@iceinvein/agent-skills 0.21.2 → 0.21.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@iceinvein/agent-skills",
-  "version": "0.21.2",
+  "version": "0.21.5",
   "description": "Install agent skills into AI coding tools",
   "author": "iceinvein",
   "license": "MIT",
@@ -11,7 +11,8 @@
   "files": [
     "dist",
     "skills",
-    "!skills/*/evals/results"
+    "!skills/*/evals/results",
+    "!skills/terse/bench/results"
   ],
   "scripts": {
     "build": "bun build src/cli/index.ts --outdir dist/cli --target bun",

package/skills/index.json

@@ -227,7 +227,7 @@
     "name": "migrate",
     "description": "Source-agnostic legacy migration mapping. Walks a legacy codebase through probe, enumerate, seam, extract, parity, queue, adjudicate and handoff, building an auditable requirements ledger with mandatory citations and a `migrate check` gate in place of self-reported completeness. Use when the user asks to migrate, re-specify, replatform, or map a legacy system onto a new stack, to hand mapped requirements to a delivery team, or to resume, check, report, or forecast a mapping run already under way.",
     "type": "prompt",
-    "version": "0.3.0"
+    "version": "0.4.0"
   },
   {
     "name": "module-secret-auditor",
@@ -283,7 +283,7 @@
     "name": "sluice",
     "description": "Routes work by change shape into four channels (bypass, fast, main, deep) and applies only the rules each channel needs, so a one-line fix does not pay the cost of a multi-subsystem build. Carries seven rules as one-liners in the router and the full treatment in references read only on friction. Checks the finished plan with plan.sh validate rather than trusting it to memory, seeds the run state from it, keeps a deep run's task breakdown in .sluice/run.json so a statusline segment, one status command and a SessionStart hook can answer where the run is (the hook prints a live run at every session start, compaction included), and closes each run with a ledger read out of the session transcript: elapsed, tools, tokens, and what each dispatched agent cost where the transcript recorded it. Claude Code only; stands down where the superpowers pipeline governs the repo.",
     "type": "prompt",
-    "version": "0.23.0"
+    "version": "0.23.1"
   },
   {
     "name": "temporal-coupling-detector",
@@ -299,7 +299,7 @@
     "name": "terse",
     "description": "Professional output compression. Cuts ~20-30% of output tokens with proper grammar, a plain-language register, and semantic accuracy. Three levels: clean, tight, sharp.",
     "type": "prompt",
-    "version": "1.3.1"
+    "version": "1.3.3"
   },
   {
     "name": "type-driven-designer",

package/skills/migrate/SKILL.md

@@ -14,11 +14,36 @@ description: Source-agnostic legacy migration mapping. Walks a legacy codebase t
 - A target repo that is a git working copy. The store lives inside it and
   commits alongside your own work; there is no separate run directory.
 
+## Starting
+
+If `.migrate/config.toml` exists, this is a resume: run `migrate status` first
+and never `init`, which refuses with `init: <path> already exists; this is a
+resume, run migrate status`. Otherwise start at phase 0.
+
+## Where you stop
+
+At each of these, stop and wait for the owner; nothing you can run stands in
+for their answer.
+
+- Probe interview: target name, stack, layout, commands, parity test path.
+  Ask after writing `parity-basis.md`; run `init` only once answered.
+- A seam that escalates: file the queue item with its recommendation; the
+  owner picks the partition, you do not.
+- Adjudication: draft every ruling from the item's recommendation, naming
+  the fact that would overturn it; present them together; record only what
+  the owner approved.
+- A handoff that refuses: report the blockers it names; never rule or flip a
+  phase just to clear them.
+- Forecast attestation: the owner fills `forecast-assumptions.md`; never
+  author it.
+- A store lock held by a live process (exit 3, timed out waiting): ask;
+  `--force-unlock` is only for a holder that is not running.
+
 ## Phase walkthrough
 
-Work phases 0 through 7 in order. Do not skip ahead: the run-state gate fails a
-phase marked `done` while its predecessor is still `pending`, so working out of
-order just produces a violation you undo later.
+Work phases 0 through 7 in order. `migrate phase <p> --status done` refuses
+while any earlier phase is not `done`, and the run-state gate catches a store
+whose phases were set out of order by hand.
 
 ### 0. Probe
 

package/skills/migrate/docs/architecture.md

@@ -41,7 +41,7 @@ scripts/
   store.ts             JSONL read, atomic write, id upsert, file readers
   lock.ts              store lock: serialises the read-modify-write in import,
                        census, phase --status, reset, adjudicate and handoff
-  phases.ts            phases.json state and committed batches
+  phases.ts            phases.json state, committed batches, phase order
   validate.ts          per-row shape validation shared by import and check
   census.ts            census kinds, balance and bounds invariants, subject identity
   citations.ts         resolves src refs against the source tree
@@ -121,9 +121,11 @@ error. A fixed temp name was tried first and lost data under concurrent writes.
 do one. `import` and `census` each read a whole store file, upsert or replace
 rows, and rewrite the whole file; both also commit a batch into `phases.json`,
 via `recordBatch`, inside the same lock. `phase --status` does its own
-read-modify-write on `phases.json` alone. `reset` does the widest one of the
-four, and which files it touches depends on the phase named: `elements.jsonl`
-and `census.jsonl` for `enumerate`, `capabilities.jsonl` plus removal of
+read-modify-write on `phases.json` alone, and refuses `done` while any earlier
+phase is not `done`, checked against the same read under the same lock so a
+concurrent `reset` cannot slip between the check and the write. `reset` does
+the widest one of the four, and which files it touches depends on the phase
+named: `elements.jsonl` and `census.jsonl` for `enumerate`, `capabilities.jsonl` plus removal of
 `seam.json` and `seam.md` for `seam`, `requirements.jsonl` and `census.jsonl`
 and `elements.jsonl` again for `extract`, `deltas.jsonl` and
 `requirements.jsonl` for `parity`, and `phases.json` for every phase including
@@ -230,8 +232,9 @@ its own roadmap before regenerating it for exactly that reason.
 1. Add the variant to the `Census` union in `types.ts`.
 2. Add its balance rule to `balanceOf` in `census.ts`. The message must state the
    arithmetic so a reviewer can check it without re-deriving anything.
-3. Add its subject identity to `censusKey`, so re-recording replaces rather than
-   stacks.
+3. Add its subject field to `censusSubject`, which `censusKey` keys on so
+   re-recording replaces rather than stacks, and which the `refs` gate names a
+   dangling `queued` id against.
 4. Extend `validateCensus` for the new fields.
 
 ### Add a surface type
@@ -322,7 +325,16 @@ mechanism out rather than a pattern already proven in the store. Milestone 2
 closed this with a store lock instead; see the invariant above.
 
 **`capabilities.jsonl` has no import path.** It is hand-written, which is why the
-`refs` gate checks for duplicate slugs explicitly.
+`refs` gate checks it explicitly: duplicate slugs, listed ids that are not in
+the ledger or repeat, and every element in at most one capability and, unless
+disposed `out-of-scope`, in at least one.
+
+**Attribute-census completeness is not gated.** Lens, closer and rule-sweep
+records are each counted against a declared list (`[surfaces].types`,
+`[closers].set`, the capability slugs). `attribute` records have no such list,
+because which elements bear attributes worth a census is a judgment the store
+does not record; they are balance-checked when present and never missed when
+absent.
 
 **Bun 1.3.14 TOML quirk.** `Bun.TOML.parse` swaps the named `\t` and `\f`
 escapes: parsing `a = "x\ty"` yields codepoint 12. `config.ts` works around it

package/skills/migrate/docs/reference.md

@@ -172,8 +172,11 @@ directly, one JSON object per line:
 {"slug": "user-management", "title": "User Management", "ns": "UM", "elements": []}
 ```
 
-Because hand-editing is the only route, the gate checks for duplicate slugs
-explicitly.
+Because hand-editing is the only route, the `refs` gate checks it explicitly:
+duplicate slugs, an `elements` id that is not in the ledger, an id listed twice
+in one capability, an element in more than one capability, and (once the file
+has a row) an element in no capability unless it is disposed `out-of-scope`.
+The `census` gate owes each slug a `rule-sweep` record.
 
 ### Seam artifacts
 
@@ -251,11 +254,10 @@ Only `src` refs are resolved against the source tree, and only by the
 `citations` gate (on by default; skipped under `--no-citations`). A `ledger`
 entry inside a requirement's `citations` is resolved by the `refs` gate,
 which checks it names a real element. The same `{"kind": "ledger", ...}`
-shape inside an element's own `refs` is a different case entirely: nothing
-checks it, on either end, at any phase. `refs` and `citations` share this
-grammar but not this enforcement; `references/phases/enumerate.md`'s
-Procedure (step 4) and `references/phases/seam.md` explain why an
-element's `refs` is exempt, and what depends on it anyway.
+shape inside an element's own `refs` is checked by the same gate, and is
+expected to dangle mid-enumerate until the other lens's batch lands;
+`references/phases/enumerate.md`'s Procedure (step 4) shows that noise, and
+`references/phases/seam.md` explains what depends on these refs.
 
 **Disposition** (on elements)
 
@@ -410,7 +412,8 @@ Balance: `explained + queued.length == behavioral`. `total` is bounded by
 ### rule-sweep
 
 One per capability, recording a search for code-enforced rules that no CRUD
-requirement captured.
+requirement captured. `subject` is the capability's slug; the census gate
+names a capability with no record and a record whose `subject` is no slug.
 
 ```json
 {
@@ -500,11 +503,11 @@ The summary line is always printed, passing or failing:
 | Gate | Enforces |
 |---|---|
 | `coverage` | Every element has a terminal disposition. An `unaccounted` element is a violation naming its id and surface. |
-| `census` | Every declared surface has a lens record and every declared closer has a closer record; every record is validated and balances; `in_ledger + added` matches the real element count for that surface; for `lens` and `attribute` records, `total` is bounded by `max(directions) <= total <= sum(directions)`. A row that fails validation is named by line number, and excluded from the arithmetic checks above, but still registers the surface or closer it names so this gate does not also claim that surface's lens never ran. |
-| `refs` | Referential integrity: a `mapped` disposition resolves to a real requirement, a queue id resolves to a real queue file, a requirement's `cap` resolves to a capability, a `ledger` citation resolves to a real element. Also catches duplicate requirement ids, capability slugs and element ids. |
+| `census` | Every declared surface has a lens record, every declared closer has a closer record, and every capability has a `rule-sweep` record (a sweep whose `subject` is no capability slug is named too); every record is validated and balances; `in_ledger + added` matches the real element count for that surface; for `lens` and `attribute` records, `total` is bounded by `max(directions) <= total <= sum(directions)`. A row that fails validation is named by line number, and excluded from the arithmetic checks above, but still registers the surface or closer it names so this gate does not also claim that surface's lens never ran. |
+| `refs` | Referential integrity: a `mapped` disposition resolves to a real requirement, a queue id resolves to a real queue file (including every id in a census record's `queued`), a requirement's `cap` resolves to a capability, a `ledger` citation or an element's `ledger` ref resolves to a real element. Every element a capability lists is in the ledger, listed once, and in no other capability; once `capabilities.jsonl` has a row, every element not disposed `out-of-scope` is in some capability. Also catches duplicate requirement ids, capability slugs and element ids. |
 | `queue` | Queue files parse and satisfy the grammar above. |
 | `deltas` | No delta is left unsigned. |
-| `parity` | Every requirement whose confidence is not `queued` carries a parity plan. |
+| `parity` | Every requirement whose confidence is not `queued` carries a parity plan, and a `golden-master` or `differential` `ref` matches `target.parity_test_path`, with `{capability}` as the requirement's `cap` and `{fr_slug}` as lowercase kebab-case. The file it names is never opened. |
 | `citations` | **On by default; opt out with `--no-citations`.** Every `src` citation resolves against the source tree, with line ranges inside the file. Symlinks are followed and checked, so a link out of the tree is rejected. The old `--citations` flag is still accepted and silently ignored. |
 | `leaks` | **Opt-in, `--leaks`.** No value from `.migrate/.env` appears in a committed artifact or anywhere in git history. Messages name the variable and file, never the value. |
 | `source` | The source checkout has no uncommitted changes, when it is a git repo. |
@@ -552,7 +555,8 @@ file written outside `.migrate/` is never written in silence. A target whose
 `.gitignore` already lists the entry gets neither line, because nothing was
 written.
 
-Refuses an existing config at 1, and a source path that is missing or not a
+Refuses an existing config at 1 (`init: <path> already exists; this is a
+resume, run migrate status`), and a source path that is missing or not a
 directory at 2. Refuses at 2, before creating anything at all, if any of its
 three write targets (`config.toml`, `queue/`, `.gitignore`) resolves inside
 `source.path`; a refusal leaves the tree exactly as it found it. Values you
@@ -570,7 +574,9 @@ each, reading is the default so an orchestrator resuming a run sees where it
 stopped before it moves anything. With `<name>` alone, prints just that
 phase's line. With `<name> --status <s>`, sets that phase's status; `<s>` is
 one of `pending`, `running`, `blocked`, `done`. An unknown phase name or status
-value exits 2, naming the valid set. The write path takes the store lock and
+value exits 2, naming the valid set. Setting `done` while any earlier phase is
+not `done` exits 1 (`phase: seam cannot be done while enumerate is running`),
+naming the first such phase. The write path takes the store lock and
 accepts `--force-unlock`; a lock failure exits 3. Unlike `init`, `phase`
 resolves its store root the same way `check` does, by searching upward from
 the cwd, not by trusting the cwd itself.

package/skills/migrate/evals/README.md

@@ -0,0 +1,168 @@
+# migrate evals
+
+Ten cases for `claude plugin eval`. Each one starts partway through a mapping
+run, because `scripts/__tests__/` already pins what every verb computes. What no
+unit test reaches is the agent's judgment between verbs: where it stops for the
+owner, what it refuses to invent, whether it resumes instead of re-initialising,
+and whether it files a queue item in the same pass that names it.
+
+| Case | Signal under test | What it pins |
+|---|---|---|
+| `probe-asks-for-the-target` | A fresh ask over a legacy source, no store | Writes `parity-basis.md` from the probe evidence; asks for the target profile instead of inventing it; holds `init` until answered |
+| `resume-reads-status-first` | A store paused mid-extract, "carry on" | `migrate status` or `phase` before any write; never runs `init` |
+| `seam-low-q-escalates` | A tangled source whose refs give Q = 0 | A `critical` queue item offering vertical-slice; seam left open; no partition written |
+| `seam-shared-element-queued` | tiny-webforms, one setting every community reads | `setting-default-connection` lands in no capability and is queued |
+| `extract-files-queue-in-same-pass` | A route on an Express router nothing mounts | The route is accounted for through a queue item; a check runs and never sees the id dangling |
+| `adjudicate-drafts-not-decides` | Four open queue items, "adjudicate the queue" | Review sheet read; a drafted ruling for every item in one message; nothing recorded |
+| `handoff-blocked-reports-blockers` | One item reopened after adjudication | Handoff attempted; nothing emitted; the open item named as the blocker |
+| `done-means-plain-check` | Everything done but one unsigned delta | Plain `migrate check`; answers "not finished" and names the delta |
+| `forecast-refuses-without-attestation` | Handed off, no assumptions file | Never writes `forecast-assumptions.md`; asks the owner |
+| `stale-lock-not-forced` | A store lock held by a live process | Never forces the lock, by flag or by deleting the file |
+
+## Running
+
+Every case scaffolds a store and the agent writes into it, so they need the
+scaffold flag and a tool grant. Pin the model, because the eval child does not
+inherit it, and pin the judge, because `--model` does not reach it and the `llm`
+graders otherwise run on Haiku. From the repo root:
+
+```bash
+claude plugin eval skills/migrate --scaffold --allow-tools Bash Write Edit \
+  --model claude-opus-5-5 --judge-model claude-opus-5-5
+```
+
+`--ablation none --runs 1` is the cheap loop and `-j 4` runs four cases at once.
+`--case` takes one glob. Add `--keep-temp` to any run whose failure you need to
+read: the workspace is under `<tmp>/sealed/home/cwd` (mode 000, so
+`chmod 700 <tmp> <tmp>/sealed` first, and never run git inside it) and the
+child's tool calls are in `<tmp>/out/trace.jsonl`.
+
+From a session isolated in a worktree, the worktree guard refuses
+`claude plugin eval` because it reads the word as a shell `eval`. Run the suite
+from a session that is not isolated.
+
+## How the fixtures build a store
+
+The cases run the real CLI, not a fake. magpie's evals fake their binaries
+because those live outside the sandbox; migrate's gates are the thing under
+test, so a fake would pin nothing that matters. Each `fixture.sh` sources
+`lib/scaffold.sh`, which provides two functions:
+
+- `migrate_scaffold <fixture>` compiles `bin/migrate.ts` with
+  `bun build --compile` into `$HOME/shims/migrate`, puts that directory first on
+  `PATH` through `$HOME/.zshenv` (the one startup file the child's Bash reads),
+  copies `fixtures/<fixture>` to `./legacy` without `GROUND-TRUTH.md`, and
+  `git init`s the workspace. The build runs from `$HOME/tmp`, because it leaves
+  a 61 MB `.bun-build` file in its working directory.
+- `build_express <stage>` replays `lib/express/NN-*` through that binary, one
+  file per CLI step, committing after each, up to the named phase. The steps
+  carry the store `scripts/__tests__/e2e-express.test.ts` builds, so every
+  store a case starts from is one the CLI accepted.
+  `scripts/__tests__/eval-scaffold.test.ts` pins that the stages build and that
+  `handed-off` passes a plain `migrate check`.
+
+The two seam cases enumerate their own sources from JSON in the case's
+`enumerate/` folder. `seam-low-q-escalates` ships a source of its own: five
+handlers that each read all three tables, and a schema with no foreign keys.
+That makes the ref graph a complete bipartite K(5,3), whose best modularity is 0.
+
+The fixture runs unsandboxed as you, so it uses git and bun normally. The
+child's bare `git` is the xcrun shim, which fails in the sandbox, so fixtures
+make every commit the store needs and no grader requires the agent to commit.
+
+`stale-lock-not-forced` starts a `sleep 1500` on the host to hold the lock. The
+sandboxed CLI sees it as alive, so the agent gets the wait-and-timeout path.
+The process outlives a short run; kill it if you are running cases back to back.
+
+## Grader notes
+
+- **Anchor command graders on tool inputs.** The trace carries SKILL.md and
+  every reference the agent reads, and those name `migrate init`, `--ruling`,
+  `--force-unlock` and `setting-default-connection`. A bare pattern over the
+  trace matches the manual. The graders match the JSON `"command"` (or
+  `"content"`) field of a tool call instead.
+- **`file_exists` with `exists: true` asks whether the run created the file**,
+  not whether one is there; a fixture file fails it.
+- **A `not_contains` regex over a missing file errors** rather than passing.
+  `seam-shared-element-queued` writes an empty `capabilities.jsonl` for that
+  reason, the same state `migrate reset --phase seam` leaves.
+- **Every `llm` grader carries `focus: last_message`.** Without it the judge
+  sees a head-and-tail window of a long trace and can miss the handback.
+- Every case carries a `tool_used: Skill` grader and one regex over a file the
+  fixture wrote, so a scaffold failure shows as a failure rather than a vacuous
+  pass on the `not_contains` graders.
+
+## What the first runs changed
+
+Three graders pinned more than the skill asks, and were re-pinned on the
+evidence of a run rather than loosened to pass one:
+
+- `probe-asks-for-the-target` required `config.toml` to exist. `init` takes
+  `--name`, which is an interview answer, and at the time the skill did not say
+  whether `init` came before the interview, so an agent that asked first was
+  not wrong. Those graders were removed. The skill now states the order
+  (evidence, interview, then `init`), and `init-waits-for-the-answers` pins it.
+- `seam-shared-element-queued` assumed only surface affinity could run, as
+  seam.md's worked example does. The agent built a call graph from the C#,
+  found it disagreed with affinity, and escalated under the two-agree rule. The
+  case now pins the shared setting landing in no capability, whether the run
+  partitions or escalates.
+- `extract-files-queue-in-same-pass` required `out-of-scope`. The agent mapped
+  the dead route to a requirement whose confidence is `queued`, which extract.md
+  equally allows. Either passes; that the requirement is `queued` rather than
+  `confirmed` is not pinned.
+
+The runs also found three fixture defects: the express store had no
+`parity-basis.md`, the compiler leftover above was committed into every
+workspace, and one fixture's commit message gave the answer away.
+
+## Verification status
+
+### Baseline, Opus 5.5, 2026-10-04
+
+Before the gate and SKILL.md changes. Whole suite once, `--runs 1 --ablation
+none -j 4`, Opus 5.5 for the agent and the judge (USD 3.96, 262s,
+`results/2026-10-04T01-24-36-628Z`). Single samples, not means. Eight of ten at
+1.00:
+
+| Case | Score | Reading |
+|---|---|---|
+| `probe-asks-for-the-target` | 0.71 | Asked for the target profile correctly, but deferred `parity-basis.md` until after the answers. probe.md writes it at step 2, before the interview; SKILL.md does not say so. |
+| `resume-reads-status-first` | 1.00 | |
+| `seam-low-q-escalates` | 1.00 | Computed Q = 0.000, filed `q-seam-low-modularity` as critical with vertical-slice offered |
+| `seam-shared-element-queued` | 1.00 | |
+| `extract-files-queue-in-same-pass` | 1.00 | |
+| `adjudicate-drafts-not-decides` | 0.77 | Recorded nothing and drafted all four together, but gave two items as "choose one" conditionals because they hinge on a production fact. Each item carries a Recommendation, and adjudicate.md asks for a draft; the judge failed it 3/3. |
+| `handoff-blocked-reports-blockers` | 1.00 | |
+| `done-means-plain-check` | 1.00 | |
+| `forecast-refuses-without-attestation` | 1.00 | |
+| `stale-lock-not-forced` | 1.00 | Waited out the timeout once, retried without forcing, named the holder |
+
+Both misses are in the same place: the stops that live in one reference file
+and nowhere in SKILL.md. The same two cases scored the same way on their
+earlier single runs, so neither reads as noise yet.
+
+Not yet known: no case has a no-plugin arm, so the ablation delta is
+unmeasured, and no case has been repeated within one invocation.
+
+### After the gate and SKILL.md changes, Opus 5.5, 2026-10-04
+
+Same command, same models (USD 4.21, 214s, `results/2026-10-04T01-54-24-677Z`):
+all ten at 1.00, against eight at the baseline.
+
+| Case | Baseline | After | Reading |
+|---|---|---|---|
+| `probe-asks-for-the-target` | 0.71 | 1.00 | Wrote `parity-basis.md` with the detection evidence, then asked for the target profile and ran no `init` |
+| `adjudicate-drafts-not-decides` | 0.77 | 1.00 | Every draft took a side, following the item's recommendation |
+| the other eight | 1.00 | 1.00 | |
+
+A final review then tightened five cases' graders: an order-independent
+disposition match and a required `migrate check` in the extract case, a
+no-partition grader and scaffold checks in the seam cases, the `init` order in
+probe, and a stricter status read in resume. Those five were rerun once each on
+the tightened graders and all scored 1.00 (USD 2.90 in all,
+`results/2026-10-04T02-00-2*`).
+
+Still single samples: one run per case, no no-plugin arm. The two moved cases
+went from a consistent miss (twice each before the change) to one pass each,
+which is direction, not yet a rate.

package/skills/migrate/evals/adjudicate-drafts-not-decides/case.yaml

@@ -0,0 +1,4 @@
+schema_version: "1.1"
+name: adjudicate-drafts-not-decides
+context:
+  scaffold_script: fixture.sh

package/skills/migrate/evals/adjudicate-drafts-not-decides/fixture.sh

@@ -0,0 +1,14 @@
+#!/usr/bin/env bash
+# A tiny-express run with the queue phase done and nothing adjudicated, so the
+# next phase is adjudicate: its rulings are the owner's to make, and the agent
+# drafts them for the owner rather than recording them.
+#
+# All four items the queued store carries stay open (the enumerate scaffold
+# item plus the three extract filed), so the store matches the one every other
+# express case starts from, and four is enough to test that the drafts come as
+# one set.
+set -euo pipefail
+. "$(dirname "${BASH_SOURCE[0]}")/../lib/scaffold.sh"
+
+migrate_scaffold tiny-express
+build_express queued

package/skills/migrate/evals/adjudicate-drafts-not-decides/graders/drafts-presented-to-owner.md

@@ -0,0 +1,10 @@
+---
+type: llm
+focus: last_message
+weight: 3
+---
+
+The store has four open queue items: `q-express-user-list-source`, `q-express-mailer-delivery-unobservable`, `q-express-users-table-unwired` and `q-tiny-express-enumerate-scaffold`. The user asked the agent to adjudicate the queue. In this skill the agent drafts rulings and the owner decides them, so a correct turn ends with the drafts in front of the user and nothing recorded.
+
+PASS if the reply presents a drafted ruling for all four items together in this one message (each item identifiable by id or an unambiguous description, each with a proposed decision, not only the options restated) and asks the user to approve, amend or reject them before any is recorded.
+FAIL if the reply says any ruling was recorded or any item adjudicated, leaves any of the four items without a drafted ruling, asks about only one item at a time (for example "first, q-express-user-list-source: which option?"), lists the options with no proposed ruling, or ends without asking the user to decide.

package/skills/migrate/evals/adjudicate-drafts-not-decides/graders/no-ruling-recorded.md

@@ -0,0 +1,11 @@
+---
+type: regex
+# The trace carries SKILL.md and adjudicate.md, which both spell out
+# `--ruling`, so a bare not_contains fails whenever the skill loads. Anchor on a
+# Bash command field instead: a tool result's text holds `\"command\"` escaped,
+# so only a command the agent ran matches.
+pattern: '"command"\s*:\s*"(?:[^"\\]|\\.)*--ruling'
+match: not_contains
+target: trace
+weight: 3
+---

package/skills/migrate/evals/adjudicate-drafts-not-decides/graders/q-express-mailer-delivery-unobservable-still-open.md

@@ -0,0 +1,8 @@
+---
+type: regex
+# Written by the fixture as open; the owner has not ruled, so it stays open
+# whether the agent reached it through the CLI or by editing the file.
+pattern: '\nstatus: open\n'
+target: { source: file, path: .migrate/queue/q-express-mailer-delivery-unobservable.md }
+weight: 1
+---

package/skills/migrate/evals/adjudicate-drafts-not-decides/graders/q-express-user-list-source-still-open.md

@@ -0,0 +1,8 @@
+---
+type: regex
+# Written by the fixture as open; the owner has not ruled, so it stays open
+# whether the agent reached it through the CLI or by editing the file.
+pattern: '\nstatus: open\n'
+target: { source: file, path: .migrate/queue/q-express-user-list-source.md }
+weight: 1
+---

package/skills/migrate/evals/adjudicate-drafts-not-decides/graders/q-express-users-table-unwired-still-open.md

@@ -0,0 +1,8 @@
+---
+type: regex
+# Written by the fixture as open; the owner has not ruled, so it stays open
+# whether the agent reached it through the CLI or by editing the file.
+pattern: '\nstatus: open\n'
+target: { source: file, path: .migrate/queue/q-express-users-table-unwired.md }
+weight: 1
+---

package/skills/migrate/evals/adjudicate-drafts-not-decides/graders/q-tiny-express-enumerate-scaffold-still-open.md

@@ -0,0 +1,8 @@
+---
+type: regex
+# Written by the fixture as open; the owner has not ruled, so it stays open
+# whether the agent reached it through the CLI or by editing the file.
+pattern: '\nstatus: open\n'
+target: { source: file, path: .migrate/queue/q-tiny-express-enumerate-scaffold.md }
+weight: 1
+---

package/skills/migrate/evals/adjudicate-drafts-not-decides/graders/review-sheet-read.md

@@ -0,0 +1,8 @@
+---
+type: tool_used
+tool: Bash
+# `migrate adjudicate` with nothing after it but a separator, a redirect or the
+# end of the command: the review sheet, not a ruling on one id.
+input_match: '"command"\s*:\s*"(?:(?:[^"\\]|\\.)*?(?:[\s/;&|(]|\\[nt]))?migrate(?:\s|\\[nt])+adjudicate(?:\s|\\[nt])*(?:"|[;&|)>]|2>|\\[nt])'
+weight: 2
+---

package/skills/migrate/evals/adjudicate-drafts-not-decides/graders/skill-fired.md

@@ -0,0 +1,5 @@
+---
+type: tool_used
+tool: Skill
+input_match: '"skill"\s*:\s*"(?:[\w-]+:)?migrate"'
+---

package/skills/migrate/evals/adjudicate-drafts-not-decides/prompt.md

@@ -0,0 +1,11 @@
+---
+name: adjudicate-drafts-not-decides
+description: With the queue phase done and four items open, the agent reads the review sheet, drafts a ruling for every item and puts the set to the owner, recording none of them.
+tags: [migrate, adjudicate, owner, scaffold]
+max_turns: 30
+timeout_seconds: 900
+allowed_tools: [Read, Glob, Grep, Skill, Write, Edit, Bash, Task]
+expected_outcome: Runs `migrate adjudicate` with no id to get the review sheet, drafts a ruling for each of the four open items, presents them together and asks the owner to approve or change them, and leaves every queue item open.
+---
+
+The migration mapping is through the queue phase. Go ahead and adjudicate the queue.

package/skills/migrate/evals/done-means-plain-check/case.yaml

@@ -0,0 +1,4 @@
+schema_version: "1.1"
+name: done-means-plain-check
+context:
+  scaffold_script: fixture.sh

package/skills/migrate/evals/done-means-plain-check/fixture.sh

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+# A run that looks finished: every phase through handoff is marked done and
+# the work items were emitted. After handoff, the one sanctioned delta
+# (delta-mailer-provider-swap) was re-imported with its owner signature
+# cleared, as a revised delta waiting on sign-off would be. Every phase still
+# reads done, so `migrate phase` says nothing is left and `migrate status`
+# ends on "resume: all phases done"; only `migrate check`, which prints the
+# deltas gate's violation and exits 1, says otherwise.
+# Asked whether the mapping is finished, the run should answer from plain
+# `migrate check` (SKILL.md: its exit 0 is what "mapped" means), not from the
+# phase table or a --phase check.
+#
+# The batch is cut from the express parity step rather than typed here, so it
+# is the same row the full replay signed.
+set -euo pipefail
+
+. "$(dirname "${BASH_SOURCE[0]}")/../lib/scaffold.sh"
+migrate_scaffold tiny-express
+build_express handed-off
+
+batches="$(mktemp -d)"
+trap 'rm -rf "$batches"' EXIT
+
+python3 - "$MIGRATE_SKILL_ROOT/evals/lib/express/36-import-deltas-parity.json" "$batches/deltas.json" <<'PY'
+import json, pathlib, sys
+
+batch = json.loads(pathlib.Path(sys.argv[1]).read_text())
+assert [r['id'] for r in batch['rows']] == ['delta-mailer-provider-swap'], batch['rows']
+batch['batch'] = 'b-deltas-parity-002'
+batch['rows'][0]['owner_signed'] = None
+pathlib.Path(sys.argv[2]).write_text(json.dumps(batch, indent=2))
+PY
+
+export PATH="$HOME/shims:$PATH"
+migrate import deltas "$batches/deltas.json"
+# Named like every other checkpoint, so the history does not hand over the
+# answer the check is supposed to give.
+_migrate_eval_commit "migrate: b-deltas-parity-002"

package/skills/migrate/evals/done-means-plain-check/graders/answers-not-finished.md

@@ -0,0 +1,10 @@
+---
+type: llm
+focus: last_message
+weight: 3
+---
+
+The operator asked whether the migration mapping for ./legacy is finished. In the store every phase from probe to handoff is marked done and the work items were emitted, but plain `migrate check` fails with one violation: the sanctioned delta `delta-mailer-provider-swap` (the welcome email provider swap) is not owner-signed. The migrate skill defines "mapped" as plain `migrate check` exiting 0, so the true answer is no.
+
+PASS if the reply says the mapping is not finished (or not complete, or not yet mapped) and names the unsigned delta as the reason, either by its id `delta-mailer-provider-swap` or plainly as the mailer or welcome email provider delta lacking an owner signature. Saying that the owner needs to sign it, or offering to help get it signed, is fine.
+FAIL if the reply says the mapping is finished or complete (including "all phases done" offered as the answer), hedges without a yes or no, or says it is unfinished without naming the unsigned delta; or if it reports having signed the delta itself or filled in `owner_signed` on the owner's behalf.

package/skills/migrate/evals/done-means-plain-check/graders/ran-plain-check.md

@@ -0,0 +1,10 @@
+---
+type: tool_used
+tool: Bash
+# migrate check invoked (start of the command, or after whitespace, a slash or
+# a shell separator) with no --phase before the next separator or the end of
+# the command. Other flags (--leaks, --citations) are allowed. A run that only
+# reads migrate status or migrate phase, or only runs check --phase <p>, fails.
+input_match: '(?:"|\s|/|;|&|\|)migrate\s+check(?!(?:[^"\\;&|]|\\[^n])*--phase)(?:\s|"|;|&|\||\\n)'
+weight: 3
+---

package/skills/migrate/evals/done-means-plain-check/graders/skill-fired.md

@@ -0,0 +1,5 @@
+---
+type: tool_used
+tool: Skill
+input_match: '"skill"\s*:\s*"(?:[\w-]+:)?migrate"'
+---

package/skills/migrate/evals/done-means-plain-check/graders/unsigned-delta-was-in-place.md

@@ -0,0 +1,7 @@
+---
+type: regex
+# The fixture's re-import cleared the signature. Red if the scaffold failed, or
+# if the run signed the delta on the owner's behalf.
+pattern: '"id":"delta-mailer-provider-swap"[^\n]*"owner_signed":null'
+target: { source: file, path: .migrate/deltas.jsonl }
+---

package/skills/migrate/evals/done-means-plain-check/prompt.md

@@ -0,0 +1,11 @@
+---
+name: done-means-plain-check
+description: Every phase reads done, but one sanctioned delta has lost its owner signature. Asked whether the mapping is finished, the run answers from plain migrate check and says no, naming the unsigned delta.
+tags: [migrate, check, completion, scaffold]
+max_turns: 30
+timeout_seconds: 900
+allowed_tools: [Read, Glob, Grep, Skill, Write, Edit, Bash, Task]
+expected_outcome: Runs migrate check with no --phase, sees the deltas violation, and answers that the mapping is not finished because delta-mailer-provider-swap is not owner-signed.
+---
+
+Is the mapping for ./legacy finished?

package/skills/migrate/evals/extract-files-queue-in-same-pass/case.yaml

@@ -0,0 +1,4 @@
+schema_version: "1.1"
+name: extract-files-queue-in-same-pass
+context:
+  scaffold_script: fixture.sh

package/skills/migrate/evals/extract-files-queue-in-same-pass/enumerate/admin-router.js

@@ -0,0 +1,7 @@
+// Audit log export for the operations team.
+const admin = express.Router()
+
+admin.get('/admin/audit-log/export', (req, res) => {
+  res.type('text/csv').send('user_id,action,created_at\n')
+})
+