assess

Purpose

Performs a comprehensive technical assessment of a codebase using 9 specialist agents that each deeply evaluate one dimension of code health, followed by a CTO orchestrator that synthesizes findings into a prioritized report with opinionated recommendations.

Correctness gate: every finding is a claim about the codebase. Per the /verify-claims protocol, each specialist finding carries a grounded confidence % and its basis (verified — you observed it in the code — vs inferred), never asserts a finding it has not verified, and marks a low-confidence judgment as tentative rather than stating it as fact. The CTO orchestrator must not raise a finding's confidence beyond what its evidence supports when prioritizing.

Provider-native delegation

Dispatch each specialist through the selected host's available subagent facility. Preserve the specialist persona, fresh assessment context, report output, and the assessment gates regardless of host. Claude delegation: Claude uses the Agent tool; the Claude model names below apply only to that facility. A Codex-selected run uses its available subagent facility and configured Codex provider policy, without translating Claude model names.

Invocation:

• Onboarding: Runs as part of /conduct after bootstrap for existing projects
• On-demand: User invokes /assess anytime for a health check
• Selective: /assess --area security runs only the security specialist

Practices

1. Determine Scope

Check invocation arguments:

Argument	Behavior
(none)	Full assessment — all 9 specialists + CTO synthesis
--area <name>	Single specialist only (security, data-integrity, dependencies, architecture, duplication, testing, infrastructure, observability, devex)

2. Prepare Assessment Directory

Create .pipeline/assessment/ if it doesn't exist. Clear any stale reports from previous runs.

3. Gather Shared Context

Before dispatching specialists, gather context they all need:

1. Codebase file listing — find . -type f filtered to source/config files (exclude vendor, node_modules, .git, tmp, log)
2. Tech-context — if loaded in session from /bootstrap, include it
3. Existing decisions — list files in .docs/decisions/ so architecture specialist can cross-reference

4. Dispatch Specialists (3 Batches of 3)

Dispatch specialists in parallel within each batch, sequential between batches. Each specialist is a subagent using the corresponding agent persona.

Batch 1 (parallel):

Specialist	Agent File	Claude model	Output
Security Auditor	agents/cto-security.md	opus	.pipeline/assessment/cto-security.md
Data Integrity Reviewer	agents/cto-data-integrity.md	opus	.pipeline/assessment/cto-data-integrity.md
Dependency Auditor	agents/cto-dependencies.md	sonnet	.pipeline/assessment/cto-dependencies.md

Rate limit cooldown: sleep 30 seconds before dispatching Batch 2.

Batch 2 (parallel, after Batch 1 completes):

Specialist	Agent File	Claude model	Output
Architecture Coherence	agents/cto-architecture.md	opus	.pipeline/assessment/cto-architecture.md
Code Duplication	agents/cto-duplication.md	sonnet	.pipeline/assessment/cto-duplication.md
Test Strategy	agents/cto-testing.md	sonnet	.pipeline/assessment/cto-testing.md

Rate limit cooldown: sleep 30 seconds before dispatching Batch 3.

Batch 3 (parallel, after Batch 2 completes):

Specialist	Agent File	Claude model	Output
Infrastructure	agents/cto-infrastructure.md	sonnet	.pipeline/assessment/cto-infrastructure.md
Observability	agents/cto-observability.md	sonnet	.pipeline/assessment/cto-observability.md
Developer Experience	agents/cto-devex.md	sonnet	.pipeline/assessment/cto-devex.md

Subagent prompt template:

You are the [Specialist Name]. Read your persona from agents/[agent-file].

Assess the codebase in [project directory]. Here is the file listing:
[file listing]

[Tech-context if available]

Write your findings to .pipeline/assessment/[agent-name].md following the output format
defined in your persona file. Be thorough — this is a deep assessment, not a surface scan.

For --area mode: Dispatch only the named specialist. Skip Steps 5-6 (no CTO synthesis for single-area assessments). Present the specialist report directly.

4b. Structure Verification

After specialist reports complete, verify that the actual codebase structure matches documented architectural decisions. This catches assumption drift — where code evolves away from what was decided, or decisions were never documented.

Checks:

1. ADR compliance — For each ADR in .docs/decisions/, verify the decision is reflected in code. Flag violations where code contradicts a documented decision.
2. Undocumented decisions — Flag architectural patterns in code that have no corresponding ADR (e.g., a new external integration, a new persistence strategy, a service decomposition that was never decided on paper).
3. Stale ADRs — Flag ADRs whose referenced code no longer exists or has been substantially rewritten without updating the decision record.

This feeds into the CTO orchestrator synthesis as a "structure verification" section.

5. Dispatch CTO Orchestrator

Rate limit cooldown: sleep 30 seconds before dispatching the CTO orchestrator.

After all 9 specialist reports are written, dispatch the CTO orchestrator:

You are the CTO Orchestrator. Read your persona from agents/cto-orchestrator.md.

Here are the 9 specialist reports:
[inline all 9 reports from .pipeline/assessment/]

Use the template at [absolute path of this skill's templates/technical-assessment.md.template].
Project name: [project name]
Date: [YYYY-MM-DD]

Write the final assessment report to .docs/decisions/technical-assessment-[YYYY-MM-DD].md.

Claude model: opus (cross-referencing 9 reports requires deep reasoning).

6. Present Findings

After the CTO orchestrator writes the final report:

1. Present the report to the user via review_artifacts
2. Highlight the verdict (HEALTHY / NEEDS_WORK / CRITICAL)
3. Highlight critical findings count and quick wins count
4. If CRITICAL: recommend running /assess --area <name> for deeper dives on problem areas

7. Record Assessment

Save assessment metadata to .pipeline/conduct-state.json:

{
  "assess": "done",
  "assess_date": "YYYY-MM-DD",
  "assess_verdict": "HEALTHY|NEEDS_WORK|CRITICAL"
}

Verdict Definitions

Verdict	Criteria	Implication
HEALTHY	No critical findings. Systemic patterns are minor.	Proceed to feature work.
NEEDS_WORK	No critical findings, but systemic patterns need addressing.	Feature work can proceed with caution. Assessment feeds /explore.
CRITICAL	Critical findings exist OR 3+ systemic patterns at important+ severity.	Address findings before feature work. Assessment findings become the backlog.

Verification

• [ ] .pipeline/assessment/ directory created
• [ ] All 9 specialist reports written (or single report for --area mode)
• [ ] CTO orchestrator report written to .docs/decisions/technical-assessment-YYYY-MM-DD.md
• [ ] Report follows template structure from templates/technical-assessment.md.template (relative to this skill's own directory)
• [ ] Verdict is one of: HEALTHY, NEEDS_WORK, CRITICAL
• [ ] Assessment metadata saved to .pipeline/conduct-state.json
• [ ] Report presented to user for review