@christang/keel 5.51.0 → 5.53.0

package/assets/bootstrap/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- keel:start version=5.51.0 -->
+<!-- keel:start version=5.53.0 -->
 ## Keel Bootstrap
 
 - Start every session with `keel context`; OpenSpec artifacts and Git are the only durable authority — never native memory, goals, or transcripts.

package/package.json

@@ -2,7 +2,7 @@
   "name": "@christang/keel",
   "displayName": "Keel",
   "description": "Keel OpenSpec execution discipline CLI for Claude Code, Codex, and OpenCode.",
-  "version": "5.51.0",
+  "version": "5.53.0",
   "license": "MIT",
   "repository": {
     "type": "git",

package/plugins/keel/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "keel",
-  "version": "5.51.0",
+  "version": "5.53.0",
   "description": "Keel OpenSpec execution discipline: stateless continuity, task capsules, deterministic gates, and expectation alignment for Codex and Claude Code.",
   "author": {
     "name": "TanglmChris",

package/plugins/keel/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "keel",
-  "version": "5.51.0",
+  "version": "5.53.0",
   "description": "Keel OpenSpec execution discipline: stateless continuity, task capsules, deterministic gates, and expectation alignment for Codex and Claude Code.",
   "author": {
     "name": "TanglmChris",

package/scripts/validate_plugin.py

@@ -37,8 +37,8 @@ REQUIRED_SCRIPTS = [
     "scripts/validate_plugin.py",
 ]
 
-PACKAGE_VERSION = "5.51.0"
-PROTOCOL_VERSION = "5.51.0"
+PACKAGE_VERSION = "5.53.0"
+PROTOCOL_VERSION = "5.53.0"
 LEGACY_MANAGED_START = "<!-- keel:start version=2.1 -->"
 OPENSPEC_SCHEMA_NAME = "keel-spec-driven"
 # Mirrors KEEL_PACKAGE_NAME in scripts/install_to_repo.py, one of the two
@@ -24945,13 +24945,20 @@ def strategy_probe_task(
         "    - README.md",
         "  - Touch:",
         "    - src/example.js",
-        f"  - {form}:",
     ]
-    if strategy is not None:
-        label = "Strategy" if form == "Verify" else "Verification Strategy"
-        lines.append(f"    - {label}: {strategy}")
-    if reason is not None:
-        lines.append(f"    - Reason: {reason}")
+    # In the expanded v3 form the strategy and its reason are task-level fields
+    # beside `Commands`, not entries inside it.
+    if form != "Verify":
+        if strategy is not None:
+            lines.append(f"  - Verification Strategy: {strategy}")
+        if reason is not None:
+            lines.append(f"  - Verification Reason: {reason}")
+    lines.append(f"  - {form}:")
+    if form == "Verify":
+        if strategy is not None:
+            lines.append(f"    - Strategy: {strategy}")
+        if reason is not None:
+            lines.append(f"    - Reason: {reason}")
     lines.extend(f"    - {entry}" for entry in commands)
     lines.extend(
         [
@@ -25401,6 +25408,410 @@ def validate_quoted_marker_is_not_a_disposition_scenario() -> int:
     return 0
 
 
+# A drift is a hard stop, so whatever the message says is what the author acts
+# on. Saying only that a value moved produced a wrong record in this repository:
+# the cause was recorded as a Review edit, which is measurably not covered by
+# the fingerprint at all.
+def validate_drift_names_where_to_look_scenario() -> int:
+    label = "drift-names-where-to-look"
+
+    def fixture(root: Path, name: str, anchor_value: str) -> Path:
+        repo = root / name
+        task = strategy_probe_task(strategy="vertical-tdd")
+        # Cite a design statement so the capsule holds two distinct sources.
+        task = task.replace(
+            "    - E1: the task proves its own behavior",
+            "    - E1: the task proves its own behavior\n    - D1",
+        )
+        task = task.replace(
+            "    - Contract: pending",
+            f"    - Contract: keel-task-capsule/v1 sha256:{anchor_value}",
+        )
+        write_gate_fixture(
+            repo,
+            tasks=task,
+            design=(
+                "## Context\n\nfixture\n\n## Decisions\n\n"
+                "- D1 — the public behavior is proven through the CLI. "
+                "Basis: fixture.\n"
+            ),
+        )
+        return repo
+
+    with tempfile.TemporaryDirectory(prefix="keel-drift-") as raw:
+        root = Path(raw)
+
+        # The true anchor, so the matching case can be compared against it.
+        settled = fixture(root, "settled", "0" * 64)
+        started = run_keel(
+            settled, "gate", "task-start", "--change", "demo", "--task", "1.1",
+            "--json", "--no-guard",
+        )
+        true_value = json.loads(started.stdout)["contract"]["fingerprint"]["value"]
+
+        drifted = fixture(root, "drifted", "0" * 64)
+        result = run_keel(drifted, "context", "--change", "demo", "--json")
+        payload = json.loads(result.stdout)
+        if payload.get("status") != "blocked":
+            report(
+                f"{label}: a task whose anchor does not match was not blocked; "
+                f"status {payload.get('status')!r}."
+            )
+            return 1
+        reason = " ".join(str(x) for x in (payload.get("reasons") or []))
+
+        for needed in ("0" * 8, true_value[:8]):
+            if needed not in reason:
+                report(
+                    f"{label}: the drift reason no longer reports both "
+                    f"fingerprints; {reason!r}."
+                )
+                return 1
+        for source in (
+            "openspec/changes/demo/tasks.md",
+            "openspec/changes/demo/design.md",
+        ):
+            if source not in reason:
+                report(
+                    f"{label}: the drift reason does not name the authority "
+                    f"source {source!r}; {reason!r}."
+                )
+                return 1
+        # The list is the authority set, not the change directory.
+        if "proposal.md" in reason:
+            report(
+                f"{label}: the drift reason names a file the capsule resolved "
+                f"no authority text from; {reason!r}."
+            )
+            return 1
+        lowered = reason.lower()
+        for word in ("evidence", "review", "checkbox"):
+            if word not in lowered:
+                report(
+                    f"{label}: the drift reason does not say that {word} is "
+                    f"outside the fingerprint; {reason!r}."
+                )
+                return 1
+        # D3: it names a search set, never a culprit.
+        for guess in ("changed field", "the field that changed", "caused by"):
+            if guess in lowered:
+                report(
+                    f"{label}: the drift reason claims which field moved, which "
+                    f"it cannot know; {reason!r}."
+                )
+                return 1
+
+        matching = fixture(root, "matching", true_value)
+        ok = json.loads(
+            run_keel(matching, "context", "--change", "demo", "--json").stdout
+        )
+        if ok.get("status") != "ready":
+            report(
+                f"{label}: a task whose anchor matches was not ready; "
+                f"{ok.get('status')!r} {ok.get('reasons')!r}."
+            )
+            return 1
+        ok_reason = " ".join(str(x) for x in (ok.get("reasons") or []))
+        if "fingerprint" in ok_reason.lower():
+            report(
+                f"{label}: a matching anchor now reports drift wording; "
+                f"{ok_reason!r}."
+            )
+            return 1
+
+    report(f"{label} scenario passed.")
+    return 0
+
+
+# Issue #112's minimal reproduction. An unfilled slot in an `M<n>` declaration
+# stopped the contract compiling, the completion path fell back to the expanded
+# v3 `Commands` field a compact task never declares, and every reference to an
+# `M<n>` was reported as naming a check the task does not declare — first, and
+# as somebody else's fault. The reporter calls it the only diagnostic in 149
+# invocations that made them edit the wrong file.
+def validate_reference_outlives_its_declaration_scenario() -> int:
+    label = "a-reference-outlives-its-declaration"
+
+    def complete(root: Path, name: str, checks, findings, form="Verify"):
+        repo = root / name
+        task = strategy_probe_task(
+            strategy="evidence-first",
+            reason="fixture; nothing here can fail first",
+            form=form,
+            commands=checks,
+        )
+        task = task.replace("- [ ] 1.1", "- [x] 1.1")
+        task = task.replace("      - Findings: none", f"      - Findings: {findings}")
+        for entry in checks:
+            lbl = entry.split(":", 1)[0].split(" ")[0]
+            task = task.replace(f"    - {lbl}: pending", f"    - {lbl}: pass. ran it.")
+        write_gate_fixture(repo, tasks=task)
+        started = run_keel(
+            repo, "gate", "task-start", "--change", "demo", "--task", "1.1",
+            "--json", "--no-guard",
+        )
+        payload = json.loads(started.stdout)
+        value = ((payload.get("contract") or {}).get("fingerprint") or {}).get(
+            "value"
+        ) or "0" * 64
+        tasks_path = repo / "openspec/changes/demo/tasks.md"
+        tasks_path.write_text(
+            tasks_path.read_text(encoding="utf-8").replace(
+                "    - Contract: pending",
+                f"    - Contract: keel-task-capsule/v1 sha256:{value}",
+            ),
+            encoding="utf-8",
+        )
+        result = run_keel(
+            repo, "gate", "task-complete", "--change", "demo", "--task", "1.1",
+            "--json",
+        )
+        return json.loads(result.stdout)
+
+    slotted = (
+        "M1: the first check asserts the public behavior",
+        "M2: the second check runs make sta RUN=<experiment_id> and succeeds",
+    )
+    clean = (
+        "M1: the first check asserts the public behavior",
+        "M2: the second check runs the suite and succeeds",
+    )
+
+    with tempfile.TemporaryDirectory(prefix="keel-reference-") as raw:
+        root = Path(raw)
+
+        reported = complete(root, "slotted", slotted, "one. Resolved here: M2")
+        text = problem_text(reported)
+        if "<experiment_id>" not in text:
+            report(
+                f"{label}: the unfilled slot is no longer reported; {text!r}."
+            )
+            return 1
+        if "not a check this task declares" in text:
+            report(
+                f"{label}: a reference to M2 was reported as undeclared because "
+                f"M2's own declaration failed to compile; {text!r}."
+            )
+            return 1
+
+        settled = complete(root, "clean", clean, "one. Resolved here: M2")
+        if settled.get("status") != "pass":
+            report(
+                f"{label}: the same task without the slot no longer passes; "
+                f"{problem_text(settled)!r}."
+            )
+            return 1
+
+        # The set narrowed to the truth, not to everything.
+        absent = complete(root, "absent", clean, "one. Resolved here: M9")
+        if "not a check this task declares" not in problem_text(absent):
+            report(
+                f"{label}: a reference to a check the task never declared was "
+                f"accepted; {problem_text(absent)!r}."
+            )
+            return 1
+
+        legacy = complete(
+            root, "legacy", clean, "one. Resolved here: M2", form="Commands"
+        )
+        if legacy.get("status") != "pass":
+            report(
+                f"{label}: an expanded v3 task lost its declared labels; "
+                f"{problem_text(legacy)!r}."
+            )
+            return 1
+
+    report(f"{label} scenario passed.")
+    return 0
+
+
+# The red-green obligation is a static function of the strategy and the tags,
+# and the `Verify` block is complete at task-start while the Evidence is all
+# pending. Issue #112 hit it three times, this repository twice more: the rule
+# was first heard after the capsule was written, the task implemented, the
+# checks run, and the Evidence recorded.
+def validate_obligation_is_stated_early_scenario() -> int:
+    label = "the-obligation-is-stated-early"
+
+    def start(root: Path, name: str, strategy: str, reason=None, checks=None):
+        repo = root / name
+        write_gate_fixture(
+            repo,
+            tasks=strategy_probe_task(
+                strategy=strategy,
+                reason=reason,
+                commands=checks
+                or (
+                    "M1: the first check asserts the public behavior",
+                    "M2 (regression): the second asserts something green stays green",
+                ),
+            ),
+        )
+        result = run_keel(
+            repo, "gate", "task-start", "--change", "demo", "--task", "1.1",
+            "--json", "--no-guard",
+        )
+        return json.loads(result.stdout)
+
+    with tempfile.TemporaryDirectory(prefix="keel-obligation-") as raw:
+        root = Path(raw)
+
+        redgreen = start(root, "redgreen", "vertical-tdd")
+        if redgreen.get("status") != "pass":
+            report(
+                f"{label}: stating the obligation changed the verdict; "
+                f"{problem_text(redgreen)!r}."
+            )
+            return 1
+        warnings = " ".join(str(x) for x in (redgreen.get("warnings") or []))
+        if ".red" not in warnings or ".green" not in warnings:
+            report(
+                f"{label}: task-start did not state the red-green obligation; "
+                f"{warnings!r}."
+            )
+            return 1
+        obligation = next(
+            (w for w in redgreen["warnings"] if ".red" in str(w)), ""
+        )
+        if "M1" not in obligation:
+            report(
+                f"{label}: the obligation does not name the check that owes it; "
+                f"{obligation!r}."
+            )
+            return 1
+        if "M2" not in obligation or "regression" not in obligation:
+            report(
+                f"{label}: the obligation does not name the exempt check; "
+                f"{obligation!r}."
+            )
+            return 1
+
+        quiet = start(
+            root,
+            "quiet",
+            "evidence-first",
+            reason="fixture; nothing here can fail first",
+        )
+        quiet_warnings = " ".join(str(x) for x in (quiet.get("warnings") or []))
+        if ".red" in quiet_warnings:
+            report(
+                f"{label}: a strategy without red-green was told about it; "
+                f"{quiet_warnings!r}."
+            )
+            return 1
+
+        # A task that fails task-start for another reason still fails with its
+        # own problem: the warning is not a verdict and cannot mask one.
+        broken = start(root, "broken", "made-up-thing")
+        if broken.get("status") != "fail":
+            report(f"{label}: an unsupported strategy stopped failing.")
+            return 1
+        if "unsupported" not in problem_text(broken):
+            report(
+                f"{label}: an unsupported strategy lost its own diagnostic; "
+                f"{problem_text(broken)!r}."
+            )
+            return 1
+
+    report(f"{label} scenario passed.")
+    return 0
+
+
+# Measured on issue #112's own shape: a two-check vertical-tdd task with no
+# red-green Evidence produced four problems in 827 characters, of which the
+# same 84-character rule sentence was four copies. It scales with the number
+# of checks, and the reporter estimates a third of failure output is this.
+def validate_explanation_is_printed_once_scenario() -> int:
+    label = "an-explanation-is-printed-once"
+    shared = "Tag the check `(regression)` if it asserts that something already green stays green."
+    with tempfile.TemporaryDirectory(prefix="keel-explanation-") as raw:
+        root = Path(raw)
+        repo = root / "repo"
+        task = strategy_probe_task(
+            strategy="vertical-tdd",
+            commands=(
+                "M1: the first check asserts the public behavior",
+                "M2: the second check asserts the public behavior",
+            ),
+        )
+        task = task.replace("- [ ] 1.1", "- [x] 1.1")
+        for lbl in ("M1", "M2"):
+            task = task.replace(f"    - {lbl}: pending", f"    - {lbl}: pass. ran it.")
+        write_gate_fixture(repo, tasks=task)
+        started = run_keel(
+            repo, "gate", "task-start", "--change", "demo", "--task", "1.1",
+            "--json", "--no-guard",
+        )
+        value = json.loads(started.stdout)["contract"]["fingerprint"]["value"]
+        tasks_path = repo / "openspec/changes/demo/tasks.md"
+        tasks_path.write_text(
+            tasks_path.read_text(encoding="utf-8").replace(
+                "    - Contract: pending",
+                f"    - Contract: keel-task-capsule/v1 sha256:{value}",
+            ),
+            encoding="utf-8",
+        )
+
+        rendered = run_keel(
+            repo, "gate", "task-complete", "--change", "demo", "--task", "1.1"
+        ).stdout
+        copies = rendered.count(shared)
+        if copies != 1:
+            report(
+                f"{label}: the shared rule explanation appears {copies} times "
+                "in the rendered text; it belongs once per run."
+            )
+            return 1
+        problems = [
+            line for line in rendered.splitlines() if line.startswith("Problem:")
+        ]
+        if len(problems) != 4:
+            report(
+                f"{label}: expected the four per-label problems, got "
+                f"{len(problems)}: {problems!r}."
+            )
+            return 1
+        for needed in ("M1.red", "M1.green", "M2.red", "M2.green"):
+            if not any(needed in line for line in problems):
+                report(
+                    f"{label}: {needed} lost its own problem line; {problems!r}."
+                )
+                return 1
+        if len(rendered) >= 827:
+            report(
+                f"{label}: the rendered failure is {len(rendered)} characters, "
+                "no shorter than the 827 measured before deduplication."
+            )
+            return 1
+
+        payload = json.loads(
+            run_keel(
+                repo, "gate", "task-complete", "--change", "demo", "--task",
+                "1.1", "--json",
+            ).stdout
+        )
+        carried = [
+            entry for entry in (payload.get("problems") or [])
+            if entry.get("code") == "missing-strategy-evidence"
+        ]
+        if len(carried) != 4:
+            report(
+                f"{label}: the JSON result no longer carries all four problems; "
+                f"{len(carried)}."
+            )
+            return 1
+        if not all(shared in json.dumps(entry, ensure_ascii=False) for entry in carried):
+            report(
+                f"{label}: the JSON result was deduplicated; a consumer reading "
+                "problems individually would get a payload whose content "
+                "depends on position."
+            )
+            return 1
+
+    report(f"{label} scenario passed.")
+    return 0
+
+
 # A scenario name, as the registry spells one. Two registered names carry no
 # hyphen — `cli` and `uninstall` — so requiring one would leave exactly those
 # two unchecked, and allowing single words was measured to add no false
@@ -25635,6 +26046,10 @@ SCENARIOS: tuple = (
     ("a-strategy-is-declared", validate_strategy_is_declared_scenario),
     ("the-weakest-strategy-states-its-reason", validate_weakest_strategy_states_its_reason_scenario),
     ("a-quoted-marker-is-not-a-disposition", validate_quoted_marker_is_not_a_disposition_scenario),
+    ("drift-names-where-to-look", validate_drift_names_where_to_look_scenario),
+    ("a-reference-outlives-its-declaration", validate_reference_outlives_its_declaration_scenario),
+    ("the-obligation-is-stated-early", validate_obligation_is_stated_early_scenario),
+    ("an-explanation-is-printed-once", validate_explanation_is_printed_once_scenario),
     (
         "authored-scenario-names-are-registered",
         validate_authored_scenario_names_scenario,

package/src/core/context.js

@@ -85,6 +85,32 @@ function taskHasCompletionEvidence(record, contract) {
   );
 }
 
+// Two hashes and nothing to search was the whole message, and it produced a
+// wrong record here: an anchor that moved was written up as a Review edit,
+// which is measurably not covered at all. The capsule already knows the answer
+// — every authority entry carries the source its text was resolved from — so
+// the search set is free. What is not free is the field that moved: only the
+// previous fingerprint is retained, not the capsule behind it, so this names
+// where to look and never what changed.
+function driftSearchSet(contract) {
+  const sources = [
+    ...new Set(
+      (contract.capsule.authority || [])
+        .map((entry) => String(entry.source || "").split("#")[0])
+        .filter(Boolean)
+    ),
+  ];
+  const where = sources.length > 0
+    ? `The fingerprint covers text resolved from: ${sources.join(", ")}.`
+    : "The fingerprint covers this task's own resolved authority text.";
+  return (
+    `${where} Evidence, Review, and the task checkbox are not covered — `
+    + "editing them does not move it. Reauthorize by re-running "
+    + "`keel gate task-start` and recording the new anchor, after confirming "
+    + "the change to the authority above was intended."
+  );
+}
+
 function taskSelection(repo, change, record, source, requestedAction = null) {
   const tasksPath = path.join(repo, "openspec", "changes", change, "tasks.md");
   const contract = compileTaskContract(repo, change, record);
@@ -100,7 +126,8 @@ function taskSelection(repo, change, record, source, requestedAction = null) {
   if (anchor && anchor !== contract.fingerprint.value) {
     return blocked(
       `Task contract fingerprint drift for ${change}#${record.id}: recorded `
-        + `sha256:${anchor}, current sha256:${contract.fingerprint.value}.`,
+        + `sha256:${anchor}, current sha256:${contract.fingerprint.value}. `
+        + `${driftSearchSet(contract)}`,
       [relativePath(repo, tasksPath)]
     );
   }

package/src/core/gates.js

@@ -9,6 +9,7 @@ const {
   ACCEPTED_REVIEW_STATUSES,
   RED_GREEN_VERIFICATION_STRATEGIES,
   compileTaskContract,
+  declaredCommandLabels,
   field,
   isConcrete,
   isPassingReviewStatus,
@@ -21,8 +22,15 @@ const GATE_STAGES = new Set(["task-start", "task-complete", "change-close"]);
 
 class GateInputError extends Error {}
 
-function problem(code, message) {
-  return { code, message };
+// `note` is the rule behind the problem rather than the problem itself. Every
+// problem carries its own copy in the JSON result, because a consumer reading
+// problems one at a time must not get a payload whose content depends on
+// position; the text renderer prints each distinct note once, because there
+// the repetition is what crowds out the specific lines. Measured on a
+// two-check task: the same 84-character sentence four times in 827 characters,
+// and it scales with the number of checks.
+function problem(code, message, note = null) {
+  return note ? { code, message, note } : { code, message };
 }
 
 function gateResult(
@@ -212,6 +220,34 @@ function anchoredFingerprint(previous) {
 // way to acknowledge a `needs-review`, so making it one would leave a
 // legitimate split unstartable. The reader is given the other task's id and
 // compares two things, rather than being told something is wrong.
+// What completion will require, said while the Evidence is still all pending.
+// The `Verify` block is complete at task-start and the obligation is a static
+// function of the strategy and the tags, so the only reason it was first heard
+// at task-complete is that nobody said it earlier — after the capsule was
+// written, the task implemented, the checks run, and the Evidence recorded.
+// A warning, never a refusal: a task whose author has not yet decided which
+// check is a regression is not malformed, it is unfinished.
+function redGreenObligation(compiled) {
+  if (!compiled || compiled.diagnostics.length > 0) return [];
+  const strategy = compiled.capsule.verification.strategy.toLowerCase();
+  if (!RED_GREEN_VERIFICATION_STRATEGIES.has(strategy)) return [];
+  const commands = compiled.capsule.verification.commands;
+  const owing = commands.filter((item) => !item.regression).map((item) => item.label);
+  const exempt = commands.filter((item) => item.regression).map((item) => item.label);
+  if (owing.length === 0) return [];
+  return [
+    `${strategy} will require concrete .red and .green Evidence at completion `
+      + `for ${owing.join(", ")}`
+      + (exempt.length > 0
+        ? `; ${exempt.join(", ")} ${exempt.length > 1 ? "are" : "is"} exempt as `
+          + "(regression)"
+        : "")
+      + ". Tag a check `(regression)` now if it asserts that something already "
+      + "green stays green, rather than discovering the obligation once the "
+      + "checks have been run.",
+  ];
+}
+
 function taskShapeWarnings(repo, selection, task, compiled) {
   if (!compiled || compiled.diagnostics.length > 0) return [];
   const strategy = compiled.capsule.verification.strategy.toLowerCase();
@@ -272,7 +308,10 @@ function taskStart(repo, options) {
     selection.change,
     [task.id],
     problems,
-    taskShapeWarnings(repo, selection, task, compiled),
+    [
+      ...redGreenObligation(compiled),
+      ...taskShapeWarnings(repo, selection, task, compiled),
+    ],
     problems.length === 0
       ? compiled
       : null
@@ -887,9 +926,13 @@ function attributeChanged(repo, task, changedList, contract, change, tasks) {
 
 function completionChecks(repo, task, contract = null, changeVerify = null, change = null) {
   const problems = [];
+  // Not the compiled capsule alone: when a check's declaration does not
+  // compile there is no capsule, and reconstructing the set from the expanded
+  // v3 `Commands` field a compact task never declares makes every reference
+  // look undeclared. The task's own declarations answer in both cases.
   const commands = contract
     ? contract.capsule.verification.commands.map((item) => item.label)
-    : commandLabels(task);
+    : declaredCommandLabels(task);
   // With no contract, the labels came from the expanded v3 `Commands` field,
   // which a compact task never declares — so their absence is a fact about the
   // fallback, not about the task. The compiler's own diagnostics are already in
@@ -926,8 +969,9 @@ function completionChecks(repo, task, contract = null, changeVerify = null, chan
             problem(
               "missing-strategy-evidence",
               `${strategy} requires concrete ${label}.${phase} Evidence for `
-                + "the same behavior check. Tag the check `(regression)` if it "
-                + "asserts that something already green stays green."
+                + "the same behavior check.",
+              "Tag the check `(regression)` if it asserts that something "
+                + "already green stays green."
             )
           );
         }
@@ -1626,6 +1670,11 @@ function renderGate(result) {
         : ""),
   ];
   for (const item of result.problems) lines.push(`Problem: ${item.message}`);
+  for (const note of [
+    ...new Set(result.problems.map((item) => item.note).filter(Boolean)),
+  ]) {
+    lines.push(`Note: ${note}`);
+  }
   for (const warning of result.warnings) lines.push(`Warning: ${warning}`);
   if (result.contract) {
     lines.push(

package/src/core/task-contract.js

@@ -229,6 +229,21 @@ function verification(task) {
   };
 }
 
+// The labels a task declares, read from the form the task itself uses. Label
+// parsing does not depend on a check being concrete, so this answers exactly
+// when the compiler cannot: a task whose `M2` declaration carries an unfilled
+// slot still declares `M2`, and a reference to it is not a reference to
+// something that does not exist.
+function declaredCommandLabels(task) {
+  const source = fieldValues(task, "Verify").length > 0
+    ? fieldValues(task, "Verify")
+    : fieldValues(task, "Commands");
+  return source
+    .map((entry) => entry.match(/^(M[1-9]\d*)(?:\s*\([^)\n]*\))?\s*:/))
+    .filter(Boolean)
+    .map((match) => match[1]);
+}
+
 function commandLabelProblems(task) {
   // A task that declared no verification form at all is reported once, by
   // requiredFieldProblems, as the one field it is missing. Its orphan Evidence
@@ -1195,6 +1210,7 @@ function loadTaskContract(repo, change, taskId) {
 
 module.exports = {
   ACCEPTED_REVIEW_STATUSES,
+  declaredCommandLabels,
   RED_GREEN_VERIFICATION_STRATEGIES,
   SUPPORTED_VERIFICATION_STRATEGIES,
   compileTaskContract,