Non-interactive modes for /ba:review (C1)

[azevedo]

Status

Deferred until concrete pipeline / CI / programmatic need exists.

Context

ce-code-review has four modes:

Mode	Interactive?	Purpose
Interactive (default)	Yes	User-driven flow
Autofix	No	Applies safe_auto fixes silently, surfaces non-safe findings for review
Report-only	No	Read-only structured report
Headless	No	Programmatic — emits structured JSON to stdout for downstream consumers

The three non-interactive modes ("externalization modes" in ce) share: structured JSON output, run-artifact persistence (/tmp/compound-engineering/ce-code-review/{run_id}/), and a validator pass (Stage 5b) that doesn't run in interactive.

Decision

Deferred.

Rationale

• No PR-based CI. User is solo contributor, no PRs.
• No current pipeline composing /ba:* commands programmatically.
• The strongest sub-argument for C1 collapsed during design:
  • A2 (run-artifact persistence) was reshaped as an interactive opt-in feature — no longer needs C1 as a carrier.
  • B8 (autofix-class) is itself deferred — no carrier needed.
• Building C1 now risks designing for a workflow that doesn't yet exist — exactly the YAGNI failure correctly applied to B7 in the same conversation.

Design captured (if revisited)

Modes to support

At minimum:

• Interactive (existing)
• Report-only (read-only, structured-markdown or JSON output)
• Headless (programmatic, JSON to stdout)

Autofix mode only if B8 is accepted in the same change (since autofix_class is the routing field).

Invocation surface options

Option	Pros	Cons
Argument flag: /ba:review HEAD..HEAD --mode=headless	Parseable from $ARGUMENTS; one command	CLI-shaped, not slash-command-shaped
Separate commands per mode (/ba:audit, /ba:autofix)	Cleaner surface per mode	Multiplies command file count
SDK / Task-tool only	Cleanest separation	No human-typed access to non-interactive features

No commitment made; revisit when the concrete consumer is known.

Output format

• Headless mode: Structured JSON to stdout. Hard dependency on C2's schema.
• Report-only mode: Markdown with structured frontmatter (machine-readable header, prose body).

Behavioral changes in non-interactive modes

• Skip reviewer-selection prompt (default to all built-in + discovered reviewers)
• Persist run artifact (likely reuse A2's directory format / location)
• Suppress conversational output
• Apply the C2 validator pass (which is the same machinery as interactive, just gated as a stricter contract for stdout consumers)

Trigger conditions for revisit

Any of:

• Concrete pipeline composing /ba:* commands programmatically (e.g., a workflow that runs /ba:execute then /ba:review then some downstream step that consumes review output)
• CI integration plan that wants exit-code semantics or structured reports
• External MCP/SDK consumer wanting structured /ba:review output
• Autofix mode becoming a priority (would bring B8 along)

Candidates that would revive on C1 acceptance

• B8 (autofix-class field): if autofix mode is in scope for C1, B8 ships with it
• A3 (contract tests): gains a stable output substrate to assert against (though C2 alone provides this)

References

• ce-code-review four-mode system (Interactive, Autofix, Report-only, Headless)
• A2 issue — already reshaped to interactive scope; no longer dependent on C1
• B8 issue — deferred separately; tied to autofix mode if ever revived
• B4+B5+B6+C2 bundle issue — C2 schema is the prerequisite for headless JSON output