| name | evaluate-candidate-implementation |
| description | Implementation phase of candidate test-task evaluation, run as a dedicated subagent, after the functional outcome is understood. Evaluate technical implementation quality — architecture, code structure, and the technical depth behind acceptance-criteria coverage — comparing against any reference repositories the story points at (for example the investors-mcp reference fork). Then score kit-usage conformance: determine which soofi-xyz kit agent(s)/skill(s) should have built a task like this by consulting arceus and README.md, consult each relevant builder agent as a read-only reviewer for a 0–100 coding score, and aggregate. Return separate implementation-quality and kit-usage scores with file-level evidence. Use after evaluate-candidate-product so implementation is judged third, never before the outcome. |
Evaluate Candidate Implementation
When to Use This Skill
Use this skill as the implementation phase, run as a dedicated subagent, only after evaluate-candidate-product has judged the functional outcome. Implementation is evaluated third: how the work was built never outranks whether it delivers the outcome.
This skill produces two scored dimensions from the model in evaluate-candidate-intent:
- Implementation quality (weight 7) — architecture, code structure, and the technical depth behind acceptance-criteria coverage.
- Kit-usage conformance (weight 5) — whether the candidate built the task using this soofi-xyz kit, and how correctly.
Step 1 — Evaluate Implementation Quality
Review the candidate repository and PR for technical quality, anchored to the intent and evidence model:
- Architecture — are the services, data model, and boundaries appropriate for the outcome?
- Code structure — module layout, separation of concerns, readability, error handling, tests.
- Acceptance-criteria depth — for criteria
evaluate-candidate-product marked Pass/Partial, is the underlying implementation sound or a thin shim?
- Reference repositories — where the story points at a reference (for example the investors-mcp reference fork), compare the candidate's integration against it: correct interfaces, respected access boundaries, no bypasses. Read the reference repo; do not assume.
Record concrete file paths and snippets. Do not credit claims that the code does not support, and do not reward polish that does not serve the outcome.
Step 2 — Determine the Expected Kit Toolchain
You cannot score kit-usage without the right answer for this task.
- Read the intent, evidence model, and weighting from
evaluate-candidate-intent.
- Consult
arceus (the router) and README.md to determine which kit agent(s)/skill(s) are the correct fit for a task like this. Treat arceus's routing as the reference for "what good looks like."
- Record the expected toolchain: primary agent, supporting agents, and the skills they load. Quote at most one sentence per source.
Step 3 — Gather Evidence of Kit Usage
Inspect the submission for signals the kit was used and its patterns followed:
- services/infrastructure choices versus what the expected agents prescribe;
- file/module layout and conventions matching the kit's skills;
- adherence to the
apply-engineering-guidelines baseline;
- commits/docs that reference kit agents/skills (helpful, not required — judge the code).
Do not credit a kit mention not reflected in the code, and do not penalize a strong, conformant implementation merely because it does not name the agents.
Step 4 — Consult Each Relevant Builder Agent for a Coding Score
For each expected agent from Step 2, consult that agent (spawn it as a subagent) as a read-only domain reviewer. Ask each to assess, from its specialty:
- are the services/infrastructure the ones it would have chosen?
- does the implementation follow its required patterns, contracts, and guardrails?
- what is correct, what is wrong or risky, and what is missing?
Request from each: a score 0–100, top strengths and deviations with file references, and a one-line confidence note. Consultations are read-only — reviewers score, they do not modify the candidate repo.
Step 5 — Aggregate the Scores
- Implementation quality (0–100) — your own technical judgment from Step 1, weighted toward what serves the outcome.
- Kit-usage conformance (0–100) — weight each consulted agent by how central it is (primary dominates). If the candidate built it well but without the kit, report a low kit-usage score and say so plainly; functional quality is captured separately by
evaluate-candidate-product.
- If expected agents cannot be consulted, mark those portions
Blocked with a reason rather than inventing a score.
Output Format
# Candidate Implementation — Quality & Kit Usage
## Implementation Quality
- Score (0–100): <n>
- Architecture / code structure / AC depth / reference-repo comparison — bullets with file paths
## Expected Toolchain
- Primary agent: <name> — <why, one sentence>
- Supporting agents/skills: <names>
## Consulted Agent Scores
| Agent | Score (0–100) | Strengths | Deviations | Confidence |
|---|---|---|---|---|
## Kit-Usage Conformance
- Aggregated score (0–100): <n> — <how weighted; built-with-kit: yes/partial/no>
## Dimension Subtotals
Map the two 0–100 scores onto the weighted model using anchored bands (0/25/50/75/100%); `Points` = Band × Weight, rounded.
| Dimension | Weight | Band | Points | Why |
|---|---|---|---|---|
| Implementation quality | 7 | <%> | n | <one line> |
| Kit-usage conformance | 5 | <%> | n | <one line> |
## Notes / Blockers
<reference repos read; agents that could not be consulted and why>
Quality Bar
- Implementation is judged after the functional outcome, never before.
- Implementation-quality findings cite file-level evidence and reference repositories where the story points at them.
- The expected toolchain is grounded in
arceus / README.md, not memory.
- Kit-usage aggregates consulted scores, weighting the primary agent highest, and reports built-without-kit honestly.
- All consultations are read-only and do not modify the candidate repository.