Current native delivery — 2026-09-19¶
Persona: main-session. Work order: GHI #1046. Baseline:
da741ef0d6c924c4b6b62bd75f8c1431faddcae2.
All four sampled native instruction chains arrive whole. Installed
codex-cli 0.155.1 rendered fresh prompts with codex debug prompt-input from
the root, commands, tests and governance-docs directories. No model task was
started. Existing trust/configuration were retained; only sqlite_home was
redirected to temporary storage. Raw prompts remain temporary, not published.
| Starting directory | Source bytes | Native inner-block bytes | Exact full chain |
|---|---|---|---|
| Root | 19,872 | 19,872 | yes |
| src/gzkit/commands | 55,368 | 55,374 | yes |
| tests | 35,959 | 35,961 | yes |
| docs/governance | 27,331 | 27,335 | yes |
Inner blocks include two newline separator bytes between files. The comparison checks equality with the complete ordered source chain, not a substring or a size threshold. delivery.json retains per-file hashes, offsets, byte counts and whole-chain equality, plus hashes of native prompts.
A separate invocation overrode project_doc_max_bytes=12000 on the command line
only. It returned exactly the first 12,000 source bytes, and full-chain equality
failed as expected. This negative control verifies the measurement detects
truncation. Neither the configured repository cap nor trust settings changed.
The September 12 audit observed commands/tests truncation on CLI 0.154.0 with a 48,511-byte root. Current root is 19,872 bytes. The old truncation finding is historical, not a current failure in these sampled paths. Both repository and CLI changed, so this is an observed before/after state, not an isolated causal experiment.
Every current probe also carries a 22,187-byte native skills-instructions block. That is catalog/discovery material, not the sum of available skill bodies. Do not compare it with the repository-only Claude metadata estimate as though they measured the same consumer or with a byte/4 estimate of model tokens.
This witnesses installed CLI prompt assembly. It does not inspect the complete active desktop conversation, establish automatic hook execution, prove native Claude delivery, test instruction refresh after a tool changes directory, or show that a model obeys the delivered instructions. Existing hook registrations and source-size measurements are not substituted for those observations.