[GH-ISSUE #4996] [Docs] Qualify JEB Linux CE launch and native method references #3087

Closed
opened 2026-10-31 00:36:21 +00:00 by bsakaguchi · 1 comment
Owner

Originally created by @gensokyo-dev on GitHub (Oct 18, 2026).
Original GitHub issue: https://github.com/morluto/rea/issues/4996

Before submitting

  • Searched existing JEB issues/PRs; #3754 covers reconnect, #4600 metadata, #2987 setup discovery.
  • No credentials, private paths, proprietary decompilation, or sensitive target data included.

Documentation location

docs/jeb-analysis.md, plus the JEB lane in docs/testing.md.

Problem

The guide generalizes an earlier licensed-build CLI limitation to JEB 5.48, omitting the current official Linux CE direct -c --mcp launch route. Native item selection also needs qualification: the audited SQLite method reference returned by JEB is sqlite3_libversion_number; raw ELF-address attempts did not complete successfully. A numeric reference did work for the authored ET_EXEC pilot, so there is no universal numeric-address rule.

Expected content

Document the verified CE launcher with selected JDK, explicit loopback/input and official first-use setup; retain the programmatic alternative and qualify earlier licensed results. Explain engine-native item references and the current code-item-discovery gap. Preserve measured costs and failures, without speculative cache or speedup claims.

Supporting evidence

Baseline 7c458f04b9155707c8ba406f8b455b10016f1b99, Linux x86-64, Node 24.18.0, Temurin 21.0.12.1. Official CE 5.48.0.202610071539 ZIP: 697,549,250 bytes, observed SHA-256 709e7fe8e88bb0e9e2bc03b028f8b9fc78cec0289d04abe54fc26cda6369e175, via https://www.pnfsoftware.com/dl?jeb4ce. The upstream manual describes the direct launcher.

Six alternating fresh-engine tasks completed: three authored ELF and three Debian SQLite library sessions, 90 public stdio MCP calls plus six CLI parity queries. Pagination, selected reads, repeated facts/Evidence identity, native-file oracles, unchanged bytes and owned cleanup passed. All completed tasks observed one metadata request per persistent MCP client; no new wrapper cache is justified. Raw catalog serialization estimates are distinct from actual model usage, which remains unmeasured.

Additional context

A documentation-only branch is prepared. Main advanced to release 6.7.0 (a25eb5fbb); rebuilding and uncached typecheck/lint plus docs checks passed, as did an additional fixture control. The additional SQLite control was safely guard-stopped when host available memory fell below 8 GiB, with no OOM; its retry is pending with stricter admission. Earlier EULA/activation, timeout, raw-address and test-oracle failures remain recorded as incomplete attempts.

CE Java/Android, GUI, other hosts, >100 units and restart recovery are NOT_ENOUGH_INFO: unverified in this audit. Existing #3754 is not duplicated. Production runtime/contracts are unchanged by this documentation work.

Originally created by @gensokyo-dev on GitHub (Oct 18, 2026). Original GitHub issue: https://github.com/morluto/rea/issues/4996 ### Before submitting - [x] Searched existing JEB issues/PRs; #3754 covers reconnect, #4600 metadata, #2987 setup discovery. - [x] No credentials, private paths, proprietary decompilation, or sensitive target data included. ### Documentation location [docs/jeb-analysis.md](https://github.com/morluto/rea/blob/7c458f04b9155707c8ba406f8b455b10016f1b99/docs/jeb-analysis.md), plus the JEB lane in docs/testing.md. ### Problem The guide generalizes an earlier licensed-build CLI limitation to JEB 5.48, omitting the current official Linux CE direct `-c --mcp` launch route. Native item selection also needs qualification: the audited SQLite method reference returned by JEB is `sqlite3_libversion_number`; raw ELF-address attempts did not complete successfully. A numeric reference did work for the authored ET_EXEC pilot, so there is no universal numeric-address rule. ### Expected content Document the verified CE launcher with selected JDK, explicit loopback/input and official first-use setup; retain the programmatic alternative and qualify earlier licensed results. Explain engine-native item references and the current code-item-discovery gap. Preserve measured costs and failures, without speculative cache or speedup claims. ### Supporting evidence Baseline `7c458f04b9155707c8ba406f8b455b10016f1b99`, Linux x86-64, Node 24.18.0, Temurin 21.0.12.1. Official CE `5.48.0.202610071539` ZIP: 697,549,250 bytes, observed SHA-256 `709e7fe8e88bb0e9e2bc03b028f8b9fc78cec0289d04abe54fc26cda6369e175`, via https://www.pnfsoftware.com/dl?jeb4ce. The [upstream manual](https://www.pnfsoftware.com/jeb/manual/mcp/) describes the direct launcher. Six alternating fresh-engine tasks completed: three authored ELF and three Debian SQLite library sessions, 90 public stdio MCP calls plus six CLI parity queries. Pagination, selected reads, repeated facts/Evidence identity, native-file oracles, unchanged bytes and owned cleanup passed. All completed tasks observed one metadata request per persistent MCP client; no new wrapper cache is justified. Raw catalog serialization estimates are distinct from actual model usage, which remains unmeasured. ### Additional context A documentation-only branch is prepared. Main advanced to release 6.7.0 (`a25eb5fbb`); rebuilding and uncached typecheck/lint plus docs checks passed, as did an additional fixture control. The additional SQLite control was safely guard-stopped when host available memory fell below 8 GiB, with no OOM; its retry is pending with stricter admission. Earlier EULA/activation, timeout, raw-address and test-oracle failures remain recorded as incomplete attempts. CE Java/Android, GUI, other hosts, >100 units and restart recovery are NOT_ENOUGH_INFO: unverified in this audit. Existing #3754 is not duplicated. Production runtime/contracts are unchanged by this documentation work.
bsakaguchi 2026-10-31 00:36:21 +00:00
Author
Owner

@XinwangTao commented on GitHub (Oct 18, 2026):

Solid verification — the explicit loopback binding and caller-selected input make the documented route reproducible instead of relying on whatever the CE distribution guesses, and keeping CE native coverage separate from licensed/Demo Android and GUI lanes is the right scoping for an audit-backed doc update. One small thing worth calling out in the doc text: the verification used -XX:ActiveProcessorCount=1, so the "one CPU / no-swap guard" limits were part of the evidence; readers on multi-core hosts should expect different resource profiles, and the docs should say the guard was deliberately minimal rather than representative. Also good that #3754 reconnect work stays untouched — the launch route and session lifecycle are orthogonal concerns.

<!-- gh-comment-id:6247603660 --> @XinwangTao commented on GitHub (Oct 18, 2026): Solid verification — the explicit loopback binding and caller-selected input make the documented route reproducible instead of relying on whatever the CE distribution guesses, and keeping CE native coverage separate from licensed/Demo Android and GUI lanes is the right scoping for an audit-backed doc update. One small thing worth calling out in the doc text: the verification used -XX:ActiveProcessorCount=1, so the "one CPU / no-swap guard" limits were part of the evidence; readers on multi-core hosts should expect different resource profiles, and the docs should say the guard was deliberately minimal rather than representative. Also good that #3754 reconnect work stays untouched — the launch route and session lifecycle are orthogonal concerns.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
morluto/rea#3087
No description provided.