omnius 1.0.892 → 1.0.894

package/docs/rea.md

@@ -0,0 +1,39 @@
+# REA reverse engineering
+
+Omnius includes [`rea-agents` 6.7.0](https://www.npmjs.com/package/rea-agents/v/6.7.0)
+as a required dependency, with the matching MIT-licensed investigation skill
+and reference guides. The [REA documentation](https://rea.tools/) describes its
+supported targets and analysis providers.
+
+```sh
+omnius rea status --json
+omnius rea analyze-javascript-application "/absolute/path/to/app" --json
+omnius rea --help
+```
+
+`omnius rea` runs the installed package directly. It preserves trailing arguments
+and exit status, requires no model server, and performs no additional npm
+download or setup. Static JavaScript/Electron inspection does not need a native
+decompiler. Native routes use existing provider configuration; Omnius does not
+install Ghidra, Java, Hopper, IDA, or other engines automatically. REA's optional
+native providers keep their own prerequisites.
+
+The `omnius-rea` and `reverse-engineer-anything` skills appear in normal skill
+discovery from any working directory. Only their metadata enters the skill
+listing; their bodies and reference guides load when selected. Project-local
+skills retain precedence. The shipped `assets/rea/UPSTREAM.json` records the
+upstream release integrity and checksums of the original skill and license.
+
+For an MCP connection, `omnius rea mcp-config` prints the exact installed stdio
+binding without changing any client configuration. `omnius rea --mcp` starts that
+server with unmodified protocol streams. Use the existing MCP onboarding path
+when a task benefits from that connection. Ordinary tasks do not start REA or
+load its full tool catalog.
+
+The upstream `omnius rea mcp add` and `omnius rea mcp doctor` commands retain
+their registration and connection-diagnostic behavior and arguments.
+
+Begin with the user's explicit artifact or capture. Keep static observations,
+runtime observations, inferences, and unknowns distinct. Preserve evidence
+identities and coverage limits; execute a target only within the requested
+investigation's scope.

package/docs/test-three-premature-completion-rca-20261021.md

@@ -0,0 +1,190 @@
+# Premature completion and stale activity in `test_three`
+
+The inspected session had stopped making requests because it had committed a
+successful terminal result. It was not waiting for a model at collection time.
+That result did not establish that the original request was fulfilled.
+
+## Evidence and scope
+
+Collection was read-only. No input was injected, no project source was changed,
+and no inference, model load, service restart, or GPU deployment was performed.
+The private evidence case is
+`test-three-premature-completion-20261022T013213Z`: 78 initial files, 2,093,283
+bytes, with source stability checks and SHA-256 hashes. Three supplemental
+files preserve font bindings and the persisted agent view. Raw conversations,
+tool outputs, and local runtime records are deliberately excluded from Git.
+
+Initial manifest SHA-256:
+`d7d851dbcfc3e24a923d7bd9815c16e8231a518956adf1559225edb6c6b0adcf`.
+Supplemental manifest SHA-256:
+`7e0a9cf823ee7cf384c80916fedf945f6b0b983e74c0394639153b9e5eb596a5`.
+
+The terminal receipt identifies the loaded package as `omnius@1.0.893`, with
+module SHA-256
+`7a890817c1b4386b83e8f9ff2b59a387ffc8b1e1e45d77993734b24da8236105`.
+The source baseline inspected for this repair is commit `196cecff`.
+
+## Timeline
+
+Times below are UTC on October 21–22, 2026; subtract seven hours for local PDT.
+
+| Time | Observed event | Implication |
+| --- | --- | --- |
+| Oct 21 22:01:54 | Original detailed task committed an incomplete result. Its objective had generic effect criteria but no native semantic requirements. | The original request remained unfinished; generic observation, mutation, and verification are insufficient to establish reference parity. |
+| Oct 21 23:16:34 | Last observed application source edit. | Later completion cannot be explained by newly implementing the missing features. |
+| Oct 22 00:52:02 | A new run started with goal `continue until completion` and no objective contract. | The follow-up displaced the original objective instead of continuing it. |
+| 00:55:20 | Parent delegated a read-only discovery brief with that narrowed goal. | The child was not evaluating the complete original acceptance requirements. |
+| 00:56:17 | Child described the existing page as complete and reported default framework fonts. | A child verdict was mistaken for evidence of original requirement coverage. |
+| 00:56:24 | Parent ran a successful production build. | Compilation was demonstrated; visual and reference requirements were not. The build receipt itself is genuine. |
+| 00:57:55 | Parent committed a completed terminal receipt with no unresolved completion obligations. | A real terminal commit recorded an insufficiently grounded success. |
+| 00:59:40 | Optional memory evaluators wrote a failed lifecycle state under the main session identity. | Background bookkeeping obscured the main lifecycle after its terminal result. |
+| 01:28–01:33 | No active or queued inference was observed for this session. Persisted child view had `stage: completed`, `status: running`, and no completion timestamp. | The lingering activity row was stale; it did not establish live progress. |
+
+## Causal defects
+
+### Incomplete results disappeared from frontend outcome handling
+
+`startTask` invoked its result callback only inside `if (result.completed)`.
+Normal task admission cleared `lastTaskMeta`. After an ordinary incomplete
+result, the frontend therefore retained null metadata. Its finalizer tested
+`lastTaskMeta?.verified === false`; null did not satisfy that test, and the
+work queue recorded completion. The same missing callback omitted the original
+request and incomplete response from the in-memory conversation used for the
+next admission review.
+
+This is a confirmed production path, not an assumption that every normal
+promise resolution means success. Both the callback placement and the null
+comparison are needed to explain the failure.
+
+### Idle continuation had no retained objective authority
+
+The interaction router supports an active scope and a reviewed continuity
+decision. Ordinary idle TUI admission did not supply that scope or carry its
+continuity decision into an objective continuation. It passed the latest input
+as `actualUserGoal` and created a fresh runner without an objective contract.
+Only an explicit pending interruption had an identity-preserving path.
+
+Consequently, conversational history and handoff prose could suggest earlier
+work, but the runner lacked the original request, accepted steering, and
+requirement identity as completion authority. Matching the word `continue`
+would not repair this: the host must decide whether a message continues,
+replaces, queries, or starts work against the real conversation scope.
+
+### Completion accepted an empty requirement set
+
+The original contract already expressed only broad effect categories. The
+latest run had no contract at all. Canonical requirement coverage review was
+conditional on native requirements being present, and generic ledger claims
+were advisory. Completion readiness consequently received no semantic
+obligations and reported ready.
+
+An offline reproduction using the retained ledger and the production readiness
+and terminal-record modules reproduced this result: zero obligations, zero
+canonical claims, `intrinsicReady: true`, and a completed terminal record.
+This does not make all heuristic summary claims authoritative, nor does an
+audit ledger remaining `open` by itself prove an invalid terminal. The defect
+is the missing human-origin requirement authority at admission and closeout.
+
+A passing build cannot establish that a requested font, layout, reference
+structure, live preview, or other distinct requirement was implemented. In
+this case, inspected source still imported default framework fonts while the
+reference-font assets existed unused. That is concrete counterevidence to the
+unqualified completion claim.
+
+### Internal helpers inherited the main lifecycle identity
+
+The two memory evaluators created internal runners without independent session
+identities or a background workload class. Non-subagent runners inherited
+`OMNIUS_SESSION_ID` even when their artifact mode was internal. Their lifecycle
+constructors recovered and persisted the main session file. Subsequent owner
+checks rejected their terminal commits, and their failure projections then
+overwrote that same lifecycle file.
+
+The main immutable completion receipt survived. There is no evidence that the
+evaluators changed its run-specific objective ledger or fulfilled the original
+task. Their separate faults were lifecycle contamination, incorrect workload
+priority, and misleading later failure state.
+
+### Child view retirement was not an outcome
+
+The saved child row combined a completed stage with a running lifecycle status.
+Conversely, removing a view could manufacture a successful status for a child
+that had never reported an outcome. Presentation must follow an explicit
+terminal result or classify an unresolved retired view as interrupted; a
+stage label and a teardown callback cannot certify completion.
+
+## Repair and validation
+
+The repair separates authored task authority, actual tool evidence, independent
+review, and frontend outcome projection:
+
+- TUI action admission reviews the complete original request and accepted
+  corrections. A separate critic checks that the interpretation preserves that
+  scope. Continuation routing receives the retained objective; unrelated new
+  work receives a new objective. Native source bytes, conversation ownership,
+  and predecessor ledger hashes are checked before reuse.
+- Every returned runtime outcome reaches frontend settlement. Missing results,
+  incomplete results, stopped work, and rejected promises cannot become queue
+  success or successful learning records merely by resolving a promise.
+- Required human-requirement admission is enforced in tool execution,
+  completion readiness, and terminal commit. Generic build success and child
+  summaries cannot substitute for the requested semantic comparison.
+- Completion claims receive an independent, evidence-bound challenge against
+  the full request and accepted changes. The worker receives specific questions
+  about existing artifacts and missing results, and can answer with authentic
+  evidence or repair the artifact. The actual reviewer verdict and exchange are
+  retained as audit records, not promoted into tool-effect evidence.
+- Identical rejected claims reuse their real challenge rather than launching
+  another reviewer. Rewording alone has a bounded review budget, and repeated
+  unchanged claims have a bounded feedback budget. Actual new evidence permits
+  a fresh review. A supported result can pass immediately; cancellation,
+  superseding input, and update/resume release temporary review reservations.
+- Internal memory evaluators have independent sessions, lifecycle storage,
+  background admission, and owned cancellation. All-failed evaluations preserve
+  the local fallback rather than publishing a fabricated score.
+- Child views settle from actual child outcomes. Teardown without a result
+  means stopped, and late progress cannot overwrite a settled stage.
+
+Regression fixtures are neutral and use CPU-only mock transports. They do not
+continue the inspected application task or send requests to local inference
+services.
+
+A read-only probe of the new recovery selector against the actual `test_three`
+queue and ledgers selected the original detailed request, not the later
+`continue` goals. It identified exactly the two previously applied human
+corrections whose objective lineage was lost. Their relationship is reviewed
+before adoption; sharing a conversation does not automatically authorize it.
+No task was resumed or modified by this probe.
+
+The archived-ledger reproduction also confirms the changed boundary: the old
+unstructured call remains compatible for advisory/internal use, while the TUI's
+explicit required-admission policy produces a non-overridable missing-human-
+requirements obligation and refuses success for that same actual ledger.
+
+Final verification covered 623 distinct regression cases:
+
+| Scope | Result |
+| --- | --- |
+| Core admission, evidence, independent challenges, continuation, interruption and update compatibility | 336 passed across 22 files |
+| TUI admission and settlement, child views, background helpers, daemon admission Stop, and required web transport lifecycle cases | 196 passed across 19 files |
+| Child execution and actual outcome publication | 28 passed |
+| Requirement review and exact citation parsing | 63 passed |
+
+The core count combines the broad run with a focused rerun after correcting a
+pre-existing test expectation. A clean baseline reproduced the stale assertion:
+the third request offered a bounded repair for missing mutation evidence. The
+updated test checks that exact feedback, the missing effect, no invented
+mutation, and one incomplete terminal result. No failing case was excluded.
+
+All 11 workspaces rebuilt from cleared TypeScript build caches. The complete
+publish bundle and fresh dependency lock passed the artifact audit. The packed
+881-file development tarball also passed the tarball audit with zero warnings.
+After the final rebuild, nine actual adapter/runner and standalone control
+cases passed again against the compiled outputs; these reruns are not added to
+the distinct-case count above.
+
+This verifies the host protocol and the reproduced failure paths. Backend
+responses in the regression suites are controlled fixtures, so this is not a
+benchmark of an actual model's judgment. The inspected live process and its
+application were not restarted, resumed, or altered, and no npm release was
+published as part of this RCA.