Tars Plan

Use the TARS cognitive compiler bridge to execute a goal.

TARS compiles goals into structured execution plans (Workflow-of-Thought DAGs). You provide the reasoning intelligence. TARS handles pattern selection, tool dispatch, validation, and learning from outcomes.

## Workflow

### Step 1 — Compile the plan

Call the `tars_compile_plan` MCP tool:

```json
{"goal": "$ARGUMENTS"}

TARS analyzes the goal, selects the best reasoning pattern (Chain-of-Thought, ReAct, Graph-of-Thought, Tree-of-Thought) based on goal characteristics and past success history, and returns a plan manifest.

Step 2 — Read the manifest

The manifest contains:

• plan_id — unique identifier for this execution
• pattern — the reasoning pattern TARS selected
• nodes — a map of node objects, each with an id, type, and prompt or tool_spec
• entryNode — the node to start from
• edges — maps each node id to its next node(s)

Node types:

• Reason — requires your reasoning (analysis, synthesis, planning)
• Tool — requires calling a TARS tool or external action
• Validate — requires checking an output against acceptance criteria
• Memory — requires querying or storing knowledge

Step 3 — Walk the DAG

Start at entryNode. For each node, based on its type:

Reason nodes: Read the prompt field. Use your own reasoning to produce an answer. This is where your intelligence drives the plan.

Tool nodes: Call tars_execute_step:

{"plan_id": "<id>", "node_id": "<node_id>", "input": "<your input or prior node output>"}

Validate nodes: Call tars_validate_step:

{"plan_id": "<id>", "node_id": "<node_id>", "content": "<the output to validate>"}

If validation fails, re-examine the upstream node output and retry.

Memory nodes: Call tars_memory_op:

{"plan_id": "<id>", "node_id": "<node_id>", "op": "query|store", "key": "<key>", "value": "<value if storing>"}

Step 4 — Follow edges

After completing a node, look up its id in edges to find the next node(s). If multiple next nodes are listed, execute them in order (they may be independent branches). Continue until no further edges remain.

Step 5 — Complete the plan

When all nodes are done, call tars_complete_plan:

{"plan_id": "<id>", "output": "<final synthesized result>"}

Example

Goal: "Find unused functions in src/Tars.Core"

1. Compile: Call tars_compile_plan with {"goal": "Find unused functions in src/Tars.Core"}.

Returns:

{
  "plan_id": "p-a1b2c3",
  "pattern": "ReAct",
  "entryNode": "n1",
  "nodes": {
    "n1": {"id": "n1", "type": "Tool", "tool_spec": {"name": "codebase_search", "args": {"scope": "src/Tars.Core", "query": "public function definitions"}}},
    "n2": {"id": "n2", "type": "Reason", "prompt": "Given the function list and call graph, identify functions with zero callers outside their own module."},
    "n3": {"id": "n3", "type": "Validate", "prompt": "Confirm each candidate is truly unused — check for reflection, serialization, or interface implementations."}
  },
  "edges": {"n1": ["n2"], "n2": ["n3"]}
}

2. Execute n1 (Tool): Call tars_execute_step with {"plan_id": "p-a1b2c3", "node_id": "n1", "input": ""}. TARS returns a list of function definitions and their call sites.

3. Execute n2 (Reason): Read the prompt. Analyze the tool output. Produce a list of candidate unused functions with reasoning for each.

4. Execute n3 (Validate): Call tars_validate_step with {"plan_id": "p-a1b2c3", "node_id": "n3", "content": "<your candidate list>"}. TARS checks against golden traces and acceptance criteria. Returns pass/fail per candidate.

5. Complete: Call tars_complete_plan with the validated final list.

Automatic behaviors

You do not need to manage these — TARS handles them behind the scenes:

• Pattern selection: TARS picks the optimal reasoning pattern based on goal analysis and historical success rates. The manifest reflects this choice.
• Regression detection: Validation nodes compare results against golden traces from prior successful runs. Regressions are flagged automatically.
• Learning: Plan outcomes (success, failure, timing) feed back into pattern selection weights so future plans improve.