docs(user-guide): document append_to_context for iterator context accumulation

Relates to EPMCDME-10900

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: 2 people

docs/user-guide/workflows/configuration/context-management.md

@@ -233,6 +233,7 @@ task: |
 - ⚠️ **Simple lists**: Not added to context as variables (use output_key or wrap in dict)
 - ⚠️ **Plain strings**: Not added to context as variables (use output_key or output as JSON)
 - 💡 **Best practice**: Always output structured JSON from assistants for maximum flexibility
+- 💡 **Parallel iterations**: Use `append_to_context: true` with `output_key` to collect all iteration results into a list — without it, only the last iteration's value is retained
 
 #### Accessing Context in Templates
 
@@ -337,6 +338,40 @@ states:
       include_in_llm_history: false  # Don't show raw IDs to LLM
 ```
 
+**append_to_context** (boolean, default: `false`)
+
+Controls whether the current state's output is appended to an existing list in the context store rather than overwriting it.
+
+- `true`: Output is accumulated — the value under `output_key` in the context store grows into a list across multiple executions (most useful with `iter_key` iterations)
+- `false` (default): Output overwrites the previous value under the same key
+
+**When to use `true`:**
+
+- Collecting results from all parallel `iter_key` iterations into a single list
+- Any scenario where multiple states write to the same key and all values must be preserved
+
+**Requirements and constraints:**
+
+- Use together with `output_key` to specify which context key accumulates the results
+- When `append_to_context: true`, the top-level state key (set by `output_key`) is **not** written — read the accumulated list from the context store only via `{{output_key}}`
+- Has no effect when `store_in_context: false`
+
+```yaml
+states:
+  - id: process-item
+    assistant_id: processor
+    task: "Analyze {{task}} and return a result object"
+    next:
+      state_id: aggregate
+      output_key: results
+      append_to_context: true  # Each parallel branch appends; context_store["results"] = [r1, r2, r3]
+
+  - id: aggregate
+    assistant_id: aggregator
+    task: "Summarize all results: {{results}}"
+    # {{results}} contains the full list from every iteration
+```
+
 **clear_prior_messages** (boolean, default: `false`)
 
 Clears all prior messages from the message history, creating a "fresh start" for the LLM context.
@@ -543,6 +578,18 @@ next:
 
 Use case: When iterating over items, prevent context from accumulating across iterations. Each iteration gets only the current item, not data from previous iterations.
 
+**Pattern 7: Accumulate All Iteration Results Into a List**
+
+```yaml
+next:
+  state_id: aggregate
+  iter_key: items
+  output_key: all_results
+  append_to_context: true
+```
+
+Use case: When iterating in parallel and you need to collect every branch's output into a single list. Without `append_to_context: true`, only the last branch's value is kept (last-value-wins). With it, `context_store["all_results"]` contains a list with one entry per iteration.
+
 ### 6.3 Dynamic Value Resolution
 
 Dynamic value resolution enables you to use context store values throughout your workflow configuration using template syntax.

docs/user-guide/workflows/configuration/state-transitions.md

@@ -361,6 +361,21 @@ The previous state can output various formats, and `iter_key` adapts accordingly
 - Two formats: dictionary key (`"items"`) or JSON Pointer (`"/data/items"`)
 - The extracted value must be a list or will be wrapped as single-item list
 
+**append_to_context** (boolean, default: `false`):
+
+- When `true`, each iteration's output is **appended** to a list in the context store instead of overwriting the previous value
+- When `false` (default), standard overwrite semantics apply — the last iteration's value wins on duplicate keys
+- Use together with `output_key` to control which context key accumulates the collected results
+- When `append_to_context: true` is combined with `output_key`, the top-level state key is **not** set — values are only available via the accumulated list in the context store
+
+```yaml
+next:
+  state_id: collect-results
+  iter_key: items
+  output_key: processed_items
+  append_to_context: true  # Each iteration appends its output; context_store["processed_items"] becomes a list
+```
+
 #### Multi-Stage Iteration:
 
 For multi-stage iteration (when you have multiple sequential states processing each item), the **same `iter_key` must be present in every state** included in the iteration chain.
@@ -550,6 +565,57 @@ states:
 4. After all branches complete, contexts and message histories are merged
 5. Merged results flow to `merge-results`
 
+**Example 7: Accumulating Results Across Iterations**
+
+When all iteration results must be preserved, use `append_to_context: true` so each parallel branch contributes to a shared list rather than overwriting it.
+
+```yaml
+states:
+  - id: get-tickets
+    tool_id: jira-api
+    tool_args:
+      jql: "project = PROJ AND status = 'Open'"
+    # Outputs: [{"id": "PROJ-1", "title": "Bug A"}, {"id": "PROJ-2", "title": "Bug B"}, {"id": "PROJ-3", "title": "Bug C"}]
+    next:
+      state_id: analyze-ticket
+      iter_key: .  # Iterate over the entire list
+
+  - id: analyze-ticket
+    assistant_id: analyzer
+    task: |
+      Analyze ticket {{id}}: {{title}}
+      Return a JSON object: {"ticket_id": "...", "severity": "low|medium|high", "summary": "..."}
+    # Iteration 1 returns: {"ticket_id": "PROJ-1", "severity": "high", "summary": "..."}
+    # Iteration 2 returns: {"ticket_id": "PROJ-2", "severity": "low", "summary": "..."}
+    # Iteration 3 returns: {"ticket_id": "PROJ-3", "severity": "medium", "summary": "..."}
+    next:
+      state_id: create-report
+      output_key: analyses
+      append_to_context: true  # Accumulate all results; context_store["analyses"] = [{...}, {...}, {...}]
+
+  - id: create-report
+    assistant_id: reporter
+    task: |
+      Create a severity report based on all ticket analyses.
+      Analyses: {{analyses}}
+    # Receives the full list of all three analyses in {{analyses}}
+    next:
+      state_id: end
+```
+
+**How it works:**
+
+1. Three parallel branches process one ticket each
+2. Each branch writes its output with the `analyses` key via `append_to_context: true`
+3. The reducer appends each result to the list — no overwriting occurs
+4. `create-report` receives `analyses = [result_1, result_2, result_3]` in its context
+
+:::tip
+`append_to_context: true` is the recommended way to collect results from all parallel iterations into a single list. It replaces the workaround of using unique per-iteration keys (`result_1`, `result_2`, ...).
+:::
+
+---
+
 #### Context Isolation and Merging:
 
 Iterations have important context management characteristics that ensure proper isolation and aggregation:
@@ -571,9 +637,17 @@ Iterations have important context management characteristics that ensure proper
 
 - When all parallel iterations complete (fan-in), their context stores are **merged**
 - The merge uses `add_or_replace_context_store` reducer
-- For duplicate keys across iterations, the **last value wins** (last iteration overwrites previous)
+- **Default (overwrite)**: for duplicate keys across iterations, the **last value wins** (last iteration overwrites previous)
+- **Accumulation mode**: when `append_to_context: true` is set on the iterating state, each iteration's output is **appended** to a list under the specified key — no values are lost
 - The merged context is then passed to the next state after iteration
 
+**Choosing between overwrite and accumulation:**
+
+| Mode                | Config                     | Result for key `output` after 3 iterations      |
+| ------------------- | -------------------------- | ----------------------------------------------- |
+| Overwrite (default) | `append_to_context: false` | `output = "result_3"` (only last)               |
+| Accumulation        | `append_to_context: true`  | `output = ["result_1", "result_2", "result_3"]` |
+
 **Message History Merging:**
 
 - Similarly, message histories from all iterations are also merged
@@ -611,9 +685,9 @@ states:
 **Important Context Merging Notes:**
 
 - Iterations are isolated during execution but merged after completion
-- Context keys set by multiple iterations will have only one final value (last wins)
-- To preserve all iteration results, use unique keys (e.g., `result_1`, `result_2`) or aggregate into lists
-- Message histories are fully preserved from all iterations
+- By default, context keys set by multiple iterations will have only one final value (last wins)
+- To preserve all iteration results, set `append_to_context: true` combined with `output_key` — the context key will accumulate all values as a list
+- Message histories are fully preserved from all iterations regardless of the merge mode
 
 #### Important Notes:
 
@@ -626,5 +700,7 @@ states:
 - Cannot combine `iter_key` with `state_ids` (parallel transitions) or `condition`/`switch`
 - Each iteration branch has isolated context and message history during execution
 - After all iterations complete, contexts and message histories are merged using LangGraph reducers
+- Use `append_to_context: true` to accumulate all iteration outputs into a list; without it, only the last iteration's value is retained for duplicate keys
+- `append_to_context: true` has no effect when `store_in_context: false`
 
 ---

faq/how-do-i-collect-all-results-from-parallel-workflow-iterations.md

@@ -0,0 +1,31 @@
+# How do I collect all results from parallel workflow iterations?
+
+Use `append_to_context: true` together with `output_key` in the `next:` block of the
+iterating state. This tells the workflow engine to accumulate each iteration's output into
+a list rather than overwriting the previous value.
+
+```yaml
+states:
+  - id: process-item
+    assistant_id: processor
+    task: "Analyze {{task}} and return a result object"
+    next:
+      state_id: aggregate
+      iter_key: items
+      output_key: all_results
+      append_to_context: true
+
+  - id: aggregate
+    assistant_id: aggregator
+    task: "Summarize all results: {{all_results}}"
+    # {{all_results}} contains every iteration's output as a list
+```
+
+By default (`append_to_context: false`), when multiple parallel branches write to the
+same context key, only the last branch's value is kept. Setting `append_to_context: true`
+preserves all values in `context_store["all_results"]` as an ordered list.
+
+## Sources
+
+- [State Transitions — Iterative Transitions](https://docs.codemie.ai/user-guide/workflows/configuration/state-transitions)
+- [Context Management — Context Control Flags](https://docs.codemie.ai/user-guide/workflows/configuration/context-management)

faq/what-happens-to-context-keys-when-multiple-iterations-write-to-the-same-key.md

@@ -0,0 +1,21 @@
+# What happens to context keys when multiple iterations write to the same key?
+
+By default, when parallel `iter_key` iterations all write to the same context key, the
+**last iteration's value wins** — earlier values are silently overwritten during the
+fan-in merge.
+
+To preserve every iteration's output, set `append_to_context: true` on the iterating
+state. With this flag, each branch appends its output to a list instead of overwriting:
+
+| Mode                | Config                     | Result after 3 iterations writing to `output`   |
+| ------------------- | -------------------------- | ----------------------------------------------- |
+| Overwrite (default) | `append_to_context: false` | `output = "result_3"` (only last)               |
+| Accumulation        | `append_to_context: true`  | `output = ["result_1", "result_2", "result_3"]` |
+
+Use `append_to_context: true` together with `output_key` to name the accumulation key.
+The accumulated list is accessible via `{{output_key}}` in subsequent states.
+
+## Sources
+
+- [State Transitions — Context Isolation and Merging](https://docs.codemie.ai/user-guide/workflows/configuration/state-transitions)
+- [Context Management — Context Control Flags](https://docs.codemie.ai/user-guide/workflows/configuration/context-management)