ReWOO

Sequential workers

Implement the ReWOO pattern in this codebase, using the LLM client and language we already use.

## Goal

Answer a question with two model calls on the happy path: a planner that writes every tool call upfront with #E placeholders, and a solver that answers from the collected evidence.

## How it works

1. Call the planner model once with the Planner prompt, passing the question and the tool descriptions.
2. Parse the reply into an ordered list of steps, each with a reason, a variable (#E1, #E2...), a tool name and an input string.
3. Validate the plan before running anything: every tool name exists, every variable is unique, and every #E an input mentions belongs to an earlier step.
4. Run the steps in order with plain code and no model calls: substitute the stored evidence for each #E in the input, call the tool, and store the result under its variable.
5. Call the solver model once with the Solver prompt, passing the question and the plan with every step's evidence filled in, and return its answer.

## Prompts

### Planner prompt

```markdown
For the question below, make a plan of tool calls that will gather
everything needed. Write the whole plan now. You will not see any
tool output before the plan is executed.

Tools:
{{tool_descriptions}}

Format, one pair per step:
Plan: a short reason for this step
#E1 = Tool[input]

A later step may use the output of an earlier one by writing its
variable (#E1, #E2, ...) inside its input.

Question: {{question}}
```

### Solver prompt

```markdown
Solve the question using only the plan and the evidence below. If the
evidence is missing or contradictory, say so instead of guessing.

Question: {{question}}

Plan and evidence:
{{plan_with_evidence}}

Give the final answer with one sentence of justification.
```

## Stop rule

Stop after the solver's single call; the planner is never called again (an LLM used as a worker tool counts as a tool, not a planner call). If the plan fails validation twice, stop before any tool runs and return the validation errors with a failed status.

## Requirements

- Cap the plan length (default 8 steps) and give each tool call a timeout (default 30 seconds).
- If the plan cannot be parsed or fails validation, retry the planner once with the errors appended; do not run a partly valid plan.
- When a tool call errors or times out, store the error text as that variable's evidence and keep going, so the solver can say what it could not determine.
- Log the plan, each substituted tool input and each piece of evidence, so a wrong answer can be traced to the step that caused it.
- Add a test with a stub LLM that plans 3 steps and fake tools; assert exactly two model calls (one planner, one solver), and that the third tool received the first two results in place of #E1 and #E2.

Parallel evidence

Implement the ReWOO pattern (Parallel evidence variant) in this codebase, using the LLM client and language we already use.

## Goal

Answer a question with one planner call and one solver call, running the planned tool calls in dependency waves so that independent calls overlap.

## How it works

1. Call the planner model once with the Planner prompt, passing the question and the tool descriptions, and parse the steps (variable, tool, input) as in the sequential version.
2. For each step, record the set of #E variables its input mentions; reject the plan if any refers to a later or missing step.
3. Start every step whose referenced variables all have evidence, concurrently, with each input's #E placeholders substituted.
4. As each call finishes, store its result under its variable and start any step that has just become ready; repeat until no step is waiting.
5. Call the solver model once with the Solver prompt, passing the question and the plan with all evidence filled in, and return its answer.

## Prompts

### Planner prompt

```markdown
For the question below, make a plan of tool calls that will gather
everything needed. Write the whole plan now. You will not see any
tool output before the plan is executed.

Tools:
{{tool_descriptions}}

Format, one pair per step:
Plan: a short reason for this step
#E1 = Tool[input]

A later step may use the output of an earlier one by writing its
variable (#E1, #E2, ...) inside its input.

Question: {{question}}
```

### Solver prompt

```markdown
Solve the question using only the plan and the evidence below. If the
evidence is missing or contradictory, say so instead of guessing.

Question: {{question}}

Plan and evidence:
{{plan_with_evidence}}

Give the final answer with one sentence of justification.
```

## Stop rule

Stop after the solver's single call. If a step can never start because its input depends on a failed step, mark it skipped rather than waiting forever.

## Requirements

- Cap the concurrent tool calls (default 4) and the plan length (default 8), and give each call a timeout (default 30 seconds).
- If a tool call errors, store the error text as its evidence and mark every step that depends on it as skipped, with the reason, in the evidence.
- If the plan cannot be parsed or has a dependency cycle, retry the planner once with the error appended, then stop before any tool runs.
- Record each call's start and end time, so the overlap is visible in logs.
- Add a test with a stub planner and two fake 100 ms search tools; assert both searches started before either finished and the calculator started only after both.