[cryptotavares]

Summary: Print one-line JSON when stdout is a pipe or file, reducing bytes sent to agents. Keep terminal output indented and add --pretty and MM_PRETTY overrides.

Stack context: mm CLI v2, Phase 1, P1-15. Base: cryptotavares/remove-ax-tree-path; stacked on #44, which reviewers should read first. Structured errors (P1-16) and retry changes (P1-17) remain separate PRs.

Problem

The CLI indents every JSON result, including output that agents read through a pipe. Nested observations carry whitespace that JSON parsers do not need. Human terminal users still need readable output.

The existing historical proof below measures indentation at 37–47% of the sampled describe-screen and click output bytes. This PR changes serialization at the CLI boundary; it does not reduce daemon response content.

Solution

Resolve the output policy in main, then apply it to the final JSON serialization in sendRequest.

flowchart TD
    A[CLI arguments] --> B[Consume every --pretty token]
    B --> C[Choose indentation: flag, then environment, then terminal]
    C --> D[Run command and receive daemon response]
    D --> E[Assemble result, observations, and metadata]
    E --> F{Output is a string?}
    F -->|yes| G[Write raw text and newline]
    F -->|no| H[Choose compact or two-space JSON serialization]
    H --> I[Write JSON and newline]

    classDef new fill:#fff3cd,stroke:#d39e00
    class B,C,H new

Loading

Yellow boxes are new.

--pretty takes precedence over MM_PRETTY. The environment accepts 1 or true for indentation, and 0 or false for compact output. Values are trimmed and case-insensitive. An unset or unrecognized value falls back to process.stdout.isTTY === true.

extractPrettyFlag removes every exact --pretty token after the existing project and observation flag parsing. The flag controls CLI stdout only. Request bodies and daemon, tool, and platform contracts retain their existing behavior.

What changes

Condition or boundary	Before	After
JSON output to a pipe or file, without an override	Two-space indentation	One line per result, followed by a newline
JSON output to a terminal, without an override	Two-space indentation	Same indentation
--pretty, including with MM_PRETTY=0	No global formatting flag	Two-space indentation
MM_PRETTY=1 or true, without the flag	No environment override	Two-space indentation
MM_PRETTY=0 or false, without the flag	No environment override	Compact JSON, including on a terminal
Final output value is a string	Raw text and newline	Same text behavior

Both JSON formats serialize the same assembled value. Existing result fields, observations, and meta keep their content and order. For example, the compact click test expects {"clicked":true,"observations":{"nodes":[{"ref":"e1"}]}} followed by a newline.

Decisions for reviewers

Decision	Reason
Select the default from stdout, rather than the command or platform	Agents normally use pipes. Terminal users retain readable output across browser, iOS, and Android commands.
Apply formatting after response assembly	Keeps this change at the presentation boundary and preserves existing observation and timing metadata handling.
Use a flag, an environment override, then a terminal fallback	Allows explicit indentation anywhere and explicit compact output on a terminal. Invalid environment values use the default.
Consume --pretty globally	Matches the existing --observe placement convention. Literal --pretty text requires run-steps or the HTTP API.
Keep P1-15 limited to successful output formatting	Error formatting and retry policy belong to P1-16 and P1-17.

Review guide

Read the files in this order:

#	File	What to check
1	src/cli/mm.ts: extractPrettyFlag, resolvePretty, and main	Every flag occurrence is consumed; flag, environment, and terminal precedence stays explicit.
2	src/cli/mm.ts: setPrettyOutput, formatJsonOutput, and sendRequest	Only the final JSON serialization changes. Plain strings and response assembly retain their paths.
3	src/cli/mm.test.ts	Policy cases, repeated flags, exact compact/pretty output, environment isolation, and mocked terminal integration.
4	src/cli/mm.ts: printHelp; README.md; SKILL.md	Document both overrides and tell agents to parse JSON rather than depend on indentation.

Testing

Fresh checks at remote head d43c8dd9486a4ebcec5d19933474198e17cdab8d:

• CI=true yarn vitest run src/cli/mm.test.ts --coverage.enabled=false --maxWorkers=2 --minWorkers=1 passes: 1 file, 333 tests.
• yarn eslint src/cli/mm.ts src/cli/mm.test.ts passes.
• yarn tsc --noEmit -p tsconfig.build.json passes.
• yarn prettier README.md SKILL.md --check and git diff --check pass.
• yarn tsc --noEmit -p tsconfig.json fails with 11 diagnostics in seven test files outside this PR's delta. They cover mobile driver fixtures, server tests, mock-network narrowing, screenshot fixtures, and an in-page type predicate. Those files and TypeScript configurations are unchanged from the base.

The Proof and Unit tests sections below retain the original historical evidence. The fresh checks above do not repeat that live validation or the full suite.

Breaking changes

• JSON results printed to a pipe or file are one line instead of 2-space indented. Callers that parse JSON (JSON.parse, jq) are not affected. To get the old output, pass --pretty or set MM_PRETTY=1.
• --pretty is now consumed by the CLI anywhere in the command line. To type the literal text --pretty, use run-steps or the HTTP API (same as --observe).

Consumer impact

Consumer	Change needed	Done on branch
Extension test/e2e/playwright/llm-workflow/	None (no script reads mm output as text)	n/a
Mobile tests/llm-workflow/	None (no script reads mm output as text)	n/a

Proof

Tested with a yalc build of this branch in the Extension (e2e build, headless Chrome) and Mobile (dev client, iOS sim + Android emulator). "Pretty" is the compact output re-serialised with JSON.stringify(x, null, 2), which is what the CLI printed before. That keeps the comparison exact even if the screen changes between calls.

Platform	Output	Before (pretty)	After (default)	Saving
Extension	describe-screen	9,816 B	6,176 B	37.1 %
Extension	click (full observation)	9,444 B	5,675 B	39.9 %
iOS	describe-screen	25,354 B	14,893 B	41.3 %
iOS	click (full observation)	26,879 B	14,845 B	44.8 %
Android	describe-screen	81,678 B	48,091 B	41.1 %
Android	click (full observation)	74,935 B	39,579 B	47.2 %

With this PR the iOS wallet describe-screen meets the ≤ 15 kB target from #44.

Check (3 platforms)	Result
Default output is one line and valid JSON	pass
--pretty and MM_PRETTY=1 output is byte-identical to the old format	pass
stdout is a TTY: indented by default	pass
MM_PRETTY=0 on a TTY: one line	pass

Unit tests

Test Files  52 passed (52)
     Tests  1827 passed (1827)
All files  97.53 % stmts | 94.91 % branches | 96.76 % funcs | 97.67 % lines

New tests: extractPrettyFlag, resolvePretty (precedence, env values), formatJsonOutput, sendRequest compact/pretty output, and main end to end (pipe, --pretty, --pretty with --observe, MM_PRETTY=1, TTY, TTY with MM_PRETTY=0). MM_PRETTY is stubbed in every test, so a developer's own value cannot change the results.

[seaona]

LGTM