[Helweg]

Summary

Deliver the approved agent-reliability work across the shared operations, human CLI, MCP, OpenCode, Pi, and omp:

• Explicit source freshness: index_status({ checkFreshness: true }) / ocbi status --check-freshness --json, separate from catalog readiness and coordinator activity. Default status remains cheap; inspection does not refresh indexes.
• Source-linked pre-edit tests: opt-in codebase_edit_context({ includeTests: true }), reusing AST-backed JS/TS direct-target and one-hop caller discovery. Complete budgeted citations and selected-only structured metadata; static links are not runtime coverage.
• Current indexed working-tree impact: pr_impact({ workingTree: true }) / ocbi impact --working-tree --json for staged, unstaged, and untracked files, with explicit refresh requirements, bounded local Git status, provenance, and UNKNOWN for incomplete graph evidence.

Correctness and review

• Resolve canonical native symbol IDs, then match exact target paths and declaration identity. Large declarations start at their signature rather than a ranked middle-body slice; missing source is withheld. Cap source blocks at 512 tokens to preserve room for relationships/tests. Production and evaluation share the same per-indexer lookup.
• Keep selection-hook test references metadata-only, count retained evidence, reserve final heading/truncation overhead for closed whole-line source blocks, and propagate cancellation after empty lookups.
• Read structural architecture files from the SQLite catalog without requiring vectors.
• Reject concurrent Git/source/index changes, including database republication with unchanged source bytes/catalog counts. Reject invalid UTF-8 paths; retain only local rename endpoints for nested projects.
• Native-backed regressions cover real source/citation identity, CRLF/Unicode, tight token budgets, scope, freshness, host surfaces, and working-tree graph transitions. An independent pre-review control reproduced incorrect suffix-file source selection and undetected concurrent graph publication.
• Documentation and Unreleased changelog updated. No version bump, storage migration, installed-extension modification, merge, or release.

Verification

• npm run build, then npm run build:ts for final TypeScript-only repairs; Rust source unchanged.
• npm run typecheck
• npm run lint
• Final full coverage suite with short temporary roots: TMPDIR=/tmp npx vitest run --coverage — 3,082 passed, 7 skipped; 154 files passed / 1 skipped on macOS ARM64, Node 26.12.1. The initial full gate also ran npm run test:run.
• npm run smoke:package — both package identities; strict standalone and SDK-interoperability declarations, ESM/CJS loading, CLI aliases and MCP initialize/tools-list.
• Actual final built CLI and stdio MCP scenario: cached/current/stale/refreshed freshness, stale-impact rejection, indexed dirty-tree symbols and transitive callers, source-linked metadata-only test references, preserved HEAD and Git status.
• Final built stdio MCP canonical-source scenario: exact target beats a same-named suffix file, a fragmented 240-line declaration starts at its signature, test references remain metadata-only, and missing source receives no source credit or fake body.
• Final built stdio MCP minimum-budget scenario: 16/64/96-character identities with actual native callers/callees, CRLF/Unicode sources and 128-token responses retain closed whole-line source blocks, and source credit matches the visible body.
• Final isolated pre-edit evaluation, unchanged budget: Hit@5/MRR@10/graph-neighbor recall 1.0; response average 478, p95 707.7, max 734 tokens; p95 latency 213.1 ms. Inspected every query against the retained baseline:

Query	Baseline tokens	Final tokens	Target hit retained	Expected graph neighbor retained
Resolver + caller	425	471	Yes	Yes
Large evaluator target	360	734	Yes; now starts at canonical signature	Not requested
Unresolved fallback	229	229	Yes	Not requested

Linux-only invalid-byte filename coverage cannot run on macOS/APFS and is left to Linux CI. The pre-edit gate caught and drove repair of declaration-source loss and excessive source-block size; failed reports were preserved, and thresholds/baselines were not relaxed. These deterministic-mock evaluations validate harness/routing behavior, not semantic embedding quality. CI must verify the final PR revision on supported Node/platform jobs before merge.

The initial CI matrix exposed source-fence clipping during final response fitting and a path-length-dependent graph-omission fixture. Both were repaired with failing-before/passing-after source regressions and a caller identity that explicitly exceeds the small response budget. The corrected revision must complete fresh CI; this is not an unchanged rerun.

Release Notes

Category: feature. Semver: minor. This PR is for review and CI only; do not merge or publish as part of this task.

[Helweg]

Validation complete for 3d74979. All 15 PR checks are SUCCESS, including Node 22/24 CI, all six platform/path jobs, CodeQL, and release labels. Fresh CI: https://github.com/Helweg/open-codebase-index/actions/runs/40674544143 . The initial red run was repaired in code, not retried unchanged. Local final acceptance: 3,082 tests passed with coverage and short temporary roots; both packed package identities passed; actual built stdio MCP verified closed whole-line source and correct credit at 128 tokens with long identities and real graph neighbors. The unchanged pre-edit gate passed all three queries with Hit@5/MRR/graph-neighbor recall 1.0 and p95 response size 707.7 tokens; all per-query deltas were inspected. Mock evaluation is routing/harness evidence, not semantic embedding-quality evidence. Review workers are settled/released. PR remains OPEN; no merge, tag, publication, or installed omp-extension change.