CLI: improve message formatting to match TUI and align with Claude Code/Codex

[afternoon]

Summary

Improves the visual formatting of CLI session output so it's closer to the TUI and to tools like Claude Code and Codex: a consistent, mostly-grey semantic palette, clearer indentation, and less visual noise around tool calls.

This is a pure formatting/visual change. It does not rebuild CLI output, add a TUI, or touch any other interface (e.g. desktop).

What changed

• docs/cli-formatting-research.md — report on how formatting currently works in goose-cli (console/bat/comfy-table/cliclack) and in the TUI (ui/text, colors.tsx, marked-terminal), plus a summary of Claude Code/Codex conventions, with concrete recommendations.
• docs/cli-formatting-plan.md — the implementation plan derived from that research.
• crates/goose-cli/src/session/formatting.rs (new) — pure, unit-tested helpers with no I/O:
  • A Role enum (Primary, Secondary, Muted, Accent, Success, Error) and role_style/apply to centralize color/dim/bold decisions instead of styling ad hoc at each call site. Accent (cyan) is the only role carrying a non-semantic color; Secondary differs from Muted/Primary by weight (bold), not hue, to keep the palette mostly grey.
  • Shared indentation constants (GUTTER, PARAM_INDENT) and a USER_PROMPT_GLYPH (❯).
  • ToolStatus + status_glyph/status_role/tool_result_status for consistent running/success/error glyphs.
  • format_user_message_plain, format_tool_header_plain, format_tool_status_line_plain, format_error_line_plain, indent_block/ends_at_line_start — plain-text formatting logic that's fully covered by unit tests, separate from the styling/printing.
  • 32 unit tests asserting the new spec (indentation, glyphs, role colors, tool header/footer shape, streaming line-boundary edge cases).
• crates/goose-cli/src/session/output.rs — wired the new helpers in:
  • User messages are now echoed with the accent-colored ❯ glyph and gutter-aligned continuation lines (matching the live prompt), instead of plain unstyled text.
  • Normal assistant/model replies (and rendered markdown tables) are now indented under the same shared GUTTER, instead of sitting flush-left, so they line up visually with the user-message prompt glyph. print_markdown_raw now captures bat's rendered output via print_with_writer and indents it with formatting::indent_block before printing. An at_line_start flag is threaded through the whole streaming render path so a chunk that only continues the previous chunk's still-open line doesn't get re-indented mid-sentence.
  • Tool call headers are now a single indented line with a status glyph, replacing the heavy horizontal rule; tool responses get a ● done / ✗ error status footer derived from the tool result's actual is_error field (not just whether the call completed at the protocol level), suppressed for requests whose header was never shown (e.g. the internal load tool).
  • Parameters, tool output, and error lines use the shared indentation constants and Secondary/Muted/Error roles instead of hardcoded strings/colors.
• crates/goose-cli/src/session/input.rs — the live rustyline prompt now uses USER_PROMPT_GLYPH instead of a literal string.
• crates/goose-cli/src/session/completion.rs — implemented Highlighter::highlight_prompt so the live prompt is styled with the same Accent role as the echoed prompt glyph, keeping live input and history visually consistent.

Review round 1

Reviewed by a subagent on a different model (GPT-5.5) focused on correctness, palette fidelity to the TUI, scope, and test quality. Findings addressed:

• Suppressed the tool-call status footer for requests whose header is never shown, avoiding an orphaned ● done line.
• Routed failed tool responses through the same error-line formatter used elsewhere for a consistent shape.
• Fixed format_user_message_plain to always emit the prompt glyph, even for an empty message.
• Made Role::Secondary colorless (bold-only) instead of reusing Accent's cyan at a different weight, so Accent stays the palette's only splash of color.
• Routed the remaining ad hoc style(...).dim() tool-graph step lines through the shared Role/indentation helpers.
• Fixed a test that mutated console's global color-enabled state without restoring it.
• Added exact-string and edge-case tests for the new footer/empty-message/suppression behavior.

As a follow-up, normal model replies were also brought under the shared gutter (previously the one thing still flush-left), with formatting::indent_block as a pure, independently tested helper.

Review round 2 (adversarial)

A second, adversarial subagent review specifically targeted the gutter-indent follow-up (the print_with_writer capture, streaming interaction, and table seams). It found two real regressions, which are fixed here:

• Tool failures reported as success: render_tool_response only checked whether the outer Result was Ok, but a tool can complete the call successfully at the protocol level while still reporting failure via CallToolResult::is_error (e.g. a shell command that exits non-zero). The footer now uses the new formatting::tool_result_status(result.is_error) and shows ✗ error in that case.
• Streamed replies got a spurious mid-sentence gutter: because each streaming chunk was rendered (and indented) independently, a chunk that only continued the previous chunk's still-open line (the common case for token-by-token streaming, since chunks rarely align to line boundaries) got its own leading gutter, visibly breaking sentences in two. Fixed by threading an explicit at_line_start flag through render_message_streaming/print_markdown/print_markdown_raw/print_table/flush_markdown_buffer, the same way thinking_header_shown is already threaded.

Also fixed as a smaller related issue found in the same pass: in non-terminal mode (piped/redirected stdout), tool output is printed verbatim without a guaranteed trailing newline, which could glue the status footer onto the same line as un-terminated output; render_tool_response now ensures a fresh line before the footer.

Two other findings from that review were assessed and intentionally not changed: (1) bat's NoWrapping(true) mode means it was never fitting content to the exact terminal width to begin with, so the extra 2-column gutter carries the same negligible overflow risk already accepted for the existing 4-space PARAM_INDENT tool-output indentation; (2) the thread-local set tracking suppressed tool headers is session-scoped and bounded by the number of internal load calls in a session, so a cancelled call leaving a stale entry isn't a practically meaningful leak.

docs/cli-formatting-plan.md was also updated to reconcile a stale non-goal note and an outdated Secondary color description with the shipped gutter-indent follow-up.

Testing

• cargo test -p goose-cli — all tests pass (2 intentionally #[ignore]d manual visual smoke checks), including the session::formatting unit tests.
• cargo clippy -p goose-cli --lib --tests -- -D warnings — clean.
• cargo fmt --check — clean.
• cargo build -p goose-cli — builds cleanly.

Scope

No changes outside of goose-cli session formatting and the two docs files. The TUI (ui/desktop) and desktop app are untouched.