tsa-landing

name: tsa-landing version: 2.0.0 description: | Land in a new (or familiar) codebase using the tree-sitter-analyzer MCP server. One workflow → 6 decision surfaces (project_card / entry_points / recent_signals / health / top_files / agent_next_step) → ≤2k tokens, ≤3 MCP calls.

Use when:

• First time entering an unfamiliar repository
• Returning to a repo after >1 week
• User asks "what is this project?" / "where do I start?"

Workflow: parallel-fan-out 3-4 MCP tools, fold output, return decision_surface. Replaces the typical 15k-token bootstrap (read README + ls -R + git log + AGENTS). allowed-tools: - mcp__tree-sitter-analyzer__project - mcp__tree-sitter-analyzer__health - mcp__tree-sitter-analyzer__edit - Bash - Read

tsa-landing — Land in any repo in <3s

First action on entering a new repo. Replaces 6 separate bootstrap calls (15k tokens) with 3-4 parallel calls (2k tokens). 92% token saved.

When to use

• You just entered an unfamiliar project (no Claude.md / no skim yet)
• You return after long gap and need "what's the state?"
• User asks any of: "what is this project / where do I start / what's the entry / recent changes / health"

Don't use when:

• You already have full context (just continue working)
• User wants to read source — use structure action=read directly

Procedure

Step 1 — Verify tools available

uv run python -m tree_sitter_analyzer --check-tools --format json | head -5

If fd or rg missing, stop and tell user how to install.

Step 2 — Fan-out 4 MCP calls (in single message, parallel)

Call these 4 tools in ONE message (parallel tool use):

1. project action=overview (no args) — project_card + entry_points
2. health action=project with max_files: 5 — grade distribution + weakest dimension
3. edit action=impact with mode: "branch" — recent_signals (last commit, ahead-of-main)
4. project action=workflow (no args) — current_phase + recommended_commands

Step 3 — Fold and emit decision_surface

Combine into single Decision Surface:

{
  "project_card": {
    "name": <from project action=overview project_root basename>,
    "purpose": <from project action=overview summary.purpose if present>,
    "primary_language": <from project action=overview summary.by_language[0].name>,
    "language_mix": <from project action=overview summary.by_language top 3>,
    "size": {
      "files": <from project action=overview summary.total_files>,
      "loc": <from project action=overview summary.total_lines>
    }
  },
  "entry_points": <from project action=overview entry_points>,
  "recent_signals": {
    "last_commit": <git log -1 --oneline via Bash>,
    "ahead_of_origin": <git rev-list --count via Bash>,
    "uncommitted_files": <git status --short | wc -l via Bash>,
    "branch": <git branch --show-current via Bash>
  },
  "health": {
    "verdict": <from health action=project verdict>,
    "risk": <from health action=project agent_summary.risk>,
    "grade_distribution": <from health action=project grade_distribution>,
    "weakest_dimension": <from health action=project weakest_dimension>
  },
  "top_files_to_know": [
    "AGENTS.md",
    "CLAUDE.md",
    <from project action=overview entry_points>,
    <top 3 from health action=project top_refactoring_targets>
  ],
  "agent_next_step": {
    "if_asked_what_is_this":
      "Read AGENTS.md (canonical contracts) + docs/CODEMAPS/architecture.md (topology). Stop after 2k tokens.",
    "if_asked_to_add_feature":
      "Call project action=workflow → follow phase_order. Use TDD (write test first).",
    "if_asked_to_fix_bug":
      "Call health action=patterns file_path=<file> → edit action=refactor file_path=<file>. Cross-ref nav action=lineage if symbol-level.",
    "if_asked_about_test_status":
      "Run: uv run pytest -q (5-min cap). Project enforces xdist parallel, ~5min for 15k tests."
  },
  "summary_line": "<project> files=<N> py=<X%> grade=<G> recent=<commit_subj>",
  "verdict": "INFO"
}

Step 4 — Stop after landing

Do NOT proceed to action until user gives next instruction. The landing is the deliverable.

Token accounting

Why this is a Skill, not an MCP tool

• MCP tool definition costs ~400 tokens always-loaded
• Skill description costs ~30 tokens (loads on trigger)
• 13-15× cheaper — frees context budget for actual work
• Workflow logic (parallel fan-out + folding) doesn't need MCP semantics

See also

• ~/.claude/memory/tsa_research_gold.md — Why this design (3 金矿洞察)
• ~/.claude/memory/tsa_playbook_24x7.md — full project state map
• docs/internal/AGENT_LANDING_KIT_DESIGN.md — earlier MCP-tool variant of this (deprecated in favor of Skill form)

Anti-bias note (per Anthropic verification-specialist pattern)

When in doubt about the health verdict, err toward higher-severity:

• INFO vs REVIEW — pick REVIEW
• REVIEW vs CAUTION — pick CAUTION

False positives at landing time are recoverable (user says "false alarm"); false negatives ship bugs into agent workflow.