feat: lessons from what recurs about one setting, and three more approved
Some checks failed
validation / verify (push) Failing after 10s

Evidence was grouped only by the kind of problem, and a group gets one lesson,
so what kept recurring inside "validation proposed a Check" could not be
seen. A proposed or unimplementable Check is now also counted under each
provider setting or resource type it names; a subject seen in two or more
Executions is its own group and reaches the curator past triage.

Over 32 Executions that gave five proposals. The owner approved three, now
published: L-07 location Checks, L-08 Cloud SQL backup encryption, L-09
backend service capabilities. L-10 is held for more evidence and L-11
rejected.

Also: retiring a lesson marks its proposal, so publish no longer brings a
retired lesson back. L-04 had returned to the Skill that way.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-10-10 12:47:18 +01:00
parent ee1ba38867
commit 5feaaa6271
17 changed files with 354 additions and 13 deletions

View File

@@ -81,3 +81,17 @@ Neither run raised the schema review's false alarm, which had been raised in cas
- **`status` says "compared", not "proved",** and shows a retired lesson as retired. One pair of runs is a comparison.
Tests: `tests/test_lessons_loop.py`.
## Evidence is also grouped by what it is about (2026-10-10)
Running `propose` over 32 Executions gave nothing useful. Evidence was grouped by the kind of problem, and a group gets one lesson: "validation proposed a Check the design lacked" was 20 of 32 Executions and already had L-01. What recurred inside it could not be seen.
- A proposed Check, and a Check sent back as unimplementable, is now also counted under each provider setting or resource type its text names (`_subjects` in `app/learning/evidence.py`). A subject seen in two or more Executions is its own group, with its own key. A setting's name is the provider's vocabulary, not a client's.
- These narrower groups skip triage: each recurred with the general lessons published. The curator is shown the existing lessons and may still decline.
- The curator's instructions say such a lesson may name the setting or resource type.
Over the same 32 Executions this gave 22 groups and five proposals, L-07 to L-11, all awaiting the owner: location Checks, Cloud SQL backup encryption, backend service capabilities, Pub/Sub persistence regions, and customer-managed keys.
A first run of `propose` before this change produced one proposal, about giving `edit_file` enough context. It was removed to run again cleanly and was not proposed the second time.
Tests: `tests/test_lessons_loop.py` (3 more).

View File

@@ -5,6 +5,7 @@ Deliverable: target_architecture
Evidence-key: 22ff1f5fed
Proposed: 2026-10-09
Published: 2026-10-09
Retired: 2026-10-09
## Rule

View File

@@ -0,0 +1,27 @@
# L-07: Verify location constraints
Status: ready-for-agent
Deliverable: target_architecture
Evidence-key: 493b0a95d9
Proposed: 2026-10-10
Published: 2026-10-10
## Rule
When a requirement specifies where resources must be provisioned, such as a designated region, multi-zone placement, or a low-carbon location, create a Check for the `region` or `location` argument on every applicable Element.
## Why
Validation frequently catches missing location constraints. Without this Check, resources may be provisioned in default regions, violating compliance or sustainability requirements.
## Evidence
Seen in 5 of 32 Executions, 8 times in all: validation proposed a Check about `location` that the design lacked.
## Proof
Not yet measured.
## Comments
Approved by the owner on 2026-10-10. It covers which region is set, which code can show. It does not settle whether a region is low-carbon, which code cannot show.

View File

@@ -0,0 +1,27 @@
# L-08: Cloud SQL backup encryption
Status: ready-for-agent
Deliverable: target_architecture
Evidence-key: 08d64df049
Proposed: 2026-10-10
Published: 2026-10-10
## Rule
When writing a Check for CMEK encryption on a `google_sql_database_instance`, do not require an `encryption_key_name` argument for backup storage, as the instance-level `encryption_key_name` automatically encrypts backups as well.
## Why
The database instance resource does not have a separate encryption key argument for backups. Asking for one causes validation to reject the Check as un-implementable.
## Evidence
Seen in 3 of 32 Executions, 3 times in all: a Check about `google_sql_database_instance` asked for something that resource cannot do.
## Proof
Not yet measured.
## Comments
Approved by the owner on 2026-10-10.

View File

@@ -0,0 +1,27 @@
# L-09: Backend Service capabilities
Status: ready-for-agent
Deliverable: target_architecture
Evidence-key: 1cacec9a09
Proposed: 2026-10-10
Published: 2026-10-10
## Rule
When writing Checks for a `google_compute_backend_service`, do not verify IAM roles inside the backend service or require health checks when it is configured with a Serverless NEG.
## Why
Roles are verified via IAM bindings outside the backend service, and Serverless NEGs do not support health checks. Invalid Checks cause the design to be rejected during validation.
## Evidence
Seen in 3 of 32 Executions, 3 times in all: a Check about `google_compute_backend_service` asked for something that resource cannot do.
## Proof
Not yet measured.
## Comments
Approved by the owner on 2026-10-10.

View File

@@ -0,0 +1,26 @@
# L-10: Pub/Sub persistence regions
Status: needs-info
Deliverable: target_architecture
Evidence-key: ba78caba9a
Proposed: 2026-10-10
## Rule
When a requirement dictates that a messaging queue or topic must replicate across zones or stay within a specific region, create a Check for the `message_storage_policy.allowed_persistence_regions` argument on the topic Element.
## Why
Without this Check, topics might use default storage policies that violate geographical or high-availability requirements.
## Evidence
Seen in 2 of 32 Executions, 4 times in all: validation proposed a Check about `allowed_persistence_regions` that the design lacked.
## Proof
Not yet measured.
## Comments
Held by the owner on 2026-10-10, on the agent's view: seen in 2 of 32 Executions, and L-07 partly covers it. To be looked at again if it recurs.

View File

@@ -0,0 +1,26 @@
# L-11: CMEK encryption checks
Status: wontfix
Deliverable: target_architecture
Evidence-key: 5cd8dc3cda
Proposed: 2026-10-10
## Rule
Whenever a requirement mandates encryption at rest, explicitly create a Check ensuring that the `kms_key_name` (or the equivalent CMEK argument) is configured on every applicable Element.
## Why
Without an explicit Check, elements may default to Google-managed keys, leading to validation failures against security constraints.
## Evidence
Seen in 2 of 32 Executions, 2 times in all: validation proposed a Check about `kms_key_name` that the design lacked.
## Proof
Not yet measured.
## Comments
Rejected by the owner on 2026-10-10, on the agent's view: encryption at rest is on by default, and this rule would put customer-managed keys on designs whose requirements did not ask for them.

View File

@@ -8,7 +8,7 @@ You review what went wrong, the same way, across many solution-architecture Exec
Your task message carries JSON context:
- `executions`: how many Executions the evidence comes from.
- `groups`: what recurred. Each has a `group` key, the `deliverable` it concerns, `what_happened` in general terms, how many `executions` it occurred in and how many `occurrences` in all, and up to three `examples` as they were recorded.
- `groups`: what recurred. Each has a `group` key, the `deliverable` it concerns, `what_happened` in general terms, how many `executions` it occurred in and how many `occurrences` in all, and up to three `examples` as they were recorded. Some groups are about one setting or resource type of the Cloud provider, named in `what_happened`: the same kind of problem recurring on that subject.
- `published_or_proposed`: lessons that already exist or are awaiting review.
For each group, decide whether a rule given to the Specialist that produces that Deliverable would have prevented it.
@@ -20,6 +20,7 @@ A good lesson:
- States one rule, as an instruction, that a Specialist can act on without knowing what happened: "Give every Element at least one Check before you submit", not "be more careful with Checks".
- Says why in one or two sentences: what goes wrong without it and what that costs.
- For a group about one setting or resource type, may name that setting or resource type and say what to do about it: that is the Cloud provider's vocabulary, not a client's. A general lesson that already exists does not cover it if the problem went on recurring.
- Is general. It names no customer, no project, no Execution, no resource name and no value from the examples. The examples are evidence for you; nothing from them may be quoted.
- Does not repeat what the Specialist's instructions or tools already enforce. When a submission is already refused for something, a lesson helps only if it prevents the attempt.
- Is worth its cost. Every lesson is read on every Execution from then on. Propose few.

View File

@@ -101,7 +101,12 @@ async def curate(
decided = {proposal.evidence_key for proposal in existing}
fresh = [group for group in groups if group.key not in decided][:MAX_GROUPS]
if decider is not None and existing:
fresh = await _not_already_covered(decider, fresh, existing, covered_at)
# A group about one setting is narrower than any lesson about its kind of
# problem, and it recurred with those lessons published: it goes to the
# curator, who is shown the lessons that exist and may still decline.
broad = [group for group in fresh if group.subject is None]
kept = await _not_already_covered(decider, broad, existing, covered_at)
fresh = [g for g in fresh if g.subject is not None or g in kept]
if not fresh:
return []
by_key = {group.key: group for group in fresh}

View File

@@ -10,6 +10,8 @@ from dataclasses import dataclass, field
from pathlib import Path
from typing import Any
from app.tools.schema_review import named_settings
EXAMPLES = 3
EXAMPLE_LIMIT = 400
# Ids and numbers vary between Executions; what recurs is the shape around them.
@@ -29,6 +31,9 @@ class Group:
executions: set[str] = field(default_factory=set)
occurrences: int = 0
examples: list[str] = field(default_factory=list) # client data: never published
# The provider setting or resource type the problem was about, for a group
# narrowed to one subject. None for a group of every problem of its kind.
subject: str | None = None
def as_context(self) -> dict[str, Any]:
return {
@@ -61,6 +66,10 @@ def read_candidates(roots: Iterable[Path]) -> list[dict[str, Any]]:
return records
# A subject shorter than this is a stray word, not a setting.
_SUBJECT_MIN = 4
def _general(text: str, limit: int = 110) -> str:
return " ".join(_VARIABLE.sub("…", text).split())[:limit]
@@ -104,6 +113,31 @@ def _shape(candidate: Mapping[str, Any]) -> tuple[str, str | None, str] | None:
return None
def _subjects(candidate: Mapping[str, Any]) -> tuple[str, list[str]] | None:
"""For a problem about what a Check says: (how to describe it, the settings and
resource types its text names).
Problems of one kind are one group, and one group gets one lesson. "Validation
proposed a Check" was 20 of 32 Executions and a single lesson; what recurred
inside it, such as the same setting missed again and again, was invisible. A
setting's name is the provider's vocabulary, not a client's, so it can be the
subject of a narrower group.
"""
kind = str(candidate.get("type"))
if kind == "proposed_addition":
text = str(candidate.get("statement", ""))
what = "validation proposed a Check about `{}` that the design lacked"
elif kind == "revision" and "cannot be implemented" in str(
candidate.get("reason", "")
):
text = str(candidate.get("reason", ""))
what = "a Check about `{}` asked for something that resource cannot do"
else:
return None
names = {name.split(".")[-1] for name in named_settings(text)}
return what, sorted(name for name in names if len(name) >= _SUBJECT_MIN)
def group_candidates(
records: Sequence[Mapping[str, Any]], *, min_executions: int = 2
) -> list[Group]:
@@ -138,6 +172,23 @@ def group_candidates(
and len(group.examples) < EXAMPLES
):
group.examples.append(example)
narrowed = _subjects(candidate)
for subject in narrowed[1] if narrowed else ():
about = narrowed[0].format(subject) # type: ignore[index]
subkey = hashlib.sha256(
f"{kind}|{deliverable}|{about}".encode()
).hexdigest()[:10]
narrow = groups.setdefault(
subkey, Group(subkey, kind, deliverable, about, subject=subject)
)
narrow.executions.add(execution)
narrow.occurrences += 1
if (
example
and example not in narrow.examples
and len(narrow.examples) < EXAMPLES
):
narrow.examples.append(example)
return sorted(
(g for g in groups.values() if len(g.executions) >= min_executions),
key=lambda g: (-len(g.executions), -g.occurrences, g.key),

View File

@@ -37,6 +37,9 @@ class Proposal:
status: str = NEEDS_TRIAGE
proposed: str = ""
published: str = ""
# The day it was taken out of the Skill again. A retired lesson is not
# published a second time, whatever its Status still says.
retired: str = ""
proof: str = ""
comments: str = ""
path: Path | None = None
@@ -47,7 +50,7 @@ class Proposal:
@property
def approved(self) -> bool:
return self.status == APPROVED
return self.status == APPROVED and not self.retired
def _slug(title: str) -> str:
@@ -65,6 +68,8 @@ def render(proposal: Proposal) -> str:
]
if proposal.published:
lines.append(f"Published: {proposal.published}")
if proposal.retired:
lines.append(f"Retired: {proposal.retired}")
lines += [
"",
"## Rule",
@@ -117,6 +122,7 @@ def parse(path: Path) -> Proposal:
status=status,
proposed=fields.get("Proposed", ""),
published=fields.get("Published", ""),
retired=fields.get("Retired", ""),
proof="" if proof == "Not yet measured." else proof,
comments=sections.get("Comments", ""),
path=path,

View File

@@ -111,6 +111,8 @@ def publish(
Raises NotApproved for any proposal a person has not approved: nothing reaches
an agent on a curator's say-so.
"""
if proposal.retired:
raise NotApproved(f"{proposal.id} was retired on {proposal.retired}")
if not proposal.approved:
raise NotApproved(
f"{proposal.id} is {proposal.status}; a person approves a lesson by "

View File

@@ -35,6 +35,8 @@ from __future__ import annotations
import asyncio
import sys
from collections.abc import Sequence
from dataclasses import replace
from datetime import UTC, datetime
from pathlib import Path
from app.agent import open_decider
@@ -42,7 +44,7 @@ from app.config import get_settings
from app.learning.curator import LESSON_CURATOR, curate
from app.learning.evidence import group_candidates, identifiers, read_candidates
from app.learning.proof import compare, record, render
from app.learning.proposals import ISSUES, load
from app.learning.proposals import ISSUES, load, write
from app.learning.skill import (
NotApproved,
mark_reviewed,
@@ -141,7 +143,7 @@ def show_status() -> int:
"in the Skill"
if proposal.id in live
else "retired"
if proposal.published
if proposal.retired
else "not in the Skill"
)
# One comparison suggests a direction; it does not prove a lesson.
@@ -211,6 +213,10 @@ def retire_lesson(arguments: Sequence[str]) -> int:
return 2
bundle = load_provider(get_settings().CLOUD_PROVIDER)
gone = retire(arguments[0], _skill_dir(bundle))
for proposal in load(ROOT / ISSUES):
if proposal.id == arguments[0] and not proposal.retired:
# Marked on its proposal, so the next publish does not bring it back.
write(replace(proposal, retired=datetime.now(UTC).date().isoformat()))
print(f"Retired {arguments[0]}." if gone else f"{arguments[0]} is not published.")
return 0 if gone else 1

View File

@@ -18,7 +18,7 @@ Each reference below holds the lessons for one Deliverable. Read the one for the
| Deliverable | Reference | Lessons |
| --- | --- | --- |
| deployment guidance | `references/deployment_guidance.md` | 1 |
| target architecture | `references/target_architecture.md` | 4 |
| target architecture | `references/target_architecture.md` | 7 |
<!-- lessons:end -->
This Skill is written by the lessons loop (`app/learning/`), from proposals a person approved. Do not edit it by hand: change the proposal and publish again.

View File

@@ -8,4 +8,4 @@ Rules learned from earlier work. Follow every one that applies.
**Why.** Missing required arguments in nested blocks will cause Terraform validation to fail, preventing submission.
Published 2026-10-09, from lesson proposal 05.
Published 2026-10-10, from lesson proposal 05.

View File

@@ -8,7 +8,7 @@ Rules learned from earlier work. Follow every one that applies.
**Why.** If detailed constraints are missed, the validation phase will reject the design for incomplete coverage, forcing a revision.
Published 2026-10-09, from lesson proposal 01.
Published 2026-10-10, from lesson proposal 01.
## L-02: Assign Checks to Every Element
@@ -16,7 +16,7 @@ Published 2026-10-09, from lesson proposal 01.
**Why.** The submission tool will reject the architecture if any Element lacks a Check, wasting an attempt.
Published 2026-10-09, from lesson proposal 02.
Published 2026-10-10, from lesson proposal 02.
## L-03: Verify Resource Arguments for Checks
@@ -24,7 +24,7 @@ Published 2026-10-09, from lesson proposal 02.
**Why.** If a Check specifies an argument that does not exist on the resource, it cannot be implemented, forcing an architectural revision.
Published 2026-10-09, from lesson proposal 03.
Published 2026-10-10, from lesson proposal 03.
## L-06: Do Not Mark Requirements as Not Applicable
@@ -32,4 +32,28 @@ Published 2026-10-09, from lesson proposal 03.
**Why.** The submission tool will reject the deliverable if a Requirement lacks a Check or is incorrectly marked as not applicable.
Published 2026-10-09, from lesson proposal 06.
Published 2026-10-10, from lesson proposal 06.
## L-07: Verify location constraints
**Rule.** When a requirement specifies where resources must be provisioned, such as a designated region, multi-zone placement, or a low-carbon location, create a Check for the `region` or `location` argument on every applicable Element.
**Why.** Validation frequently catches missing location constraints. Without this Check, resources may be provisioned in default regions, violating compliance or sustainability requirements.
Published 2026-10-10, from lesson proposal 07.
## L-08: Cloud SQL backup encryption
**Rule.** When writing a Check for CMEK encryption on a `google_sql_database_instance`, do not require an `encryption_key_name` argument for backup storage, as the instance-level `encryption_key_name` automatically encrypts backups as well.
**Why.** The database instance resource does not have a separate encryption key argument for backups. Asking for one causes validation to reject the Check as un-implementable.
Published 2026-10-10, from lesson proposal 08.
## L-09: Backend Service capabilities
**Rule.** When writing Checks for a `google_compute_backend_service`, do not verify IAM roles inside the backend service or require health checks when it is configured with a Serverless NEG.
**Why.** Roles are verified via IAM bindings outside the backend service, and Serverless NEGs do not support health checks. Invalid Checks cause the design to be rejected during validation.
Published 2026-10-10, from lesson proposal 09.

View File

@@ -10,9 +10,9 @@ from pathlib import Path
import pytest
from app.adk.fake_runtime import Call, FakeAgentRuntime
from app.adk.fake_runtime import Call, FakeAgentRuntime, FakeDecider
from app.adk.roles import load_roles
from app.adk.runtime import RoleSpec
from app.adk.runtime import Answer, RoleSpec
from app.learning.curator import LESSON_CURATOR, curate
from app.learning.evidence import group_candidates, identifiers, read_candidates
from app.learning.proposals import (
@@ -563,3 +563,101 @@ def test_the_published_lessons_were_reviewed_against_todays_vendored_skills():
assert upstream_reviewed(skill), "run: python -m eval.curate_lessons reviewed"
assert upstream_changes(skill, upstream_versions(bundle)) == []
# --- what recurs inside a kind of problem ------------------------------------------------
def addition(statement: str) -> dict:
return {"type": "proposed_addition", "statement": statement, "serves": "NFR-1"}
def test_the_same_setting_missed_in_several_executions_is_its_own_group():
records = [
record("run-1", addition("`deletion_protection` is true on the database")),
record("run-2", addition("deletion_protection is enabled"), addition("x y")),
record("run-3", addition("`min_tls_version` is TLS_1_2")),
]
groups = {group.subject: group for group in group_candidates(records)}
# Every proposed Check is one broad group; the setting missed twice is another.
assert len(groups[None].executions) == 3
narrow = groups["deletion_protection"]
assert len(narrow.executions) == 2 and narrow.kind == "proposed_addition"
assert narrow.description == (
"validation proposed a Check about `deletion_protection` that the design lacked"
)
assert narrow.key != groups[None].key
# Seen in one Execution only, a setting is not a pattern.
assert "min_tls_version" not in groups
def test_a_resource_that_checks_keep_asking_too_much_of_is_its_own_group():
def cannot(why: str) -> dict:
return {
"type": "revision",
"deliverable": "target_architecture",
"reason": f"Check K-1 cannot be implemented as written ({why}).",
}
records = [
record("run-1", cannot("example_database has no backup key argument")),
record("run-2", cannot("example_database has no query access flag")),
]
subjects = {group.subject for group in group_candidates(records)}
assert subjects == {None, "example_database"}
async def test_a_group_about_one_setting_reaches_the_curator_past_triage(tmp_path):
issues = tmp_path / "issues"
approve(proposal(issues), issues) # a general lesson already exists
records = [
record("run-1", addition("`deletion_protection` is true")),
record("run-2", addition("`deletion_protection` is true")),
]
groups = group_candidates(records)
# Triage says the general lesson covers everything it is asked about.
decider = FakeDecider(
{group.key: Answer("L-01", 0.95) for group in groups if group.subject is None}
)
seen: list[str] = []
def turn(task):
seen.extend(group["what_happened"] for group in task.context["groups"])
return [Call("finish")]
runtime = FakeAgentRuntime({LESSON_CURATOR: [turn]})
await curate(
runtime,
CURATOR,
groups,
executions=2,
folder=issues,
decider=decider,
)
assert seen == [
"validation proposed a Check about `deletion_protection` that the design lacked"
]
def test_a_retired_lesson_is_not_published_again(tmp_path):
"""Retiring left the proposal approved, so the next publish put the lesson
back: L-04 returned to the Skill that way on 2026-10-10."""
from dataclasses import replace
skill = lessons_skill(tmp_path)
issues = tmp_path / "issues"
made = publish(approve(proposal(issues), issues), skill, today=DAY)
retire(made.id, skill)
retired = write(replace(parse(made.path), retired="2026-10-09"), issues)
assert parse(retired.path).retired == "2026-10-09"
assert not retired.approved
with pytest.raises(NotApproved, match="was retired on 2026-10-09"):
publish(retired, skill)
assert published(skill) == {}