Opensearch Review Workflow

OpenSearch Review Workflow

Run the three declared OpenSearch review skills against one source snapshot. Validate their availability and complete execution independently of other installed OpenSearch skills.

Declared inventory and dependencies

Runnable skills, exactly:

• opensearch-optimization
• opensearch-query-review
• opensearch-schema-review

Optional synthesis agent: opensearch-expert.

opensearch-review-workflow is the orchestrator and is never recursively invoked as a component.

Cost-aware native routing

Resolve portable profiles from the current runtime capability catalog at dispatch. Use the three mandatory skill passes at frugal/low for a bounded static diff, promote only the affected query, schema, or optimization pass to standard/medium for cross-index or rollout reasoning, and use the optional opensearch-expert=standard/medium only when specialist evidence conflicts. Reserve frontier/high for confirmed irreversible data or public compatibility risk.

Every native handoff uses model_selection: explicit-native-override; spawn with fork_turns: none, the exact resolved model, and exact reasoning_effort. Never inherit the primary model or reasoning setting. Persist requested/observed routes in the ledger and reject mismatch. Timeout, quota, authentication, and missing cluster access do not justify a larger model.

Input contract

Accept a repository/diff scope and any supplied mappings, templates, settings, queries, pipelines, traffic/latency/recall measurements, cluster topology, version, and rollout constraints. Derive missing artifacts from the repository when safe. Label unavailable runtime metrics as gaps; do not fabricate cluster state or performance conclusions.

Workflow

1. Required dependency and execution inventory check

Resolve each declared runnable skill by its canonical frontmatter name in the host runtime's effective catalog after normal precedence rules, limited to discovery roots active for this target. Do not scan unrelated home or repository paths or infer identity from directory names.

• If the effective catalog is unavailable, stop with CATALOG_UNAVAILABLE.
• A missing declared skill stops preflight with MISSING_DEPENDENCY; a declared skill with ambiguous definitions after precedence stops with AMBIGUOUS_DEPENDENCY.
• Additional installed skills, including opensearch-code-authoring, are informational only. They neither block preflight nor become review components. Exclude the orchestrator itself from execution. A shared opensearch- prefix does not make a skill a review dependency.
• The planned mandatory skill list must contain each declared runnable ID exactly once, with no additional IDs. Missing, duplicate, or extra planned passes stop with INVENTORY_MISMATCH.

Record the declared IDs, resolved dependencies, additional discovered OpenSearch IDs, and planned mandatory list separately. Resolve the optional synthesis agent only if used; its absence or ambiguity is a synthesis gap, not a failed mandatory review dependency.

2. Freeze and route evidence

Freeze the target revision and shared artifact list. Give each skill the complete scope plus its specialist inputs:

• opensearch-query-review: query DSL, builders, search pipelines, relevance and pagination
• opensearch-schema-review: mappings, analyzers, templates, vector fields, reindex/alias plan
• opensearch-optimization: scale, shards, memory, indexing, vector parameters, p99 and recall

Absence of a feature is a specialist NOT_APPLICABLE result with evidence, not a skipped run.

3. Run and synthesize

Run all three skills concurrently, with a hard limit of three active runs. Each result must contain severity, confidence, artifact/path, evidence, impact, fix, rollout risk, and verification command or measurement.

After all runs complete, optionally invoke opensearch-expert when findings conflict across query, schema, and capacity concerns or a cross-cutting rollout decision is needed. Give it the original outputs and ask it to synthesize, not rerun or overwrite specialists.

Deduplicate only identical artifact plus root cause. Preserve conflicting advice, such as recall versus latency or mapping correctness versus migration risk, and state the trade-off. Require recall@k and p99 together for vector performance claims and a reindex/alias path for immutable mapping changes.

4. Verdict

Reconcile the mandatory run ledger against the declared IDs: require exactly one terminal result per declared skill. A missing, duplicate, or undeclared result makes coverage INCOMPLETE; retain valid findings and identify the discrepancy. Keep optional synthesis outside this mandatory list.

• INCOMPLETE: a mandatory skill failed, timed out, or returned unusable output
• BLOCK: supported CRITICAL risk, data-loss risk, or unsafe no-rollback rollout
• CHANGES_REQUIRED: supported HIGH finding without a CRITICAL finding
• WARNING: only MEDIUM/LOW findings
• APPROVE: all three completed and no supported findings remain

Never use optional opensearch-expert failure to discard specialist results. Mark the synthesis gap and retain the verdict warranted by those results.

Normalize supported severities to CRITICAL|HIGH|MEDIUM|LOW|NONE while retaining original labels. When a mandatory run failure makes the verdict INCOMPLETE but a completed specialist found a supported CRITICAL, add blocking_finding_present: true and preserve that finding.

Output contract

Return:

1. scope, OpenSearch version/assumptions, and immutable revision
2. inventory check with declared IDs, resolved dependencies, informational extras, and planned IDs
3. run ledger for all mandatory skills and optional agent
4. aggregate verdict
5. unified findings with contributors and unresolved trade-offs
6. query/mapping/settings diffs or parameter changes, without applying them
7. verification matrix covering correctness, _mapping/_analyze, rollout, recall@k, p99, memory, and relevant cluster statistics
8. coverage gaps and next actions

Failure and partial-result semantics

• Stop before component execution on unavailable or ambiguous required dependencies or a mismatched mandatory execution plan. Additional installed skills do not invalidate the review inventory.
• After execution begins, retain successful artifacts but return INCOMPLETE if any mandatory run fails. Never approve a partial suite.
• Mark live-cluster claims UNVERIFIED when access or metrics are unavailable. Static repository review may continue, but it cannot prove production performance or capacity.
• If the source revision changes, mark all results STALE and do not reconcile across revisions.