Delivery probe results: Claude Code, revision 4

• Harness and version: Claude Code 2.1.281, fresh headless sessions (claude -p --permission-mode default --output-format stream-json)
• Model: claude-opus-5-5 (1M context), the configured default; no override
• Operating system: Linux (Ubuntu 24.04, x86_64)
• uv and Python versions: uv 0.12.17; Python 3.12.8 and 3.13.1
• Tester (agent or person) and date: Claude Code agent (Opus 5.5), 2026-09-24, at the operator's request
• Probe package revision (see PROTOCOL.md): revision 4, at commit 69d58a6d

Result. Every case passed. Agents reached the library only by reading files, with no lookup commands. The one read rule that init wrote into the project's .claude/settings.local.json was enough for every read. The stub hop worked the same way in all runs: stub, then the real SKILL.md, then its relative links. One risk was found: before init reran after a switch to an editable install, a stub still pointed into the stale share/ snapshot, and the agent read old content with no warning (case 6).

Harness facts

• Instruction file it loads: CLAUDE.md, which here imports AGENTS.md and .cp-delivery-probe/library.md with @. An import of a missing file did not cause an error (case 8).
• Can the agent run shell commands? Yes. Read-only commands inside the project ran without approval; commands outside the project or with $ expansion needed approval, which a headless session cannot give.
• Can the agent read files outside the project, and under what setting? With a Read(//<root>/**) allow rule. In this run the rule sat in the project's uncommitted .claude/settings.local.json, the first test of a project-level rule. Without it, the Read tool and ls of the library were denied (case 2).
• Does it read project-level skill directories, and which? .claude/skills/. The stubs under .agents/skills/ were not needed by Claude Code.
• Can permission settings live in the project, and can a machine-specific part stay uncommitted? Yes: .claude/settings.local.json, which init also listed in .gitignore.

Cases

Every model case ran in a fresh session with the protocol's prompt. The raw stream-JSON logs were in the scratchpad and were deleted at cleanup.

Case	Result	What happened	Prompts or denials
0 Install and init	Pass	Python 3.12 install. Init wrote four stubs, library.md, the read rule, and a .gitignore block. --check: six ok.	—
1a Base layer, file imported	Pass	No skill stubs. With library.md imported by CLAUDE.md, the agent's first call read the instructions index by full path, then retire-widget.md and its linked shared-step.md. RETIRE-WIDGET-V1, SHARED-STEP-V1. 3 tool calls.	none
1b Base layer, file read by the agent	Pass	CLAUDE.md imported only AGENTS.md. The agent read library.md with a Bash cat, tried one $-expansion command (denied), then read the index and both files. Same tokens. 6 tool calls.	1 Bash denial, not blocking
2 Read permission	Pass	With init's rule: no denials in any case. With settings.local.json emptied: the Read and an ls of the library were denied, and the agent stopped and pointed at the missing rule. Init restored it.	Without the rule: 2 denials
3 Stub skill, run 1	Pass	Skill tool loaded the stub; the agent read the real SKILL.md in share/, then ../shared-step.md, ../../types/probe-note.md, and ./references/read-detail.md, each resolved from the real skill's directory. SHARED-STEP-V1, PROBE-NOTE-TYPE-V1, READ-DETAIL-V1.	none
3 Stub skill, run 2	Pass	Same trace as run 1.	none
3 Stub skill, run 3	Pass	Same trace as run 1.	none
4 Router through a stub	Pass	Project with no AGENTS.md or CLAUDE.md. The prompt selected delivery-probe-library without naming it; the agent read the real router skill, then calibrate-gadget.md. CALIBRATE-GADGET-V1.	none
5a Upgrade, same Python	Pass	Bump to V2 and reinstall. No init rerun: --check six ok. Fresh base and skill sessions reported V2.	none
5b Description change	Pass	After a description change and reinstall, --check reported exactly the two stubs of that skill stale, and cp-delivery-probe-library warned about them. One init rerun: six ok.	—
5c Python change	Pass	Reinstall on Python 3.13: six ok with no init rerun; a fresh skill session reported V2.	none
6 Editable install	Pass after repair; silent stale read before it	Editable install: --check reported all six outputs stale. The source was then bumped to V3 while share/ still held V2. Before repair, a fresh skill session followed the stub into share/ and reported V2, with no warning, because the agent ran no command that could warn. After one init rerun, the stubs, library.md, and the rule pointed at the source tree, and fresh base and skill sessions reported V3. Returning to a normal install: the command warned, one init rerun repaired it, and a fresh session reported V3 from share/.	none
7 Skill names	Observed	Claude Code listed the stubs as delivery-probe-library and delivery-probe-read, with no prefix.	—
8 Fresh clone before init	Pass	A copy of the test project with only AGENTS.md and CLAUDE.md. The session started without error despite the failing import. The agent found no library.md, did not guess a location, and told the user to run init.	2 Bash denials (checking whether the init command exists)

Modifications to the package

None. In the scratch copy, only the supplied tools/bump.py and the protocol's description edit (case 5b) changed files.

Consequences for the design

• Observed: stubs work in Claude Code. In five skill sessions (case 3 three times, 5a, 5c) and the router case, the agent followed the stub to the real skill and resolved its relative links from the real location, with no failed read.
• Observed: the generated library.md works as the base-layer route. Imported, it costs no tool call; read by the agent, it costs one.
• Observed: one Read rule in the project's .claude/settings.local.json is the only setting Claude Code needs.
• Observed: a normal upgrade and a Python change need no init rerun; a description change is detected and repaired by one rerun.
• Observed: after an editable/normal switch and before the init rerun, agents read the stale share/ snapshot silently. With no command in the agent's route, the check cannot warn the agent. Inferred: this risk is limited to install-mode switches, which developers make; the design should say so, and the developer procedure should rerun init right after a switch.
• Observed: a fresh clone without init fails safely: the agent says what is missing and does not guess.

Cleanup

Ran cp-delivery-probe-init --remove in both test projects and uninstalled the tool. No user-level setting was changed: ~/.claude/settings.json kept its SHA-256, and the user-level skill directories were untouched. The scratch copies and logs were deleted. No hooks were used, no credentials were touched, and the real llm-commonplace installation was not changed.