Status: implemented (same PR as this spec). F6's upstream issue: doiteng/omni#62101, covering both the list-contracts 403 and the missing currency/unit metadata.
Source: Alfredo's feedback DM (Slack, 2026-08-25, thread p1787676307172329 + the API-key thread p1787676758296559), tested against the shipped dci ai session. Every claim about existing code cites the function and file it is based on.
Scope: six small, independent changes to the dci ai surface plus one upstream issue to file. No change to non-interactive agent mode, pipes, or CI behavior except the new one-shot verbosity flags (F3), which only restrict output.
Not an issue (no action): the agentic customer-context switch ("What's driving this posthog jump?" → customer context switched: csp.doit.com → posthog.com) was called out as working well. The 403 handling in F6 is also evidence the harness degrades gracefully — both stay as-is.
Feedback: "when I look for a slash command, hitting enter won't select it. just with tab. (this behavior is different from other CLIs, like claude)"
Today: in the main input key handler, enter always submits the raw input (ai_tui.go:824 → m.submit()), while tab is the only key that accepts the highlighted completion (ai_tui.go:827-834). With the popup open and /anom typed, Enter submits the literal /anom, which fails, instead of selecting /anomalies.
Change: when the completion popup is visible and the trimmed input is not already an exact command name, enter behaves like tab: accept the highlighted completion into the input (with trailing space) and keep editing. A second Enter submits. When the input exactly matches a command (with or without args typed), Enter submits as today — so /quit⏎ and /anomalies list⏎ never need a double-Enter.
This matches Claude Code's picker: Enter and Tab both accept; Enter never submits a prefix that the popup is actively correcting. Up/Down keep moving the highlight (ai_tui.go:836-863), Esc keeps dismissing.
Tests (ai_tui_test.go): popup open + partial input + Enter → input becomes selected command, no submit; exact command + Enter → submits; no popup + Enter → submits (unchanged).
Feedback: "a terminal 'beep' when it returns the result would be helpful."
Analytical turns run 1–2 minutes (opus thinking dominates); users tab away and want a completion signal.
Change: on TurnDone (ai_tui.go:628), the TUI writes BEL (\a) — the terminal maps it to its own notification (sound, dock bounce, tab highlight), which is exactly the Claude Code convention. Details:
- Only for turns that actually did work: skip the bell when the turn produced no tool calls and finished in under ~3s (echo-fast turns would make
/helpbeep). /belltoggles it; the choice persists inai_settings.jsonnext to the key and model (same read/write path asloadAISettings,ai_command.go:65). Default on.- One-shot mode: no bell ever — stdout may be piped, and BEL in a captured stream is corruption, not UX.
Tests: TurnDone with tools → frame output contains \a; fast toolless turn → no bell; /bell persists across a settings round-trip.
Feedback: "dci ai 'query' could have a flag to be verbose/non-verbose (not outputting the whole investigation)"
Today: one-shot verbosity is auto-detected — thinking, tool starts/results, and context switches stream to stderr iff stderr is a TTY (runAIOneShot, ai_command.go:70). A human watching always gets the whole investigation; there is no way to opt out short of 2>/dev/null (which also eats real errors).
Change: two flags on the ai command, honored only in one-shot mode:
--quiet(long form only — restish's global--rsh-queryowns the-qshorthand, and pflag panics on the redefinition at parse time, which brokedci aiin v2.6.1): suppress the investigation narration (thinking deltas, tool start/result lines, context-switch notes) even on a TTY. Errors, the destructive-approval verdict line, andDCI_AI_STATSoutput still print — those are contract, not narration.--verbose: force narration on even when stderr is piped (useful when capturing a full investigation transcript to a file).
Both flags set, or either combined with interactive mode → usage error. The default stays the current TTY auto-detection, so nothing changes for existing callers.
Tests (ai_command_test.go): flag plumbing → the verbose bool passed to the event loop; --quiet still prints event.Error and the approval verdict.
Feedback: "Not showing unit symbol or indicating the currency on tables" — the screenshot shows an AI-composed markdown table (Service | Jun | Jul) with bare numbers in inconsistent scales (936,800 next to 1,278.4K), no currency, no unit.
Where the numbers come from: that table was written by the model, not by our renderer — the deterministic table path already renders money correctly when it can (renderCellText/cellCurrency/formatMoney, main.go:2827-2875; pivot money-column marking, pivot.go:187-191). The gaps are:
- The model has no rule about units. The system prompt (
aiSystemPrompt,ai_prompt.go) tells it to be concise and table-sparing, but says nothing about labeling numbers. Add a prompt rule to the stable prefix: every numeric table or figure states its unit — currency code/symbol for cost (taken from the query result'scurrencyfield, never guessed), and the metric's unit for usage; use one consistent scale per column (raw, K, or M — never mixed) and say which. Prefix change invalidates the prompt cache once per release — acceptable. - Cost results usually carry currency already — the response transform injects
currencyinto the result container the model reads (response_transform.go:53-62) — but only when the request body declared it. When the model composes a query without an explicit currency, the API applies its default and the result carries no marker; the prompt rule above then has nothing to cite. Cheap fix in the same transform: when the request declared no currency and the result has money-typed metric columns, inject the API's documented default (USD). Usage-only results stay unstamped — a currency on unit metrics would mislabel them (review finding, first cut gated only on the request having aconfigfield). - Usage metrics have no unit anywhere in the query response — a known API gap (noted at
response_transform.go:56). The prompt rule's fallback: when the unit is genuinely unavailable, the model must say so in the header ("usage, unit as reported") rather than print bare numbers. The real fix is upstream — fold into the omni issue in F6.
Tests: prompt prefix contains the units rule (ai_prompt_test.go); transform injects currency: USD when the request body had none (response_transform_test.go).
Feedback: "/mouse just for wheel scrolling is the correct behavior?"
Yes — by design (AI-SPEC §3.1, D9): mouse capture defaults off because capture broke text selection/copy in some terminals; /mouse opts into wheel scrolling for the transcript, PgUp/PgDn always work (ai_tui.go:166-168, 1231-1237). This spec keeps the default.
Two small improvements:
- Persist the toggle:
/mousecurrently resets every session (mouseOn: false,ai_tui.go:215). Store it inai_settings.jsonlike/bell(F2) and/model, so wheel-scrolling users choose once. - Reply in the thread confirming it's intentional and why (copy/paste survival), pointing at the
/helpscrolling paragraph (ai_tui.go:1591-1594), which already explains it.
Tests: settings round-trip restores mouseOn.
Feedback: "There's an issue with the list-contracts api (not the harness)" — the screenshot shows the model's own graceful summary: list-contracts posthog.com returned 403 Forbidden ("Check the active customer context and your DoiT permissions"), reading contracts needs contractsReadOnly / contractsViewer or a write-capable role.
Alfredo's doer session should plausibly hold contract read permission for a customer context it can otherwise query — this looks like a permission-model gap (or an over-strict check) in the DCI API, owned by omni, not by this repo.
Action: filed as doiteng/omni#62101, together with the usage-units API gap from F4.3. CLI-side: nothing — the harness already surfaced the failure with the exact roles required, which is the behavior we want.
Feedback (second thread): "Can we store the API Key in the settings? When running dci ai for the first time, we could ask for it if there's no env var." Storage and the guided prompt already exist (ai_settings.json, guided key setup ai_tui.go:1043-1078), but the prompt only appears after the user types their first question — Alfredo tested one-shot and never saw it; maintainer verdict: "a bit clunky today."
Change: when the session opens with no resolvable key (resolveAIKey empty), start the guided key setup immediately — banner explaining that / commands work now and AI needs a key, input pre-focused in key-entry mode, Esc dismisses to a normal session (the current type-a-question trigger stays as the fallback). Typing / on an empty key buffer also drops straight into command entry (review finding: the setup owns the keyboard, so the "/ commands work" hint must be honored, not just printed); a slash mid-key or after a question queued the setup stays key input. One-shot keeps the current hard error (ai_command.go:68) — no interactive prompts on a scriptable surface — but the message should mention both ANTHROPIC_API_KEY and that running dci ai interactively offers to save the key.
Tests: session start with empty key → key-entry mode active, banner rendered; Esc → normal session; start with key present → unchanged.
All CLI changes are patch-level (fix:/feat: per commit; released in the next batched tag per the usual cadence). Suggested order — each lands independently:
| Order | Item | Files | Effort |
|---|---|---|---|
| 1 | F1 Enter accepts completion | ai_tui.go |
XS |
| 2 | F3 --quiet/--verbose |
ai_command.go |
XS |
| 3 | F2 bell + /bell |
ai_tui.go, settings |
S |
| 4 | F7 key setup on open | ai_tui.go, ai_command.go |
S |
| 5 | F4 units prompt rule + USD default | ai_prompt.go, response_transform.go |
S |
| 6 | F5 persist /mouse + thread reply |
ai_tui.go, settings |
XS |
| — | F6 file omni issue (+ units API gap) | omni repo | issue only |
Decisions taken at implementation: (a) F1 Enter accepts the highlighted completion and never auto-runs it; only an exactly-typed command submits (case-sensitive, so /QUIT⏎ corrects to /quit instead of failing). (b) F2 bell defaults on, gated to turns that ran commands or lasted ≥3s. (c) F4's prompt-prefix edit accepted — one cache invalidation per release. (d) F6 and the units gap share one omni issue (doiteng/omni#62101).