Lens - Long-Context Analysis

Deeply analyze a large document, codebase, or transcript that's too big to reason about by skimming. Uses ACTIVE RETRIEVAL (map → query → synthesize) over Kimi K3's 1M window instead of dumping everything in at once — which avoids the "lost in the middle" failure mode of naive long-context prompting.

Usage

/lens [target + what you want to know]
/lens [files...] [question]

When to Use

• A document/codebase/log is too large to hold in your head
• You need specific insights, not a generic summary
• You suspect inconsistencies or hidden relationships across sections
• Naive "read it all" would blow context or miss the middle

/lens vs alternatives

Instructions

When user invokes /lens [target]:

Step 1: Map the structure first (don't dump)

Identify the target (file paths or pasted content). Get a structural map BEFORE deep analysis:

mcp__tachibot-mcp__kimi_long_context({
  content: "[paste content here, OR a one-line note if using files below]",
  files: ["path/to/file1", "path/to/dir/**"],
  task: "summarize",
  outputFormat: "structured",
  query: "Map the structure: list the major sections/modules/themes and what each covers. Do NOT analyze yet — just produce the table of contents."
})

Present the map to the user: "Found [N] sections: ... Diving into the relevant ones."

Step 2: Targeted analysis (active retrieval)

For the user's actual question, query specific aspects rather than re-reading everything:

mcp__tachibot-mcp__kimi_long_context({
  content: "[content]",
  files: ["..."],
  task: "analyze",
  outputFormat: "detailed",
  query: "[the user's specific question, scoped to the relevant sections found in Step 1]"
})

Use task: "extract" to pull specific facts, "find" to locate something, "compare" to contrast sections.

Step 3: Cross-check for inconsistencies (if the content makes claims)

If the content argues or claims things, get an independent second read on the key parts:

mcp__tachibot-mcp__gemini_analyze_text({
  text: "[the key claims/sections surfaced in Step 2]",
  type: "key-points"
})

Step 4: Synthesize

mcp__tachibot-mcp__nextThought({
  thought: "Synthesize the long-context analysis of [target].\n\nFindings:\n[Step 2 + Step 3 results]\n\nProduce:\n1. Direct answer to the user's question\n2. Key insights (ranked)\n3. Relations/dependencies across sections\n4. Inconsistencies or gaps found\n5. What to look at next",
  model: "gemini",
  executeModel: true,
  contextWindow: "all",
  nextThoughtNeeded: false
})

If kimi_long_context is unavailable, fall back to gemini_analyze_text (1M context) using the files parameter for the analysis steps.

Step 5: Present Results

## Lens: [target]

### Answer
[Direct answer to the question]

### Key Insights
1. [ranked insight]
...

### Relations / Dependencies
[How sections connect — what depends on what]

### Inconsistencies & Gaps
[Contradictions or missing pieces found]

### Look At Next
[Where to dig further]

Examples

• /lens src/**/*.ts how does auth flow through this codebase?
• /lens [paste 80-page spec] what contradictions exist between sections 3 and 7?
• /lens server.log find the root-cause chain for the 02:14 outage
• /lens compare these two RFCs and tell me where they disagree