (87 other cases pass); all 92 component cases pass after it. They cover trailing space, leading space, spaces after a relative dot, a space-only segment, and a fully qualified spelling with boundary spaces. A separate control keeps .Hidden distinct from .Hidden in alias matching.

The XML owner now validates and resolves the original component spelling and compares exact alias references. The redundant raw/trimmed representations and the downstream trim-based invalidity checks are removed. The existing malformed boundary control now uses XML character-reference tabs, which DEX disallows, rather than incorrectly treating U+0020 as invalid. Empty names and the descriptor separator controls remain covered. Package grammar retains its separate owner. The domain schema and retained summary consume these strings without another normalization step.

The source authorities are DEX SimpleName and Android 16's buildClassName, which applies the package to the exact string. These are decoded-XML and source-qualified cases. No authentic space-named DEX APK or device load was executed; the decoded manifest alone does not establish DEX version, class existence or loadability. Earlier real-ApiDemos measurements keep their exact source pins and are not re-labelled as measurements of this correction. The public response schema and ordinary valid-name output are unchanged by source reasoning; no new token or speed improvement is claimed.

Executed space-preservation source pin d1f15f68f65bc546590000eee86c91b1e061bf55 incorporates merged IL2CPP/main 3ce311757dd8ac444c8297b2871464536f93816d. All 135 component/JADX/protocol/MCP checks pass without changing deadlines: 47.168 s elapsed, 43.081 CPU s and 1,550.0 MiB sampled peak family RSS. Type/lint, full formatting and generated-document checks pass. Earlier combined integration had 672 passes and one Android 30-second MCP timeout; the unchanged three-case MCP file passed separately at the original deadline. These outcomes remain distinct. This repeat is a fresh runner with primed OS/file caches; validation costs are not evidence of a general performance improvement.

Final declared-value review¶

Source pin d87801811cc80dd683f482a76f9cb9ad1ab9b670 removes the last trimmed helper. Service foreground_service_type now preserves the declared Android attribute exactly: absence is null, while explicit empty, space-only, padded, token-list and resource-reference spellings remain distinct. This field records a declaration; it does not validate service-type tokens or establish runtime eligibility. The component guide now distinguishes valid DEX boundary spaces from invalid controls and from the separate manifest-package grammar.

Six source-boundary cases were added. Against the previous implementation, 95 checks pass and 3 fail (empty, space-only and padded declarations). After the owner fix, all 141 selected component/JADX/protocol/MCP checks pass at the original deadlines. The newly malformed foreground-type values are verified at the parser boundary; the existing SDK retention workflows do not constitute a real Android runtime verification of those values.

The selected suite took 51.269 s elapsed, 41.298 CPU s and 1632.789 MiB sampled peak family RSS. check:fast, full format:check and docs:check pass. Earlier real ApiDemos measurements retain their stated source pins; no new performance improvement or real DEX 040 space-named class verification is claimed.

At the superseded d799bafa5f2f7c563da04955b540345f7bae70e4 CI run, the source worker exited while starting a JavaScript reference-lifetimes test; there was no failed Android assertion or complete source-suite result. That JavaScript source tree is byte-identical to the integrated main source. This is an observed worker failure, not a reproduced main failure or a passing check. The final head requires its own CI assessment.