D14 publication¶

Measured 2026-09-12, HEAD 4535839bb47d27456ccbafea0170bc7f17b88b71, with the current working-tree root hash f53831450e7f0a5457966e83a16f8be7b6fdb016b255e58946211109cdffc1ac. Native read-only codex debug prompt-input and app-server config/read / hooks/list probes used temporary SQLite state. All requests completed successfully. No model task started; no repository, configuration or trust changes were made.

The 42,442-byte root now arrives whole. Subtree truncation remains. The effective project-document cap is still 65,536 bytes, applied cumulatively to source bytes. Blank-line separators add two bytes per join to the rendered inner block.

Start	File	Disk bytes	Included bytes	Result
Root	AGENTS.md	42,442	42,442	Whole
src/gzkit/commands	AGENTS.md	42,442	42,442	Whole
src/gzkit/commands	src/AGENTS.md	7,458	7,458	Whole
src/gzkit/commands	src/gzkit/AGENTS.md	22,651	15,636	Partial: 7,015 bytes missing
src/gzkit/commands	src/gzkit/commands/AGENTS.md	17,505	0	Not reached
tests	AGENTS.md	42,442	42,442	Whole
tests	tests/AGENTS.md	30,284	23,094	Partial: 7,190 bytes missing

Commands-start source bytes total 65,536; joined inner block is 65,540 bytes. Tests-start source bytes total 65,536; joined inner block is 65,538 bytes. Root-only inner block is 42,442 bytes.

The command chain still ends inside task-discovery material in src/gzkit/AGENTS.md, at the sentence beginning Every reader derives from TaskId.parse. The test chain now reaches task lineage material and ends in a reference beginning docs/governance/agent-contrac. The JSON retains exact tails and per-file hashes; it contains no full raw prompts.

Comparison boundaries¶

The original audit observed 48,511 bytes. The parent reports an independent working-tree reversion before publication reduced the immediate prepublication root to 47,851 bytes. The approved publication therefore reduces that immediate baseline by 5,409 bytes, not 6,069. The net historical-to-current difference is 6,069 bytes, including the unrelated 660-byte change.

Compared with the historical native probe, source-parent delivery increases from 9,567 to 15,636 bytes and test-file delivery from 17,025 to 23,094 bytes. Those are net historical comparisons, not isolated causal measurements of this publication. Commands-specific delivery remains zero.

The native skills-instructions block now measures 22,172 bytes. Its earlier 22,179-byte figure is historical; this audit makes no causal claim about that seven-byte change and does not change any skill bodies.

Current hooks: zero recognized project handlers¶

hooks/list currently returns an empty project-hooks list, with no project errors. The only returned warning concerns a plugin SessionEnd timeout; it is not a project-handler error. This differs from the historical observation of three registered, untrusted project hooks.

The current .codex/hooks.json is 592 bytes and has the older command-array / inject shape under SessionStart and UserPromptSubmit. Native registration is therefore absent in the measured state. Project .codex configuration remains active and features.hooks=true; enabled infrastructure does not establish a recognized handler. No automatic orientation or handoff-context delivery is claimed.

The prior three-handler trust result and prior orientation stdout size must not be presented as current. This probe did not run any hook, restore the earlier integration, or alter trust.

Limits¶

These measurements witness the installed CLI's startup instruction assembly for three starting directories. They do not measure the running desktop conversation's complete context, refresh behavior after later shell-directory changes, token usage, model obedience, or a Claude context-window ratio. Disk existence and root-only delivery are insufficient evidence for subtree instruction survival.