Concresca UAIX intake processing update

Report body

Concresca UAIX intake processing update

Review date: 2026-09-08. Starting source: 604c2221ed77c9b1bbad90e40c55906c8c5d3c79.

Request and scope

The user asked to make sure agent-file-handoff/Improvement is being processed and update the UAIX strategy using the selected file-handoff plus loops setup. The wizard was fetched live and read as external guidance. It is not runtime authority and does not itself host, schedule or execute an intake processor.

The initial live scan found 47 non-placeholder files, 2,206,496 bytes, in Improvement; Content contained only its placeholder. The ZIP is one pending container with 33 members, not one reviewed document. Member paths/sizes were inspected without extraction or execution; their contents remain unreviewed.

This round changes private handoff/processing sources, not production API, database, public content or visual identity. No production restart is required. Independent two-new-agent acceptance remains an open mandatory gate; the live onboarding evidence in v049-independent-dogfood-onboarding-remediation.md does not establish that external acceptance.

Current strategy implemented

The old startup/system profile described a static-only Linux workspace with no Git or live credentials, and the stored continuation resumed superseded v0.48 work. Exact prior top-level memory and manifest bytes were copied to .uai/archives/intake-strategy-before-604c222/ before replacement: 28 files, 339,985 bytes. The prior human-readable handoff was additionally recovered from the starting Git commit into the same historical custody directory. Doctrine files and the existing persona source/evidence were left unchanged. These snapshots preserve history, not current instructions.

The manifest now selects the requested existing-setup update, file handoff and loops. Twenty concise date-free typed hot files own current context, execution, intake, source review and continuation. Mandatory doctrine remains explicitly required. Exports now derive only from that roster instead of concatenating obsolete release state and dormant persona. Historical source-authority fields and canonical persona records remain preserved, not silently reinterpreted.

scripts/concresca_intake.py inventories both buckets, binds records to path and exact SHA-256, and registers new material only as unreviewed/unfinished. It never accepts, executes, archives or deletes a source. It rejects path escapes, linked paths, missing/unresolved files and unsafe ZIP members, and checks review/proof/custody references. A passing integrity check with active files is not completion; --require-empty is the explicit whole-queue gate. Substantive truth and test sufficiency still require agent review; merely creating an evidence file cannot establish a correct implementation.

Newly completed full source review

Source: Deep Research Scoping and Execution Guide.md, 28,717 bytes. SHA-256: 98e1f90e7f92b977826f104105bd44c91e94308762536502071065952b5db9be. Complete source read: executive summary, scope questions, evidence hierarchy, database/tool suggestions, search method, source rubric, timelines/depth, deliverables/templates, budgets, risk controls and next steps.

Disposition: incorporated with explicitly bounded acceptance.

Accepted into .uai/research-review.uai, .uai/file-handoff.uai, .uai/operations.uai and the derived resume: identify the user's question and owning destination; separate observation, inference and recommendation; trace load-bearing claims to actually reviewed primary evidence; record contrary evidence, source-version limitations and narrow support; keep each batch bounded; preserve exact original sources and an auditable outcome.

Not adopted:

• The numeric source-quality score and weighted rubric. Descriptive source checks are sufficient here; no universal score or participant ranking is introduced.
• Mandatory human charter completion before any useful work. Existing user scope and product owners provide the default; ask only for a material missing choice or authority boundary.
• Illustrative fixed hours, costs, source counts, calendar schedule and paid tools as requirements or market facts. No spending or account creation.
• Opaque conversation citation tokens as evidence. Named laws, databases, standards and product claims are research leads, not verified current facts adopted or published by this round.
• A new public article or bulk memory upload of the raw guide. It contributes a local processing method, not public factual knowledge or runtime authority.

Contrary/limiting evidence: Concresca already had a substantive source/authority doctrine and prior reviewed intake; a new guide does not replace those owners. The useful method is adapted to an operatorless, judgment-free service rather than adopted wholesale. No unresolved implementation is hidden by closing this method-only source.

Exact custody: reports/input/v049/Deep Research Scoping and Execution Guide.md. The source and custody SHA-256 were compared before intake-copy removal. The original remains recoverable from repository custody; it is not public-deploy content. The ledger records destinations, review and validation evidence.

Reconciled prior reviews and remaining work

Seven exact source hashes match v048-agent-integration-intake-review.md: Conformance Diagnostic, API Documentation, Credential Lifecycle, HTTP MCP Interoperability, AEO, Structured Data, and Ecosystem Registry Discovery. Five exact hashes match v048-remaining-improvement-intake-review.md: Monitoring, GEO Benchmark, Multilingual Discovery, Performance/Accessibility, and Security Readiness. Their full-review provenance is retained explicitly; this is not a claim that all twelve were freshly read in this round.

All twelve remain kept-active with concrete follow-up actions. Their earlier source/deployment and interface assumptions must be compared with current v0.49 owners before any closeout. Old statements that a current interface did not exist, or that deployment was unavailable, are not renewed as present facts. Credential lifecycle was also freshly read in full: its human approval queue, invented OAuth requirements, raw replay-secret caching, general predecessor credential grace and imported Christ.cx authority are not adopted. Stable identity, pre-send durable custody, bounded exact retry and recovery remain the useful themes; broader recovery acceptance stays unfinished.

After this bounded closeout: 46 active files remain — 12 previously reviewed but unfinished, and 34 unreviewed/partial sources including the ZIP. The complete intake is not processed. The queue JSON owns exact identities and next actions; future counts must be derived from live scan.

Verification and continuation

Focused intake tests: 15 run, 14 passed and 1 native symlink test skipped because the OS did not allow link creation. A separate mocked junction guard passed. Six current-memory tests passed, including typed roster, exact setup choice, unchanged doctrine, date-free hot state, current continuation and deterministic exports. Export regeneration/check passed. A test exposed Windows ZIP filename normalization; the scanner now validates original names before normalization can hide backslashes or NUL truncation. An inaccessible directory must raise an integrity failure, never look like an empty bucket.

The first sandbox test attempt could not access newly created temporary fixture directories. The successful run used approved isolated temporary-file access; unattended execution must not weaken sandbox permissions to bypass that limit.

The on-demand continuation selects a bounded source/batch, fully reviews it, applies useful scoped changes, validates them and records actual outcomes. Existing unrelated machine-controller drafts remain untouched. A same-task hourly heartbeat was requested, but app safety review rejected persistent change/push/deploy authority without specific recurring authorization from the user. No scheduled intake processor was created. This is a scheduling/authorization blocker, not an intake review failure. Request explicit authority before retrying; do not create an indirect scheduler as a workaround. The ordinary on-demand workflow remains available and this bounded round completed its review and implementation. For local-file scheduled work, the machine and desktop app must remain running, as described in the official scheduling guidance. No promise of continuous server-side or offline execution is made.

Release delivery must include committed-source repository and root-deploy ZIPs with filenames ending -wip.zip. Raw reports, .uai, source intake, credentials and mutable runtime state must be excluded from the root-deploy archive. The packager filename suffix now follows that exact rule. Git attributes preserve new raw custody and historical snapshot bytes without newline normalization; current export source hashes use declared UTF-8/LF normalization so ordinary Windows checkouts do not change derived logical-source identities. Unreviewed intake remains local and untracked, not silently committed or published. The committed-source ZIP is not a backup of those pending inputs.

Final live scan after verified removal of the processed intake copy: integrity PASS, 46 active files / 2,177,779 bytes, 12 reviewed-but-unfinished, 34 unread or partial, one closed record, 47 tracked identities. --require-empty correctly returned the incomplete gate; no whole-queue success was claimed.