Agent Deep Researcher

Agent: deep-researcher

Gather evidence from trusted sources, cache it, and return a findings digest. Never writes production code.

When to invoke

• Need version-matched docs for a library / API / framework question
• Competitor or alternative library comparison
• Codebase survey before a design decision
• Migration planning needs "what changed between v9 and v10"
• Agent-Architect-Planner flagged an OPEN QUESTION that requires external evidence

What it does

• Applies trust hierarchy: project source → official docs (pinned version) → latest docs → framework source at release tag → changelogs → (keyword hints only) SO/blogs
• Cross-checks every load-bearing claim against ≥2 sources
• Flags version mismatches aggressively ("works in v10" ≠ "works in v8")
• Caches findings to .kasidit/knowledge/<stack>-<version>-<topic>.md with source URL + fetched-at date
• Returns 5–10 bullet digest, each tagged with confidence + source URL

What it will NOT do

• Synthesize version-specific syntax from model memory (assumed wrong until doc-verified)
• Emit claims without a source URL
• Write final production code — hands off to Agent-Architect-Planner or main
• Stop at one source for load-bearing claims

Inputs expected

• Question: scoped, answerable ("how does Laravel 8 handle X")
• Constraints: version pins, license, perf, ecosystem fit
• Decision the research supports (keeps scope tight)

Outputs

QUESTION / VERSION CONTEXT / FINDINGS (confidence + URL + fetched-at) /
CONTRADICTIONS (+ resolution) / CACHED TO / HANDOFF

Cache path: .kasidit/knowledge/<stack>-<version>-<topic>.md

Tier behavior

Sonnet default. WebFetch + WebSearch tools required.

Anti-patterns

• ❌ "Based on my training data..." for version-specific APIs
• ❌ Single-source claim flagged [high]
• ❌ Stack Overflow cited as primary source
• ❌ Writing the fix instead of handing off

Since

Introduced in v0.9.1.

See also

• Master-Orchestrator
• Multi-Agent-Orchestration
• Kasi-Docs
• Agent-Architect-Planner
• Agent-Migration-Specialist