Référence CLI 1.9.2
Version documentée. La référence technique conserve la langue anglaise du dépôt source.
Revision: 2026-11-09
Qualification and known limits¶
This page distinguishes the CLI's behavior from the server's capabilities. A
working menu or a green unit test does not establish that every tool is enabled
for every account. Versions described in documentation can await npm approval;
check deposium --version and npm view @deposium/cli version separately.
What the current qualification checks¶
| Surface | Evidence and boundary |
|---|---|
| Command/reference coverage | 122 offline help invocations cover the 105 inventory declarations plus entrypoints/aliases; 42 Markdown files and 210 literal command examples were checked for known commands/options and broken local links. Help never proves an actual backend operation succeeded. |
| Developer demo | The source-only npm run demo:ui completes using current helper exports and synthetic fixtures on both Node versions; typecheck includes the demo and CI runs it. No server calls or model streaming are simulated as real observations. |
| Startup | Bare terminal menu, explicit aliases, offline help/version, scripts/CI/dumb terminals, unknown options and clean exit. Wordmark appears at entry, not inside results. |
| Main menu | Named space selection, failed/empty catalogues, shared Compound request contract, internal operation refusal and local navigation have CPU tests and terminal recipes. |
| Chat | Token delivery as received, activity cleanup, canonical replacements, HITL safe defaults, blocked answers, final-stream source receipts, verification notices and local source commands have regression tests. |
| Tool catalogue | Real ordinary-key catalogue: 15 entries with schemas. Full description/schema retained, including a 3,546-character description. Picker/read/back verified in a native terminal. Inspection does not execute a tool. |
| JSON | Existing space/files/search/corpus/tool tests plus dedicated general-AI and legacy subcommand contracts check parseable stdout, disabled spinners and noninteractive authentication. Mutating commands are exercised with fake transports, not by modifying production. |
| API scope | Ordinary authenticated reads are qualified separately from admin rights. An omitted tenant is server-resolved. Space choices do not grant document access. No cross-tenant exhaustive security qualification is claimed. |
| Delivery | npm staged artifact/source/hash, public version and latest are separate checks. A matching stable GitHub Release triggers docs synchronization; verify the Docs deployment afterward. |
The 1.7.2 qualification passed 878 tests across 57 files on Node 22.13 and 24. This includes 37 parameterized legacy JSON command cases, two Compound JSON cases, startup/renderer/branding checks and mutation-refusal coverage. A native terminal recipe opens the bare command with its logo, navigates and exits; a separate rendering fixture verifies formatted general-AI output without the old warning, repeated logo or confidence bar.
An ordinary-key read-only SaaS recipe also checked health, the authorized space catalogue, corpus counts, BM25 retrieval and one fresh document question. The answer matched the expected subject and supplied a final-stream citation, about 11 seconds in that single run. This is functional evidence, not a general latency benchmark or a claim that all production operations or roles were exercised.
Choose the right operation¶
- Chat: questions and synthesis using documents in an accessible space.
- Search: matching passages without a generated synthesis.
- Corpus statistics: server-reported counts, not a content summary or accuracy score.
- General AI (Compound): no selected document space; optional server reasoning/web tools.
- Query intent: classifies your question; does not summarize documents.
- Tools: reference descriptions/input schemas; selecting an entry does not execute it.
- MCP app resources: require a compatible host, not a terminal dashboard.
A model or tool name in metadata is reported by the server. It is not an independent trace proving every stage used that model. Zero reported tools does not attest a private-document search. The CLI does not supply local TurboV3/LLM engines by itself.
Still not delivered or universally qualified¶
- CLI 1.8.0 adds saved language and default-space preferences for the core flows described in Settings. Historical advanced-command output, server data and library keyboard hints are not all localized. Chat now sends the effective UI language by default.
- Some server paths deliver a whole answer in a burst. An activity indicator does not make generation faster or provide true producer streaming.
- Full narrow-table/resize/reflow behavior and standalone Bun binaries are not newly qualified across macOS, Windows and Linux by this Node release.
- Local interruption is not a receipt that remote work stopped. Help/examples do not imply production uploads, deletion, key rotation, code execution, monitor mutation or paid evaluations were run during this qualification.
- Backend-dependent operations can require additional permissions/features. A server refusal remains an error; internal routes are not made public to make an old CLI command pass.
See the command reference, interactive guide and release procedure. Historical development/audit records retain their original dates and test counts; they are not current release claims.
Preferences qualification — CLI 1.8.0¶
919 tests / 62 files pass on Node 22.13 and 24, including the existing
credential, JSON, HITL and canonical-answer tests. A real terminal recipe verifies
French startup → Settings → Back → clean exit without changing the user's
preferences. An ordinary-key read-only probe observes 22 authorized spaces and
confirms access to the selected recipe space via a fresh catalogue, with Health
ok and the requested French invocation language. No model call or business-data
mutation is needed for this preferences qualification.
The static documentation check covers 126 offline help invocations, 43 Markdown files and 221 literal examples, with no unknown command/option or missing local file target. This is distinct from executing every backend operation.
Language priority, locale precedence, automatic mode and invocation-only overrides are checked independently. Process tests use isolated temporary homes to verify persistence across launches, encrypted storage and unchanged synthetic credentials. The main menu also tests offline Settings, help/back/exit without authentication, API-only client initialization and avoiding duplicate Chat authentication. Process tests exercise real command parsing, offline settings, translated help and clean JSON, including invalid flags and conflicting scope options. No real user preferences are modified by those tests.
Catalogue fixtures cover authorized, revoked, empty and failed lookups, explicit
versus saved scope, environment precedence, noninteractive refusal and preserving
prior settings on failure. Health JSON without --context retains its existing
contract; explicit context distinguishes unavailable catalogue data from endpoint
health and strips credential-bearing URL components. Chat tests verify the actual
request language with the selected scope/history retained. Catalogue tests check
translation coverage and interpolation slots, without translating result data.
Live command input — CLI 1.9.0¶
933 tests / 63 files pass on Node 22.13 and 24. The terminal-stream fixtures exercise actual readline key events: prefix filtering,
Up/Down selection, Tab without submission, Enter, exact /source, Escape,
Unicode cursor editing, command arguments, unknown commands, bracketed multiline
paste (including insertion mid-question and pasted tabs), abort/Ctrl+C cleanup and terminal resize. Catalogue pagination is bounded;
all main-menu entries fit a normal 24-row terminal. Existing scope/history, HITL,
canonical answer, credentials and JSON regression suites remain required.
These CPU tests do not establish a production-model latency gain or native Windows/Linux terminal behavior. A standalone terminal recipe can inspect these local commands without sending a model request. The existing activity indicator and the server's actual token delivery are unchanged by completion.
Search display regression — CLI 1.9.1¶
An authorized read-only search returned twenty passages and roughly 130 kB of business data. In 1.9.0 the generic table renderer produced no output before a 2.5-second child-process deadline. With passage presentation, replaying the same private response finishes in about 105 ms, exit 0, with 12.7 kB of terminal output. This is one display replay, not a server latency or retrieval-quality benchmark. No private document text is committed. Synthetic large-Unicode fixtures run in bounded child processes so a rendering regression cannot freeze the test runner. JSON preserves the original response; preview limits do not change retrieval.