Workbench Algorithm Advisory Gate

Workbench Algorithm Advisory Gate

Use this skill when an issue declares SDD_STAGE: Algorithm Advisory Gate, asks for algorithm review before Task List, or assigns the Remote Algorithm Advisor.

Required Source Reads

1. The latest SDD_STAGE: Technical Design block.
2. docs/algorithm-advisory-gate-lane.md.
3. The scoped files named by code_scope or SCOPED_EVIDENCE.
4. The mounted data-algo skill entrypoint.

Do not read unrelated issue history or broad repo trees unless the Technical Design names that evidence.

Runtime Boundary

• Advisory-only. Do not edit repo files.
• Do not mutate global Claude Code settings, .capy/settings.json, Multica runtime config, Sanity datasets, credentials, or Workbench Max.
• Use per-lease temporary HOME and temporary Claude config when running in a VM or disposable cell.
• If the exact GPT-5.5 xhigh model setup is unavailable, return FLAG with the observed model readback.

Required Report

ALGO_ADVISORY_REPORT
target_issue:
source_stage: Technical Design
runtime:
model:
isolation:
skill_sources:
code_scope:
correctness_verdict: PASS | BLOCK
complexity_verdict: PASS | FLAG | BLOCK
data_structure_notes:
algorithmic_risks:
task_list_injections:
verification_injections:
settings_readback:
residual_risk:
VERDICT: PASS | FLAG | BLOCK

Verdicts

• PASS: no correctness blocker and complexity risk is acceptable.
• FLAG: Task List may proceed only if advisory injections become explicit tasks, checks, or residual risks.
• BLOCK: Technical Design must be revised before Task List because the advisor found a concrete correctness, data-loss, privacy, concurrency, or semantic risk.