Historical change record. The September 11 application-wide shutdown and its blanket resumption restrictions were superseded by the operator on September 15. This page records past actions and results; it does not impose current execution restrictions. See the stop-scope recovery record.
The operator authorized the engine repair after confirming that a feedback message must be able to produce independently tracked outcomes. The selected source is EventHorizon feedback 43625, cycle 98225, existing native parent issue 1563. One message reported sleep, effect lifecycle, item continuity and collaboration problems, including property renting as an example. The old agent treated the combined report as one review-required scope decision. Telegram delivery had been verified, but that did not validate the escalation itself.
The supervised validation is limited to reassessing and recording separate outcomes for this one feedback message. It does not authorize EventHorizon implementation or restarting general automation. Existing issue 1571 and 1447 gates remain unchanged. Incomplete Telegram inbound work is preserved separately and is not included.
PR 4399 merged as 3f4504f75f5a006ce6d14817a320cdbf6e4f91a0, with the exact tested source tree. REQ-0131 was updated and pushed separately before implementation. New planner results support distinct semantic outcomes with source coverage, independent classification, priority, authority, risk and decisions. Existing schema-1 results remain readable for ongoing work.
Compound results are retained before child creation, then materialized through existing native issue and native feedback receipt writers with stable source references. Each child links the original feedback/cycle/task and follows ordinary planning, review and verification independently. Parent operator priority is preserved. Already-separated outcomes cannot recursively expand. The original report stays pending until all required children are resolved and verified; historical work links are preserved. The obsolete combined question is retired only after the new links are retained.
The explicit assess-feedback-outcomes-once and apply-feedback-outcomes-once commands support one historical review. Assessment is read-only. Application records native outcomes but does not start workers. Original planning result and review history remain retained.
Full Rust tests passed (413 library, 1278 main, 86 Gitea MCP, 16 Kubernetes MCP; ignored database tests are separate). Isolated PostgreSQL regression passed in both binaries, including replay, changed-plan refusal, historical-link preservation, unchanged child states and all-child verification before parent resolution. Formatting and diff checks passed.
The exact merged image is building. No live outcome assessment or child issue creation has occurred yet. Read-only worker preflight refreshed a normal, non-fallback ChatGPT subscription at 06:03 UTC: eligible, primary usage 0 percent, secondary usage 16 percent, required model advertised. No paid API fallback is authorized.
At this historical checkpoint the backend deployment preserved server-only arguments; that restriction was superseded on September 15. Only the idle supervised worker image may be replaced; retained workspace/history remain on its existing storage. Exact-generation rollout verification and previous image/revision are retained for rollback. Build resources are bounded to four CPU, 24 GiB memory and an 18 GiB temporary filesystem. Temporary registry credentials and builder are removed on exit.
Evidence and scripts are retained under /home/ax_julian/temp/feedback-outcomes; source is /home/ax_julian/temp/autocoder-feedback-outcomes. Cleanup review is due 2026-09-22. Application architecture is documented in the repository at documentation/architecture/FEEDBACK_OUTCOMES.md.
The exact 3f4504f75f5a image (sha256:4519bc3ebdada025213f18207b17232f888c94f4d1a310cf0aeb66626efa6815) reached backend generation 2387 and the idle supervised worker. The one task feedback-outcomes-c98225 was explicitly claimed on the refreshed normal subscription. It cloned the target, then failed before starting a model turn: the observation reader joined planning tasks only through the original planning_task_id, while the supervised reassessment deliberately preserved that original owner until result review. No thread, tokens or child issues were produced.
PR 4400 fixes the narrow evidence lookup for explicitly scoped outcome assessments. It still requires the expected scope kind, cycle ID, matching repository and matching task target; ordinary planning lookup is unchanged. The extended isolated PostgreSQL test accepts the scoped assessment and rejects mismatched repositories. Both database executions and the full Rust suite passed. The corrected image is building. The failed attempt remains retained, and only the same task will be retried after deployment.
The scoped retry completed read-only and returned eight proposed outcomes. Admission rejected underscore identifiers even though the planner contract allowed a generic lowercase identifier. The result also proposed several product questions before exhausting existing-contract investigation and included unrelated systemic observations in its coverage summary. No child issues were materialized from this result.
PR 4401, merged as b00e0fe6f0667e67f3a9f93cb5684f86bd0c7665, accepts stable underscore identifiers, clarifies evidence-first decomposition without inventing product systems, and removes the stale model instruction. Full local Rust tests and formatting passed. Image delivery and a supervised reassessment of the same source are in progress. The original task result and accounting are retained for comparison; no general restart or managed-product implementation is authorized.
The refined assessment returned six outcomes from the same source: five bugs under established requirements and one property-rental feature request. Root review checked the full reported concerns against each rationale and implementation brief. Collaboration and shared play remained one outcome; missing reproduction did not become a product-choice gate for the five bugs. Property rental retained unknown priority and its own required product decision. The first assessment and its full task/accounting record remain in cycle evidence outcome_assessment_review.prior_task; the original combined planning/review remains in prior_compound_review.
Backend generation 2389 and the idle supervised worker run revision b00e0fe6f0667e67f3a9f93cb5684f86bd0c7665, image digest sha256:2e88134893013fe59d90187094d86fdaf91ff8cea0fc0917d4df52962e41b37d. The reassessment used the previously verified scope-capable runtime with the corrected persisted planning instructions; materialization used the final image. The exact backend receipt is recorded. General backend mode remains server-only.
The actual engine apply-feedback-outcomes-once command materialized:
| Native issue | Outcome | Classification | Priority |
|---|---|---|---|
| 1576 | Scene progression | Bug | Important and urgent |
| 1577 | Sleep control | Bug | Important and urgent |
| 1578 | Effect lifecycle and narration | Bug | Important and urgent |
| 1579 | Continued item usability | Bug | Important, not urgent |
| 1580 | Collaboration and shared play | Bug | Important, not urgent |
| 1581 | Property rental | Feature | Undecided |
Each card has a distinct canonical native-feedback receipt and a stable source identity linked to original feedback 43625. Parent case/cycle is outcomes_pending, preserving previous links. The obsolete combined escalation and review are closed as superseded by the split; this is not a product approval or a gameplay resolution. Replaying the materializer preserved child IDs, receipt IDs, child activity count and parent links.
An authenticated browser check of the normal EventHorizon issue board at https://codecartography.dev/issues opened all six cards and verified their type, priority and pending state with no browser errors. Property rental displayed Important/Urgent as Undecided. The board also displays the existing generic Activity: Scheduled projection even though retry/admission is not scheduled; authoritative DB checks confirmed zero running workers and only the pre-existing queued acceptance task. No implementation was started. Protected cycles for issues 1571 and 1447 match their saved baseline.
Evidence is bounded under /home/ax_julian/temp/feedback-outcomes/: verification.json, ui-verification.json, sanitized api-evidence.json, replay comparisons, retained assessment results, test logs and image receipts. Temporary authenticated sessions were revoked, isolated builders/registry credentials removed, and the test PostgreSQL port-forward closed. Evidence retention review is 2026-09-22.
This validates decomposition, independent stored classification/priority, source linkage, replay safety for child creation, stale-review retirement and normal-board visibility for one feedback message. It does not validate gameplay fixes, subsequent child planning/implementation/delivery, fleet-wide assessment quality, or Telegram inbound replies. Children await normal governed admission after explicit restart or a separately selected supervised run. Parent resolution still requires verified child outcomes. The retained first bug rationale includes contextual systemic evidence; that evidence does not expand its authorized implementation scope.