Effort is per station, and a build reads its own diff before asking you to

[unitypark]

Two gaps in the agent layer, found by reading it rather than by a failure.

1 · Effort was typed and then thrown away

ModelCallOptions.effort accepted all five levels and defaulted to high.
Three call sites passed the literal 'high'. Nothing configured it. And the
Claude Code path passed no effort at all — so in subscription mode the
setting specd thought it had chosen never reached the model. claude --effort
exists and takes the same five levels; specd simply never sent it.

One value for every station is wrong in both directions at once. xhigh is
the documented setting for coding and agentic work — it's what Claude Code
itself runs at — while cheap mechanical passes are where low belongs. specd
ran a build that writes code a human is about to merge at the same level as an
index run that summarises text.

Station	Was	Now
build, review	high (CLI: nothing)	xhigh
spec, ground	high	high
index	high	low

effortFor(station, project.effort) resolves it, projects.effort moves every
station at once, and NULL means "no preference", not low — a project that
never chose one must not be quietly moved off the defaults. The value now
reaches the API path, both CLI paths, and the runner job payload, so a
dispatched build is not silently cheaper than a local one.

2 · Nothing read the diff

Verify answers "do the tests pass" — which the test suite already answers.
Nobody answered "is this the change we approved, and is it any good". That
reading was left entirely to whoever opened the PR, cold.

ReviewAgent now reads the diff with read-only tools between verify and
publish, and answers in a schema: a verdict, a one-line summary, and findings
carrying path:line, a severity, and the acceptance criterion or design claim
each bears on. They render into the PR body above the acceptance criteria —
the new reading sits above what it's measured against.

Four things it had to get right:

• Advisory, and it says so in the body. A finding does not fail a build.
  That's the line citation drift already sits on: an unrelated opinion must not
  be able to stop an approved spec from shipping. Making findings refusable is
  a house-rule decision (2.4), not a default.
• Read-only by denial, not omission. Write and Edit are in
  --disallowed-tools and the pass runs under --permission-mode plan, so an
  over-helpful reviewer cannot quietly amend the branch it was asked to assess.
• An empty findings list is a real answer, and a clean pass is reported
  rather than omitted — silence would be indistinguishable from a pass that
  never ran, which is the same failure as a green verify for tests that didn't
  execute. Tested.
• It runs where the workspace is. The prompt is authored API-side so both
  paths ask the same thing, but a dispatched build reviews on the runner —
  the diff is on the runner's disk and gone by the time the report arrives.
  Building this only for the in-process path would have been the same
  half-wiring as the effort bug above.

What I deliberately did not do

Replace specd's prompts with stock /plan and /design commands. The
SpecAgent's prompts are what produce EARS criteria, citations, and the four
verdicts — swapping them for generic ones would trade the thing that makes a
spec checkable for convenience. Review was the actual gap.

Verify

pnpm typecheck && pnpm test (43 API files), pnpm build, pnpm site:check.
16 new tests: the station-default invariants (including one that fails if they
ever collapse back to a single value), effortFor null handling, and the PR
rendering — severity ordering, the not-run case, the clean case, and section
order. --effort flag values and --permission-mode plan were both checked
against claude --help rather than assumed.

For the reviewer

• Migration 0019_project_effort.sql adds a nullable column with a CHECK
  constraint. Additive; run pnpm db:migrate.
• Every build now makes one extra model call, at xhigh. Skipped on an
  empty diff, and soft everywhere — the commits exist and verify has spoken, so
  a review that can't run costs an opinion, not the build. But it is a real
  cost increase per build, on top of build effort rising from high to xhigh.
• The review prompt is unexercised against a real model here. Its schema
  and rendering are tested; the quality of what it finds is not, and won't be
  until it runs on a real build.
• Effort is one dial per project, not per station. Someone who wants cheap
  grounding and expensive builds can't express that yet; effortFor is the
  seam.