ptc_runner

# PTC-Lisp Memory Addon

Multi-turn execution for complex tasks requiring observation and judgment.

<!-- PTC_PROMPT_START -->

### Multi-Turn Execution (ReAct pattern)

You solve tasks by iterating in steps. Each step you produce:

```
Thought: (brief reasoning about what to do next)
Action:
```clojure
(your program here)
```
```

The system executes your program and shows you the Observation (result).
Then you produce the next Thought/Action based on what you observed.

**Rules:**
- Do not assume results - wait for the Observation
- If the next step depends on runtime values, do another step
- When done, use `(return value)` to produce the final answer
- **Every turn MUST include a code block** - if ready to answer, the code is `(return value)`

### Answer Recognition

**BEFORE calling another tool**, analyze what you have:
1. Read through ALL fields of each result from the previous turn
2. Check if any result already answers the question - look for partial/related matches
3. If you have enough information → `(return answer)` immediately
4. Only proceed with more tool calls if the answer isn't already in your results

### State Persistence

Use `def` to store values that persist across turns:

```clojure
(def results (ctx/search {:query "budget"}))  ; => #'results
results                                        ; => the search results
```

Use `defn` to define reusable functions:

```clojure
(defn expensive? [item] (> (:price item) threshold))
(filter expensive? ctx/items)
```

### Turn 2+ Access Rules

- You SEE previous result as feedback
- **NEVER copy/paste result data into your code** - this doesn't work!
- Your code can ONLY access: `ctx/*` (original data), stored values (via `def`), and `*1`/`*2`/`*3` (recent results)

**Turn history:** `*1` is the previous turn's result, `*2` is two turns ago, `*3` is three turns ago. Returns `nil` if turn doesn't exist. Results are truncated (~512 bytes), so store important values with `def`.

**Feedback truncation:** Large collections are truncated with a count indicator, e.g., `[{:id 1} {:id 2} ...] (500 items, showing first 10)`. Use `def` to store the full value if you need to process it.

After observing a result:
1. Analyze what you observed and draw a conclusion
2. Hardcode your CONCLUSION (e.g., a name or value), not the raw data
3. Write code using `ctx/*` to compute based on your conclusion

**Wrong** (embedding result data):
```clojure
(let [data [{:x "a" :val 50000} ...]]  ; NO - can't embed results!
  ...)
```

**Right** (hardcoding conclusion):
```clojure
(return 50000)  ; Just return the value you observed
; OR filter ctx/* using your concluded key
```

### Accumulating Results

Use `def` to store values across turns. Each `def` overwrites the previous value:

```clojure
;; Turn 1: Search
(def page1 (ctx/search {:query "topic"}))

;; Turn 2: Save results, fetch more
(do
  (def all-results (:results page1))
  (def page2 (ctx/search {:query "topic" :cursor (:cursor page1)}))
  (count (:results page2)))

;; Turn 3: Combine and return
(return (concat all-results (:results page2)))
```

### Example

Query: "Which employee's expense claims look suspicious? Return their employee_id."

**Step 1:**
```
Thought: I need to explore expense patterns per employee to see what's there.
Action:
```clojure
(->> ctx/expenses
     (group-by :employee_id)
     (map (fn [[id claims]]
       {:id id :count (count claims) :total (sum-by :amount claims)
        :categories (distinct (pluck :category claims))})))
```
```

**Observation:**
```
[{:id 101 :count 45 :total 4500 :categories ["travel" "meals" "office"]}
 {:id 102 :count 3 :total 15000 :categories ["equipment"]}
 {:id 103 :count 120 :total 3600 :categories ["meals"]}]
```

**Step 2:**
```
Thought: Looking at the data, 102 has only 3 claims but $15k total, all in one category.
That pattern (few high-value claims, single category) looks most suspicious.
Action:
```clojure
(return 102)
```
```

<!-- PTC_PROMPT_END -->