Repository navigation
v0.1.22
Context Goblin 0.1.22 fixes unsafe source-link reads and stale project caches, and makes benchmark reporting stricter.
- Resolve source and metadata paths before reading; exclude links outside the workspace and links into denied paths.
- Detect source additions, edits, and deletions without Git, including allowed source-link targets. Track ignored source files in Git projects as well.
- Bound fallback/ignored source scans to 2,000 files and 32 MiB; treat unverifiable freshness as stale. Regenerate caches from older formats.
- Cache package script names and the presence of AGENTS.md instead of script bodies, instruction contents, or test descriptions. OpenCode remains responsible for loading native project instructions.
- Keep context_goblin_get as the recommended single-call flow; remove the conflicting four-call recommendation from the README.
- Run integration checks with the pinned OpenCode CLI. Add timeouts, stop model runs after provider errors, preserve completed reports, and reuse only matching-version benchmark data.
- Require every requested model to pass compatibility and all efficiency checks. Count only distinct files read through the read tool, and explain that the 6/6 answer score measures checklist coverage.
Validation: 72 local tests passed, typecheck/build/report checks passed, OpenCode 1.18.20 integration smoke passed, and the npm package dry run passed.
Fresh one-shot A/B on OpenCode 1.18.20:
| Model | Files saved | Uncached input saved | Total event tokens saved | Compatibility | Efficiency |
|---|---|---|---|---|---|
| gpt-5.5 | 53% | -28% | 7% | pass | fail |
| gpt-5.6-sol | 47% | 51% | 34% | pass | pass |
The gpt-5.5 input increase is reported as a failure. These single-run results do not establish repeatable savings across both models; the repeated stability benchmark remains deferred. Detailed answers and accounting are included in the repository reports.