Haiku Tier — Surface Checklist (extended reference)

This is the extended reference the Tier 1 (Haiku) agent loads when a surface check needs more context than the inline agent definition provides. For the canonical flow + RESULT block schema, see ../../../agents/haiku-reviewer.md.

Scope

Surface review sits at the first line of defence. It catches:

• Missing / mis-named files
• Narrative-only "evidence"
• Debug code left in the diff
• Obvious copy-paste, obvious magic numbers, obvious silent-fallback patterns
• Cross-boundary issues that are visible at the diff level (wrong column name in a WHERE, .toFixed() on a nullable)

It does NOT trace call graphs, does NOT cross-verify config bidirectionally, does NOT audit architectural fit. Those belong in later tiers.

What counts as "evidence"

For a [x] box to pass at the surface tier, the evidence must be at least one of:

• file:line — a concrete pointer a reviewer can open and read
• commit hash — a hash visible in git log of the branch
• command + output — a reproducible command with captured output (tee log path is fine)
• subagent RESULT comment-id — a comment number referencing a fenced ## RESULT block

Narrative-only evidence ("implemented the feature") is a fail.

Common false positives

• Intentional duplication for defence-in-depth: a value validated at the API layer AND at the DB layer is not a DRY violation — it's a designed belt-and-braces. Flag for the subagent to confirm intent before marking as a finding.
• Argparse defaults for optional tunables: parser.add_argument("--retries", default=3) is not a "silent fallback" — the argument is optional by contract. Only flag default= patterns that hide REQUIRED config.
• Sentinel values that are the contract: return None from a "not found" function is not a silent failure — it's the advertised return contract. Flag only when the caller can't distinguish the sentinel from a valid value.

Bash output discipline

• Wrap any command expected to produce > 200 lines in a tee-style capture. Report only the path + verdict in RESULT.
• Do NOT paste pytest -v output, git diff of 50+ files, or unfiltered log tails back to the main agent.

Graph fallback

When the code-review-graph MCP is unavailable or the project uses unsupported languages (Markdown, YAML, JSON, SQL, TOML, shell), do NOT silently skip the graph-backed checks. Instead:

1. Retry once after a short delay.
2. If still unavailable, fall back to an Explore-style subagent that performs the equivalent manual audit (callers, callees, large functions).
3. Include the subagent's RESULT block in your own RESULT as [graph unavailable: <reason>] [subagent fallback: <id>].

A bare [graph unavailable] with no subagent evidence = silent skip = FAIL.