0.28 release readiness review

This is the historical 0.28 assessment. The prepared 0.29 follow-up adds controlled latency measurements, twelve repeated edit tasks and recoverable pin replacement. Its evidence updates the performance, outcome and migration limits recorded below.

The September product review asked for trustworthy pre-edit context, quieter prioritization, recoverable decisions, useful agent integration and maintainable human/agent documentation. The 0.28.2 follow-up rechecked those outcomes against first use, upgrades and the published surfaces, after a user found that the skill refresh mechanism had no discoverable update path.

Review coverage

Accepted outcome	Evidence and follow-up
Context agrees with scan	Shared analysis and existing context regression tests cover authority/agent configuration, claims, importing tests and excluded/partial targets. Hook delivery and root selection were missing from this check; 0.28.2 repairs both.
Default signal outranks repetitive noise	Bounded file ranking and conservative/default-off choices from 0.28 remain. Ranking uses the same corpus and clock; no detector or scoring changes are included in this patch.
Prior decisions survive upgrades	Pin migration preserves metadata and refuses stale plans/collisions. CI/suppression guides still encouraged blanket re-pinning; that advice is corrected. Migration’s multi-file transaction limit remains explicit.
Skills work in installed hosts	0.28.1 exercised fresh host discovery and actual scoped edits. This patch adds discoverable maintenance and real-terminal upgrade checks, and retains customizations.
Hooks deliver useful evidence	The old command printed plain PreToolUse output, used an incorrect timeout unit and could fetch independently through npm. This patch fixes the host envelope, root, evidence/status presentation and offline executable selection.
JSON is a usable public contract	A deeper check found stale declarations, hardcoded feedback versions and unversioned/prose triage output. Runtime envelopes are corrected and public type declarations are now generated and checked.
Humans and agents find consistent docs	Current workflow guides replace obsolete instructions. The built site had broken relative Markdown links; conversion and internal-link verification now cover the rendered result.
Improvements are measured honestly	Unit/integration checks, installed-package upgrades, host delivery and ranking answer separate questions. None is presented as proof of general agent productivity or defect reduction.

Remaining limits

• Agent outcome benefit is unproven. The 0.28 paired edit trial tied at 3/3 acceptance passes per arm. Broader tasks, repetitions and controlled model selection are still needed. A successful host integration check does not fill that evidence gap.
• Context latency remains material. Context and the hook analyze the repository. The 0.28 self-repo observation rose from 3.52s to 8.99s; caching/shared parsed-input reuse was explicitly deferred. No performance improvement is claimed for 0.28.2. Hooks remain optional.
• Heuristic analysis is incomplete evidence. Import/test discovery is not executed coverage. Universal analysis does not imply a language parser covered every file. Ranking labels remain partly type-level and the deep set is dominated by one fixture.
• Pins need judgment. Absence does not prove resolution. Migration replaces each selected file atomically but is not a transaction across all pin files; retain a commit or backup before applying a reviewed plan.
• Host behavior can vary. A host can fail to activate a skill or omit its instructions. Changing a file on disk does not retroactively reload a skill already read in the current conversation. Restart/reload when needed.
• Distribution scope remains npm. Homebrew, standalone/Windows binaries, additional languages, crimes ask, hosted services and paid tiers remain deferred. Adoption and retention are not established by download counts.

These are explicit limits, not completed goals. The release is intended to be shareable as a local evidence tool, with accurate instructions and working integration paths. 0.28.2 release notes record validation.