v1.3.31 - Safe reconstruction changes

Request-driven changes to existing reconstructions now have focused guidance for reconciling current intent and custom code with retained compatibility, tracing shared dependencies, selecting acceptance/regression checks, and choosing proportional data recovery. New reconstruction requirements stay distinct from reference-version and environment changes; planning remains separate from implementation authority.

The bounded Pocket Ledger example freezes the original baseline and literal outcomes before adding a completion action. It exercises retained listing/import/export, rejected edits, injected save failure/retry, reopen and previous-code readback, with six targeted faulty variants. Manual contextual judgments and runtime observations remain separate; this does not certify arbitrary model comprehension or global compatibility.

Validation includes 535 source and extracted-package tests, eight client/integration install cases and eight exact-ZIP upgrades, manifest/link/privacy checks and genuine skill validation. The CI matrix retains three existing Linux platform skips. Historical evidence and installer/updater/integration safeguards are preserved. Same-agent source-exposed review; no additional runtime dependency, provider or service.

v1.3.30 — Interface accessibility

I added focused accessibility guidance for reconstructed interfaces: keyboard actions, focus order/visibility/return, names and labels, error correction and dynamic status feedback. Observed original behavior stays separate from proposed improvements and from visual/functional fidelity.

A separate owned Pocket Board example demonstrates selected keyboard and focus paths with 22 native-key browser checkpoints and 14 focused regression tests. Historical fixtures remain unchanged. Screen-reader announcements, other devices/browsers and measured contrast remain unrun with practical manual next steps; this is not a WCAG conformance or independent-user certification.

No new dependency, auditor, runtime, service, provider or scheduler. Existing installer/updater consent, compatibility, safe-path, recovery and release reliability behavior is preserved.

v1.3.29 — Visual and partial evidence

Screenshots, recordings and incomplete descriptions now have focused guidance for building a useful brief from limited evidence:

• Preserve source/frame/crop and actual timing limits; distinguish visible facts, narrative claims, inferences, unknowns and proposed behavior.
• Keep visual similarity, demonstrated interactions and functional/file/protocol compatibility separately assessable.
• Ask about gaps that change the result while continuing useful reversible work within the requested scope and mode.

Includes a bounded synthetic Notes-panel example with frozen inputs, explicit missing/cropped states and fixed structured checks. The example is same-author/source-exposed, not an actual recording or independent model benchmark.

No new runtime, dependency, provider or service. Existing installation, optional integration, weekly-check consent and compatibility behavior are preserved.

v1.3.28 — Runnable reconstruction handoffs

Runnable results now receive proportional prerequisites, setup/run/stop, walkthrough, data and recovery guidance, with a check outside the development checkout. Function and specification-only tasks retain their existing scope.

The existing Pocket Tasks example includes a standalone PowerShell walkthrough for save/reopen, cancel, command errors, collision recovery, backup and reversible reset. It uses existing Node and no packages/services. Its same-agent fresh-directory exercise is not a clean OS or independent user test; local behavior does not imply remote or global parity. Candidate code and historical reference evidence are unchanged.

Installer, updater, consent, integration and compatibility protections remain unchanged. Release/source/package/installation/upgrade verification and the existing public CI/CodeQL gates are required for this release. No new runtime dependency or automatic software acquisition is introduced.

v1.3.27 — Unavailable backend handling

I added focused handling for inaccessible backends: distinguish observed client behavior from unknown server logic, select an explicitly labeled local substitute, supplied finite replay or deferred integration, and keep hard real-service requirements pending until verified.

The new runnable synthetic notes example uses the existing checker, with 23 frozen scenarios for local save/reopen, unavailable versus empty, malformed responses, replay mismatch and pending synchronization. Six deliberate misleading structured completion/observation mutations are rejected. This source-exposed same-agent example makes no live-provider, remote-persistence or production integration claim.

Installation, optional updates, consent, safe paths, recovery and existing compatibility behavior remain unchanged. The release updates skill guidance and adds a conditional reference without new dependencies or services.

v1.3.26 — Original-app quirk decisions

Original-app quirks now have focused preserve, improve or defer guidance. Reconstruction decisions link observed evidence, user goals, affected consumers and acceptance, while keeping inferred intent and intentional differences separate.

• Preserve required file/protocol compatibility; improve within authorized goals after checking dependencies; defer visibly with an unresolved question or revisit trigger.
• Retain scope, useful-first-slice priorities and mode boundaries without compulsory per-quirk approval gates. Planning alone does not authorize implementation.
• Add a worked mixed-profile example and 14 bounded same-agent manual scenarios, clearly separated from executable tests. Two selection-checker regressions challenge unchanged pin ordering and hidden-pin retention.

Installer/updater/integration, consent, safe paths, recovery, historical oracles and CI safeguards are unchanged. No new dependency, model API, agent, paid service or scheduler. Local denied full-suite preflights remain incomplete; supported public CI and exact artifact checks are required before publication. Existing host sessions may need to reload the updated skill.

v1.3.25 — Useful reconstruction slices

Broad reconstruction requests now start with the first useful user outcome and a short sequence of coherent working slices. Hard requirements remain binding; shared dependencies and unknowns that could change the design are distinguished from optional polish and deferred work.

• Improved journey selection, dependency ordering and evidence-linked acceptance guidance, with proportional loading and corrected-priority resume behavior.
• Added a save-first worked planning example and 14 manually assessed scenarios covering conflicting priorities, offline/file compatibility, unavailable backends, corrections and varied wording. These are same-agent, source-exposed teaching material, not a model benchmark or time-savings claim.
• Updated the usage example. Understand/future-build requests and narrow function scope retain their boundaries; a plan does not authorize implementation.

Installer, updater, integration, consent, path safety, historical fixtures and CI safeguards are unchanged. No dependency, model API, paid service, agent or scheduler is added. Validation and artifact identities are verified before publication; local denied preflights remain separately recorded limitations. Existing host sessions may need to reload the updated skill.

v1.3.24 — Measured optimization worked example

I added a runnable measured optimization example using the existing Pocket Ledger reconstruction. The protocol was frozen before implementation; literal correctness oracles, all raw paired timings, runtime/workload/source identities and executable evidence checks are retained.

The single-pass Unicode candidate passes the selected behavior tests but does not demonstrate an overall speed benefit. Typical and stress medians were higher in both final processes, while small-input medians disagreed. I retained the original implementation and the unsuccessful candidate as a teaching example. Results cover only the recorded synthetic workloads and Windows Node runtime; no universal speed or memory claim.

The companion's performance guide now explains bounded experiments, equivalent work and output consumption, precise resource metrics, failed/inconclusive outcomes and complete raw-sample review. No new dependency, model benchmark or machine-sensitive CI timing gate.

Validation: five focused source/package regressions; full existing public CI across Node20/22/24 on Linux, Windows and macOS; source/extracted package, manifest, links/privacy, genuine skill validators, eight isolated install configurations and eight exact-ZIP upgrades. Windows source/package each487/487 with0 skips; Linux retains3 casing-related skips. Local full suites remained unrun after denied junction preflights, never treated as passes. Authorship and substantive review were same-agent and source-exposed, not independent.

The installer, updater, consent/alias/integration/recovery contracts, historical fixtures and README support links remain intact. Installed-skill maintenance preserves the selected settings and backups. Existing host sessions may need a refresh to load changed guidance.

The first tag run had one Windows Node22 CLI-capture timeout in the unchanged artifact harness (486 passes,1 failure; that job's package suite unrun). One retry of that job with identical code/tests/three-second timeout passed source and extracted package487/487 with0 skips and the same ZIP digest. The original failure is retained; its root cause remains unknown. A passing retry does not establish a harness fix. All other tag jobs passed on their first attempts; no benchmark was rerun.

Release ZIP:2,220,798 bytes; SHA256 fe6c890ae0b428af92ea0db0f90c7e81e2f5654550739baab68ef971012c56da. This exact local/Windows artifact is published; other platform builders may produce different compressed ZIP bytes while verifying the same manifest/payload.

v1.3.23 — test environment diagnostics

I added bounded test-environment diagnostics to the existing source and extracted-release test entry points. npm test now observes directory-link fixture setup before starting the unchanged suite. A denied or unknown preflight reports validation as incomplete; --run-tests explicitly runs the full suite when useful and preserves original TAP, errors, counts and exit status.

Actual link-setup failures retain their original errors and add a scoped diagnostic. Cleanup failures remain visible. Probe success does not guarantee later fixture success, and EPERM does not by itself establish an administrator requirement or explain unrelated assertions. No safety assertion or required gate is skipped or weakened.

Nine diagnostics regressions cover controlled supported, denied, unexpected and mixed failures, missing execution, cleanup/reporting failures and default Node discovery. Existing CI jobs and protections are retained, with no new dependency, service, retry engine, timeout increase or automatic remote dispatch. The installed skill payload is unchanged; this release updates repository testing tools and contributor guidance.

Windows source and extracted-package tests passed 482/482 with no skips. Review and final validation: PR #53. Review was same-agent, source-exposed. Local junction preflights remained incomplete; CI passes do not relabel those denials.

Exact Windows-verified ZIP SHA-256: fc8ba18685df08ded3153977015778600c0cc2327f1bea44c3cd721b576b7207 (2,194,126 bytes).

v1.3.22 — Investigation recovery boundaries

Investigation failures now stay tied to the selected question and checkpoint, preserving applicable completed evidence while distinguishing partial capability and precise blocked questions. Recovery follows REA’s existing rules; timeout and retryable flags do not establish absent functionality or justify unchanged retries, broader access, software acquisition or paid analysis fallbacks.

The fixed artifact-capture checker rejects contradictory raw failure/partial fields alongside matching streams and exits. Seven controlled process-failure regressions verify preservation of completed work and failed attempts without false completion.

The existing public CI matrix now verifies and tests the exact extracted release archive as well as source. No new runtime dependency, retry engine or service is introduced. Installer, updater, integration consent and read-only check behavior are unchanged.

Validation uses same-agent source review and bounded fixtures; it does not certify arbitrary prose, model compliance, whole-target equivalence or a universal cause of tool failures. Historical provider observations remain historical.