Next actions and verification limits

Current review defects

• Fix the P1 remote SQL risk gap for parameterless function calls and calls within CTAS. Reproduce the exact remote planner/confirmation path; the 2026-10-31 review reconfirmed the existing defect without executing SQL on Oracle (SAFETY-01, A22).
• Fix the P1 exit-drain outcome classification: killing a live descendant after normal launcher exit must leave an interrupted/unknown write outcome, and a migration must not be marked succeeded without its completion marker (A14, A16). Cancel the separate drain timer after normal close so it cannot signal a reused process-group ID. Local shell and fake-Oracle workflow reproductions are in docs/evidence/code-review-20261031/; native Oracle/JVM and Windows qualification remain separate.
• Resolve password-file parents against canonical allowed roots and reject directory-symlink escapes while retaining no-follow/inode checks on the file (A21).
• Bind Composer predicate identifiers to verified columns/SQL words and reject unresolved bare parameterless routines in both row predicates and write authorization (CMP013, CMP014, CMP037).
• Make full APEXlang restore use the reviewed backup alias consistently during import and post-import verification; cover a backup alias different from the current configured alias (A13).
• Upgrade the vulnerable locked dependencies, starting with direct Zod 4.6.2 and the relevant MCP SDK fixes. The authorized runtime and complete dependency audits report the same nine affected packages, with no additional packages from dev dependencies; assess reachability individually, preserving the distinction between current stdio execution and HTTP/OAuth advisories. Verify oversized-input handling and the existing CLI/MCP/packaging contracts. No dependency or runtime fix was applied by this review.

Connected and host qualification

• Broaden Decisions labels beyond the four live-checked scenarios and qualify cold-process distributions, backend-cache conditions and native-host authoring separately. The 2026-10-29 run has 20 successful warm measurements per scenario, 80/80 declared group/isolation checks, and three HTTP fallbacks among 92 attempts. Preserve failures in reporting and compare only equal-quality outcomes; investigate HTTP status/retry guidance when failures recur. Review the initial 0.5 threshold against broader labels rather than treating these fixed cases as general accuracy.

• Extend A261-069 beyond the single 26.1 sales-dashboard pilot using adaptive native-host contract reads/search and additional page types. Three runs per flag reached the same offline compiler/46-check threshold: 344.31 s disabled versus 178.05 s enabled, including one timeout fallback. Retain the fixed two-template retrieval and exploratory completion-policy limits; all failed attempts are counted. Generator usage is known, but the timed-out Decisions request has no usage, so exact enabled-arm token savings remain unavailable. Review local Ukrainian dashboard ranking, which selected less relevant templates, without rewriting the measured baseline. Keep 26.2 and connected/browser qualification separate.

• The APEX 26.1.5 read-only P0/P1 comparison passed 20 samples plus one excluded warm-up for metadata and planning. Metadata median improved from 9.62 to 5.42 seconds; plan from 6.86 to 3.13 seconds. The common full working-copy apply uses four connected sessions in fake-runner tests. Its estimated saved session time is 12.45 seconds; actual apply speed remains unmeasured on Oracle. A connected import benchmark needs a separately authorized target/plan and recovery scope. Preserve fresh identity/drift/post-import observations and discard-after-connected-batch policy; the same-session guard remains deferred until transport error-stop behavior is proven.

• Reuse the APEX 26.1 server test catalog (69 baseline and quality/performance cases), nine host-browser shell/session cases, current-build MCP runner and host time/token checkpoints. The authorized 26.1.5 title-import/reset and exact 21-file final audit passed; retain their private evidence and baseline. Extend metadata object/signature fixtures, separately planned backup restore, real fault recovery, Composer runtime and fresh host/platform scenarios with their required targets and authorization. Use the separate 20-repetition read-only benchmark for cold/warm latency; model-driven workflow and historical regression comparisons require their own controlled baseline. The local Codex plugin was refreshed to 2.1.0 on 2026-10-21; open a fresh native session to verify the hyphenated-workspace fix currently confirmed only in the current build.

• Reuse the implemented qualification harness. The Claude usage adapter is implemented and verified on a historical transcript (claude-usage-real.json); the scripted two-turn check remains pending. Read-only distributions, offline compiler/reference checks and native runs have private results. Original faults-26.1-final.json failed 6/7 A261-058 inputs; repaired review-fixes/faults-26.1-R8.json fails 5/7 because next actions are absent, while CLI workspace/compiler parity now agrees. Corrected A261-059 diagnostics pass 2/6. Original extension-26.2-final.json failed target-mismatch classification and golden retrieval; the corrected target check is prerequisite-gated, and retrieval remains 17/21. See the exact triage below. Repair the local 26.2 saved connection/ORDS availability during an authorized setup task before connected parity or cost comparisons.

• Finish definition-specific write/setup/readback/reset/byte-audit fixtures for 26.2 and agent cases, AGT-012's Composer oracle, P2/prewarm measurement bodies and native connected-session limit accounting. Flags currently gate but do not implement these scenarios. Read-only plan jobs expose no completion timestamp; A261-060 and apply-only expiry/tamper checks remain unexecuted. AGT-006/013 still need correctly staged human operator coordination. Do not substitute generic resets or imports.

• Review AGT-007 technical answers independently through manualReview; missing reference/citation remains an automatic failure. Shared dictionary/repair oracles now associate column types and tolerate trailing whitespace. New three-repetition runs of the five selected AGT tasks on both hosts replace the final evidence: 11/15 Codex and 6/15 Claude passed (agent-codex-26.1-final.json, agent-claude-26.1-final.json). Claude Edit/dictionary permission hints and one Codex installed-workspace INVALID_INPUT remain distinct from answer quality; prior reports remain under review-fixes/previous-agent-*-26.1-final.json. Retain actual failures/blocked iterations and exclude them from successful latency statistics. Pin model/effort and bind the exact active native plugin payload before host speed comparisons; the historical Claude runs contain different observed models and recorded native time excludes postchecks. Broader HCX-*/HCL-*, operator and dashboard cases remain unexecuted. Keep all 26.2 results and CodeScan findings separate from the 26.1 baseline; remediation requires its own scope.

• Qualify the remaining connected preview cases on an authorized DEV/QA/TEST schema that may hold a test package:

  • COMMIT and EXECUTE IMMEDIATE DDL inside stored code (expected ORA-00034, or the tripwire);
  • autonomous transactions;
  • a page whose processes call project packages.

  The 2026-11-01 run covered data, binds, APEX session, HTP, rollback and the skip rule for anonymous code. Afterwards, consider apex_error.add_error read-back, LOB/BLOB columns and an ORDS transport that keeps one transaction.

• Exercise the database-only (importMode: none) plan in recovery (mode: recover) after an interrupted script. The plan, apply and tamper checks are unit-tested, and one plan was applied to the demo TEST database on 2026-11-01; recovery of an interrupted database-only run has not been exercised.

• Extend the local preview where design iteration needs it: facet charts, interactive grid pivot/group by, EXECUTE IMMEDIATE/package bodies/explicit cursors in the PL/SQL interpreter, calendar drag-and-drop saves, view-as on report column authorization and submit, BLOB images in cards and grid columns, map color schemes/clustering/icon styles, property checks for report columns, processes and dynamic action settings, and a per-page (not per-session) lost-update snapshot. Record repeatable rendering checks against the real 26.2 assets; keep the preview separate from Oracle evidence.

• Extend local chart AJAX qualification to overlapping refresh responses and dynamic series changes; the crypto comparison currently verifies serial peer changes and a width redraw. Review item-default re-evaluation when an AJAX request explicitly clears a defaulted item; the comparison requires an asset choice and hides its null option.

• Exercise apexrest_peer_review from fresh native Claude Code and Codex sessions on a real saved ship plan: host detection, job waiting, verdict recording and the required apply gate. Resolve Windows .cmd shims for the peer CLIs, and recheck Anthropic's current terms for headless claude -p usage on subscription plans before documenting it as subscription-covered.

• Exercise the shared six-skill/twelve-tool bundle in fresh Codex and Claude Code sessions: skill/tool discovery and the one-call scoped ship/job flow. Manifest-driven subprocess parity and managed binding checks are local evidence only; request browser verification separately when needed. Confirm a shared managed home or externally serialize independent homes.

• On an explicitly authorized development/test target, verify full and selected-file imports, source preservation, authenticated browser CRUD, denied actions, stale versions and narrow screens. Record compiler, import and browser observations separately.

• Qualify Linux/Windows/WSL2 native installations and connected SQLcl/ORDS behavior. Hosted build tests do not establish native host or Oracle behavior.

• Qualify native selected export for shared-component files; current isolated component imports disclose full APEXlang observation. The local page/LOV/dashboard and selected-page restore checks passed; broader families, independent machines and target builds remain separate qualification.

• Exercise current server-first recovery under real interrupted/lost-response imports, backup/restore and externally serialized runners. Retain unknown ownership until the outcome is reconciled.

Open source code review findings

• Fix SQL script risk inspection in packages/core/src/sql-review.ts: parameterless query function calls and calls inside CREATE TABLE AS SELECT currently produce no mutation risk. Both synthetic cases pass the remote confirmation gate without confirmation. Add negative regressions for the complete risk-to-authorization path while retaining ordinary schema creation and built-in function cases.

• Qualify process-tree termination against a real stalled Oracle JVM, on Windows and through a fresh native host; local Node process fixtures do not cover them.

Open quality findings from review qualification

Private triage and repaired harness evidence are in docs/evidence/qp-harness/review-fixes/faults-26.1-R8.json and extension-26.2-R8.json; original final reports remain preserved. These findings do not authorize runtime or catalog changes.

A261-058 unknown-connection now records CONNECTION_REQUIRED / 3 with a next action, and compiler-syntax records INVALID_PROPERTY / 1 with a next action on both interfaces. They pass. A261-059 unknown-property and invalid-region-type retain precise rejection. No runtime semantics or golden expectations were changed for these open findings.

Application verification

• Select further recipes from the remaining two community candidates (both need external AI services) added on 2026-10-14 only after per-source dependency closure, adaptation to the 26.1 recipe scaffold and compiler checks; resolve the LGPL-3.0 review for United-Codes/uc_ai before redistributing any of its files.
• Qualify the internet 26.1 recipes, including the 93 official-source adaptations added on 2026-11-01, on explicit DEV/QA/TEST applications. The editable grid and sectioned form need a project table and row processing before saving; Geocoded Address needs the Oracle geocoding service. Pinned source checks, offline compiler results and packaged MCP probes do not establish SQL execution, import, browser behavior or native-host discovery.
• Use the host in-app browser for affected interactions and relevant narrow layouts; keep observations separate from compiler/import results.
• Reconcile the nine advisory CodeScan findings in the local 26.2 CRM fixture before claiming security qualification.
• Existing projects can drop retired tests and database.testsDir settings; the loader ignores them. Regenerate old plans and reconcile unresolved deployment state rather than rewriting journals.

APEX 26.2

• Use the Oracle synchronization skill for source updates. Review new release directories, imported catalog/examples, licensing and compiler MMD separately; preserve 26.1 evidence and routing. Full inventory retrieval does not close runtime/provider/import/browser qualification.

• The 2026-11-01 check found three new 26.2 sample applications on the oracle/apex 26.2 branch (244dd3691fa5c2897a069fc912b149b9c2853679); recipes were adapted from them, but the 1,350-page public reference library was not re-imported. Decide whether to refresh its archive pin. The 2026-10-30 upstream comparison found no bundled 26.1/26.2 source changes or newer release directory. Oracle's latest APEXlang commit remains 03ee10d02273bc4002fd527089e7fbb5884f94c4 (6 October); the official generation skill is listed only for 26.1. Keep the reviewed pins and repeat the comparison for a future refresh.

• Keep the recipe gap explicit: the local component/pattern catalogs remain bound to 26.1, while 1,350 public 26.2 pages are source references. A 26.2 recipe expansion needs its own dependencies, MMD and compiler evidence. Fifteen pattern recipes remain unresolved, including master-detail concepts without implementations; do not promote them based on inventory coverage. The one unresolved component is the internal theme Button helper without a supported standalone host; use the ready native button recipes.

• Keep the managed installer default at 26.1 until broader qualification is complete. A 26.2 project needs its explicit profile, compiler and lock.

• ORDS and SQLcl MCP-mode selected-file imports require a separate transport implementation; explicit file mode must continue refusing unsupported transports.

• Keep component-order normalization restricted to classic-report columns and page, process and dynamic-action children with unique keys and explicit sequences; other region types and child families need their own Oracle results and negative tests. Add further export defaults only with a reference default and an observed native export; the latest classic-report readback is session-reported, not independently repeated in the follow-up review. Selected-scope checks do not establish unselected server freshness; full audits remain separate.

• Qualify provider calls, DDS, OCI IAM/network prerequisites and workflow migration independently of offline recipes and metadata visibility.

Runtime and Composer

• Complete Composer shared-resource ownership, typed route/consumer handling, property/fault coverage and authenticated Oracle/browser qualification. See Composer.
• Verify the generated save dialog over HTTP/HTTPS and the create-as-edit retry path. Materialization and compiler success do not establish runtime DML or authorization.
• Add recovery for LOCK_CORRUPT, clarify first-use integrity records and verify Windows launchers/rename retries.
• Qualify process-group termination of SQLcl on native Windows (taskkill /T) and confirm the exit-drain fallback against a real stalled apex import; the POSIX behavior is covered by local shell fixtures only.
• Broaden remote database dependency/operation qualification and local-transport identity coverage; production deployment remains forbidden.

Release and repository policy

Run relevant local checks before integrating changes; CI runs only on main. Repository release/integration workflows remain disabled. The retained 2.2.0 npm exception does not apply to source version 2.3.0. Current-version npm eligibility and full qualification require all nine source-bound reports under the existing configuration. Do not treat absent reports as passed gates.

• A requested 2.4.0 publication needs its own applicable readiness decision, current required reports, green final-source CI on all four jobs, matching public archive integrity and a fresh registry installation with CLI/MCP/reference probes. Version update and main integration do not establish registry publication.
• Obtain full Composer runtime/new-chat, native macOS/Linux/Windows and Oracle integration qualification during a separately requested qualification run. The retained version-bound npm exception does not mark those reports passed. Keep original 26.1 records separate from the new block compiler proofs.
• Preserve the passing Windows cleanup regression, .github/workflows/ci.yml, all three local npm gates and the default full-qualification gate. Hosted CI is separate from native model-host qualification.
• Future native qualification needs supported config profiles with the exact current plugin and accessible dedicated Linux x64 and Windows x64 hosts. Do not copy authentication tokens or treat the existing ARM64 Podman VM as Linux x64 qualification.

Preserve functional catalog provenance, licensing and compiler bindings. Preserve private application baselines, APEXlang and legacy SQL backups, migration history and unresolved journals; repository cleanup does not authorize removing active deployment state.

Documentation website

• Keep tracked Pages generated from the current Markdown/site sources; rerun site:check, documentation checks and relevant browser review after changes.
• After a separately authorized push to main, verify the live GitHub Pages website, search, navigation and assets.

View this page source