Costs, Limits and Verification
A useful demonstration should let readers reproduce its claims and identify their boundaries. This series points to the maintained QuietCli module in FrenchExDev: its complete ZIP, executable companions and verification runner. The published source manifest and reports establish a traceable relationship between the explanations, the packaged files and the code that was inspected.
They serve a narrower purpose than a comparative benchmark. The examples check contracts and failure behavior. They do not establish that one implementation consumes less CPU, less memory or fewer billed tokens than another. RTK is discussed from official sources, while the execution report explicitly records RtkExecuted: false and Benchmark: false.
Reproduce the example checks
Download the complete module ZIP (ouvre dans une nouvelle fenêtre) or use a FrenchExDev checkout. Run its verification launcher (ouvre dans une nouvelle fenêtre) in PowerShell 7.4+ Core. Git must be available, and the supplied long-job example needs Python:
$moduleRoot = 'C:/tools/QuietCli'
& (Join-Path $moduleRoot 'scripts/Invoke-Examples.ps1') -QuietCliRoot $moduleRoot -Python python3An alternate source checkout can be selected without editing scripts:
$moduleRoot = 'C:/code/FrenchExDev/FrenchExDev_i2/PoSh/FrenchExDev/FrenchExDev.QuietCli.PoSh'
& (Join-Path $moduleRoot 'scripts/Invoke-Examples.ps1') -QuietCliRoot $moduleRoot `
-Python 'C:/path/to/python.exe' `
-ReportPath './verification-local.json'The runner constructs a temporary Git repository, executes the scenarios and removes its temporary directory after checking the cleanup boundary. It writes a detailed report and emits a compact summary. The published verification report is an observation from its recorded environment; a new run can produce different durations and serialized byte counts.
For the published run, host Python was unavailable. The streaming case therefore used -PythonImage localhost/frenchexdev-pandoc-pdf:1, an image already present locally with Python 3.13.5. The adapter runs the original fixture with no network or image pull and mounts only its fixture directory read-only. This is a recorded environment choice, not a prerequisite to download that image. Use your own verified local Python image or a native interpreter; the report distinguishes the adapted run from the unchanged native Python companion.
To repeat the recorded execution on the author's machine with that existing image:
& (Join-Path $moduleRoot 'scripts/Invoke-Examples.ps1') -QuietCliRoot $moduleRoot `
-PythonImage 'localhost/frenchexdev-pandoc-pdf:1'What the checks establish
| Scenario group | Evidence required |
|---|---|
| Document analysis | Valid JSON produces the expected count; malformed JSON fails. |
| Native execution | A deliberate nonzero exit remains a failure; a timeout remains distinct. |
| Internal limits | Capture overflow and oversized records do not become complete answers. |
| Final budget | JSON fits the requested limit and reports removal of domain data. |
| Task aggregation | A consumed failed step remains visible in the enclosing task. |
| Unknown values | An unavailable count remains null, rather than becoming zero. |
| Git semantics | Clean, modified, directory-entry, rename and detached cases behave as documented. |
| Reuse and streaming | Composition preserves the inner result; both noisy-job stages establish their totals. |
These scenarios protect concrete interpretation boundaries. They are intentionally more meaningful than tests that reproduce the implementation line by line. Passing them still does not exhaust every external command format or every possible scheduling condition. A new companion should add evidence for its own domain and supported inputs.
The separate site report records chapter lengths, local links and the scoped navigation update. The browser report checks nine rendered pages, expansion and collapse, and downloadable artifacts. The PowerShell syntax report covers the published code blocks and authored scripts. These checks establish publication behavior rather than comparative performance.
Identify the sources
The source manifest records QuietCli's module version, archive hash and SHA-256 hashes of the files packaged from the workspace. The archive builder (ouvre dans une nouvelle fenêtre) includes the runtime, skill, examples, fixtures and launcher. Code links open GitHub rather than a separate blog copy, so maintenance has one source. GitHub links follow the moving develop branch; the retained package manifest identifies a particular download independently of later branch changes.
At the initial inspection of commit 50dc14e, the README referred to absent PROMPT.md and docs/STACK.md files, and the skill cited unbundled helpers. The distribution update replaces those routes with existing resources and explicit module dependencies. These historical gaps remain recorded in the source notes. The README (ouvre dans une nouvelle fenêtre) and skill (ouvre dans une nouvelle fenêtre) now describe the complete module download. The distribution report checks extracted files and execution separately from syntax and site rendering.
External references were checked on 12 October 2026: the RTK README (ouvre dans une nouvelle fenêtre), its architecture documentation (ouvre dans une nouvelle fenêtre), and the Git status format (ouvre dans une nouvelle fenêtre). The RTK links target a moving development branch. Their inspection date is recorded; they are not represented as immutable source snapshots or locally executed behavior.
Separate three kinds of cost
Initial cost includes learning the CLI format, defining properties, writing callbacks and checking assumptions. Agent-generated code still needs that work. If a simple native query already supplies a short sufficient answer, creating a companion can add unnecessary effort.
Repeated execution cost includes native work, interpreter startup, parsing, allocations, sanitization and serialization. The composition example starts a second PowerShell process to demonstrate reuse explicitly. That has a real cost even though its final answer is small. No timing advantage is claimed from choosing that example structure.
Maintenance cost includes changes in tool formats, counting scope and domain requirements. Reusable code can reduce repeated authoring effort, but only while it continues to satisfy the contract. An old source hash helps identify the inspected implementation; it does not certify compatibility with every future CLI version.
Measure only what was measured
The output limit counts UTF-8 bytes in a final JSON document. It does not count model tokens exactly. It also does not include all prompt text, model output or conversation history. A smaller result can be a useful ingredient in reducing context, while the total session effect depends on the complete workflow.
A comparative study would need equivalent inputs, preserved-information criteria, tool versions, repeatable timing conditions and a stated token-counting method. It would also need to account for authoring and reuse separately. This series supplies working examples and verification evidence so that readers can assess practical fit without substituting unmeasured performance claims for that evidence.