[DmitrL-dev]

Summary

• File views stay inline. cat, nl, sed, head, tail, and git show/diff/blame outputs up to 64 KiB are no longer packed. In a recorded 0.1.10 session, 3 packed reads (two file views, one rg) caused 5 extra artifact reads in smaller ranges, and the agent started splitting its reads to stay under the 6 KiB threshold.
• Artifact reads are never repacked. A command that reads a receipt artifact now passes through unchanged. The read is counted as artifact_rereads/artifact_reread_bytes, and --report adds net_saved_bytes = saved_bytes - artifact_reread_bytes.
• More test runners stay inline on unknown status. Unknown-status output from cargo/go/npm/pnpm/yarn/bun test, jest, vitest, Gradle, Maven, and similar runners is no longer packed, matching the existing pytest/unittest rule.
• Verifier parsing: 2>&1-style descriptor duplication is accepted. Go word flags (-short, -bench, -shuffle) are no longer read as help requests.
• Redaction:
  • indented diagnostics after an Authorization header are kept;
  • an unterminated private-key marker redacts only its key-like lines, not the rest of the output;
  • env-style secret names (AWS_SECRET_ACCESS_KEY, GITHUB_TOKEN), common token prefixes, and URL user-info passwords are now redacted;
  • matching is linear on long token-like runs.
• Verification debt: prose files (.md, .rst, .adoc) under source directories no longer create debt, and neither do failed apply_patch results, including Exit code: N headers.
• test_hook_commands skips the generator tests when scripts/generate_hook_commands.py isn't packaged with the installed plugin.
• Version 0.1.11+codex.20260928; CHANGELOG, README (EN/RU), SKILL.md, and docs/savings.md updated.

hooks.json and sol_bootstrap.py are unchanged, so the pinned loader protocol and the hook trust hash stay the same.

Verification

• python -m unittest discover -s plugins/sol-codex/scripts -p 'test_*.py': 71 tests, OK (7 Windows-only skips), Python 3.9 on macOS. New tests cover every change above.
• scripts.test_installed_hooks, scripts.test_upgrade_cache, and scripts/benchmark_receipts.py pass.
• The sanitizer handles 200 KB token runs and 5,000 unterminated key markers in under 0.5 s. An intermediate version took 21 s on 13 KB; a regression test now covers this.
• scripts/validate_repository.py and scripts.test_release_tools fail with the same 14 errors and 7 failures on clean main. They come from tracked research data and are handled in a separate PR.

Not changed on purpose

• Verifier wrapping stays limited to Bash in bypassPermissions.
• Env-assignment prefixes (FOO=1 pytest) are still rejected.
• The global threshold still overrides the Astra threshold, as documented.
• Code-mode receipt events are still counted: the hook payload does not identify code mode.

🤖 Generated with Claude Code