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>
- An Execution's files are written as each Deliverable is submitted, marked
in_progress, so a killed process or one waiting for its user leaves them on
disk. A write removes files an earlier write left that it no longer has.
- A question about delivered work is answered from that Execution's solution
guide by the reply writer; a question about the service is answered as
before. Nothing is started for either.
- Lessons: prove writes only for lessons the Skill holds; the Skill records
the vendored Skill versions the lessons were read against, status names any
that changed, and a test fails until they are reviewed.
- Four more code patterns: Firestore, BigQuery, Eventarc and GKE.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Infrastructure code is plain resources by default. An Execution may opt in
with terraform_modules; the deployment default is TERRAFORM_MODULES, off.
- A module is accepted only from the registry namespaces the provider bundle
lists (terraform-google-modules), at one exact version. Any other source is
refused before Terraform runs; until now nothing stopped generated code
naming any source for terraform init to download.
- terraform_module looks up a module's inputs, outputs and submodules from
the registry and says when it cannot be used with the pinned provider.
- skills/google-cloud-iac-modules: when a module fits, and four modules with
examples a test validates against hashicorp/google 8.6.0.
- Opted in, the IaC author and validator get the Skill and the lookup. Not
opted in, they get neither and a module block is refused in code.
- A Change request follows its Baseline. execution.json records the choice.
- case-07 runs case 03's request with modules opted into.
Azure (aztfm) and AWS (terraform-aws-modules) sources are recorded in the
plan for when those Cloud providers are added. No case has run with this.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Schema review: a dotted path is judged only when it starts at a name the
resource type has, so a database flag or a service name is no longer read
as a missing setting. Lesson L-04, which worked around that, is retired.
- Skill routing: in a Change request the question says what changed.
- Evaluation: a case may list code_standards the infrastructure code must
meet; case 06 checks its overlay's regions, naming and labels.
- Model tiers (iteration 5 step 6): a lighter model for a role's first
attempt, named in the provider bundle. Measured on case 03 and left off:
Flash saved five cents on the IaC author and took twice the requests.
- Steps 7 and 9 decided against on the numbers, recorded in the plans.
- Plans: the runs of cases 05 and 06, and the first proof of the lessons on
case 03 (no Revision, 22 minutes against 28, quality held).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Iteration 4:
- Overlay Skills (step 12): an Execution names an organisation's standards
Skill from the folders the deployment configures; recorded with its version.
An example standards Skill and an evaluation case that uses it.
- Lessons (step 10): a prove command that compares a run with lessons against
one without, and triage of evidence an existing lesson already covers.
Iteration 5:
- Step 8: the references of a Specialist's Skills its task needs are named in
its context as read_first.
- Step 14: a remote MCP server's reply that instructs the agent is withheld;
the user's own text is flagged as a risk and nothing more.
- Step 15: in a Change request, items only reworded are listed apart from
real changes.
- Step 19: each Execution's design recommendations are judged when it ends and
the scores kept in its record. The evaluation takes its rubric from the same
place.
Each question was calibrated against live Jev; none has run in a case. Steps
6, 7, 9, 13 and 18 stay planned, with the reasons in the plan map.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Capture, distil, review, publish, retire (ADR-0004). Lesson candidates of
past Executions are grouped into what recurred; a lesson-curator role,
outside any Execution, proposes a lesson per group worth one; each proposal
is an issue file a person approves or rejects; an approved lesson is
written into the provider's lessons Skill, which a Specialist is given only
once it holds lessons for that Specialist's Deliverable. LESSONS_ENABLED
off runs without them, for an evaluation with and without.
The curator proposed six lessons from 21 Executions. The owner approved all
six on 2026-10-09 and they are published: five for the target architect,
one for the IaC author. None is proved by an evaluation yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The target architect exhausted its 45 tool calls on a first attempt in
three of five runs: 31 schema lookups and 29 graph_put calls for one
design. Each time it carried on from its working graph, at a cost of four
to ten minutes.
- terraform_schema takes resource_types, a list, and outlines every
argument and block each accepts, by name, in one reply.
- The architect is told to make that one call once the Elements are
decided, and to build the graph in a handful of calls of ten to twenty
items.
Case 04 then completed at 100% for the first time, the architect
finishing in one attempt with 11 lookups and 13 graph calls.
Plans record the run and correct three step statuses.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Case 04 (interactive)
- A scripted answer may name who it is for, so facts written for
requirements discovery are not given to a Specialist confirming a draft.
- A reply that restates what the agreed requirements already say is read
as an answer, not a request for changes.
- With a user present, a first submission is refused until the user has
been shown a draft of that Deliverable (CONFIRM_DRAFTS).
Routing and reviews
- Intake: a message that starts something is read first. A question is
answered and starts nothing; a change to delivered work starts as a new
request with a note, until Change requests are built.
- Requirements and recommendations list what infrastructure code can show
(conditions) apart from what it cannot (outside_code); the latter needs
no Check and validation proposes nothing for it.
- A requirements review, "names no products", behind its own switch (off).
- Each pillar advisor reviews its pillar's checklist; the engine writes the
reviews into the design recommendations. The judged checklist check went
from 9-45% a pillar to 100% in one run of case 04.
Evaluation
- Per-case thresholds for Revisions, seconds and tokens, set from every
completed run on this build.
- A progress file after every agent run, so an interrupted run keeps its
time and tokens.
- eval/calibrate_judge.py with a 32-question set awaiting a person's
labels. Unlabelled: Jev stable at 1.00, Gemini 0.92, agreeing on 0.91.
Smaller
- An Execution that ends with warnings writes its Deliverables, with the
warnings and each status in the manifest.
- CI installs the application's dependencies before linting and testing.
- A load-balancer pattern in the IaC patterns Skill.
Plans record nine runs of 2026-10-08. Iteration 6 is complete.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
From the case 02 run cancelled at 66 minutes on 2026-10-07:
- The validator records verdicts a few Checks at a time (verdict_put) and
they are attached on submission. It timed out five times writing them in
one reply. Verdicts on the same code survive a run cut short; verdicts on
changed code are dropped. Its model timeout is 300 s.
- A schema review at submission, in code: every setting a Check names is
looked up on its Element's resource type, and an unknown resource type is
reported. Two Revisions in that run were for Checks no code could meet.
On the two designs that passed validation it raises 2 of 96 Checks, both
one metric name used as a value, so it is returned once and never blocks.
- Reviews run together; the Jev coverage review stays off.
Plans record both steps and the cancelled run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After case 02 (completed at 100%, but with two Revisions of the Target
architecture for gaps validation found one pass at a time):
- A Deliverable stands while what it was written from is unchanged. The
architecture description is written from the Elements and Relationships,
so a Revision that only adds or rewords Checks no longer runs the describer.
- Non-functional requirements and recommendations list their conditions,
one entry per separate thing that must be true; submission requires them.
- A coverage review at submission (JEV_COVERAGE_REVIEW, off by default):
for each condition, Jev chooses which Check verifies it, or none. Asked as
one broad yes-or-no per requirement it could not tell a complete design
from one with known gaps; asked as a choice per condition it raised all
four known gaps of case 02 and nothing on the complete graph. An opt-in
live test (LIVE_JEV=1) repeats that.
- The validator is told to report every gap in its first pass.
Plans record the case 02 run and both forms of the review with figures.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Start-up
- Export the env file's variables for other libraries (the Cloud provider's
client) into the process environment; the app's own settings and secrets
are never exported, and a value already in the environment wins.
- Use GOOGLE_GENAI_USE_ENTERPRISE in place of the deprecated
GOOGLE_GENAI_USE_VERTEXAI.
Target architect
- Look each distinct resource type up once instead of once per setting: in
the run of 2026-10-07 it made 39 lookups for 8 types and exhausted its
budget. terraform_schema's find takes several names; budget 30 to 45.
- A cut-off schema outline says so.
Removed, as leftovers of the original pipeline that nothing read
- terraform/ (no longer passed terraform validate), scripts/validate.sh,
tests/test_package.py and the CI steps that ran on them.
- The sample outputs under docs/ and requirements.yaml.
- CI now pins Terraform 1.14.3, as the container image does.
Plans record the case 01 run: completed at 100% in 18 minutes with no
Revision, against 60 minutes and warnings before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Design is now approved before any infrastructure code is written (ADR-0005).
A new target-architect Specialist produces one graph of Elements,
Relationships and Checks; the diagram is rendered from it, deployment
guidance must account for every Check, and validation gives a verdict per
Check. Requirements, components, products and recommendations carry ids in
data files, and a submission with a gap in the Requirements chain is refused.
Also in this commit:
- Customer cases group Executions, with an intent history per case
- lesson candidates are captured when an Execution ends
- evaluation gains chain, verdict, judged and interactive checks, and
per-case thresholds
- Terraform providers are cached between checks
- the requirements analyst proposes defaults instead of asking
- the legacy-baseline case and the checks that only read it are removed
- ADR-0006 (Change requests) accepted; ADR-0004 and ADR-0007 proposed
- plans for iterations 4 to 7 under .agents/plans/
Unproven with a real model: no evaluation run has completed on this design.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Increment 6 of .scratch/solution-architect-agent/spec.md, test-first.
- app/tools/mcp.py: MCP client and catalog exposing only allowlisted,
readOnlyHint tools as framework-neutral ToolSpecs; unreachable servers
and failed calls become discovery gaps; ADC auth with quota project.
- app/tools/local.py: Mermaid structure check and Terraform fmt,
validate and plan in a scratch copy; apply is never constructed.
- Tool factory gives roles their named local tools and bundle MCP
servers; pillar advisors get general-tier documentation tools.
- providers/gcp.yaml: 42 allowlisted tools reviewed against live
tools/list of 12 Google Cloud MCP servers (snapshot in
providers/reviews/), excluding data-reading and secret-capable tools.
- scripts/list-mcp-tools.py, docs/operations/mcp-access.md (IAM deny
policy, service account roles), cloud-discovery and iac-validation
Skills, MCP and Terraform settings, pytest asyncio auto mode.
- 31 new tests; full suite 189 passed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Increment 2 of .scratch/solution-architect-agent/spec.md.
- Vendor google-cloud-solution-architecture and the six
google-cloud-waf-* Skills from google/skills@81a31a6, unmodified,
with LICENSE, SHA256SUMS and VENDOR.md; add
scripts/sync-vendored-skills.sh to re-sync at a pinned sha.
- app/agent_skills.py: discover Skills and validate SKILL.md
frontmatter against the Agent Skills specification.
- app/providers.py + providers/gcp.yaml: load and validate a Cloud
provider bundle (Skills, output template, per-role models, tool
tiers, remote MCP servers). MCP allowlists stay empty until the
read-only review in increment 6.
- Settings CLOUD_PROVIDER and MCP_ENABLE_PREVIEW.
- 35 new tests; full suite 65 passed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>