Skip to content

Agent Zero AI framework

Make chat compaction resumable and secret-safe

Alessandro425cfc283b3c
diff --git a/plugins/_chat_compaction/AGENTS.md b/plugins/_chat_compaction/AGENTS.md
index 7f318836c..739ed901a 100644
--- a/plugins/_chat_compaction/AGENTS.md
+++ b/plugins/_chat_compaction/AGENTS.md
@@ -18,7 +18,10 @@
- Preserve chat history integrity and persistence after compaction.
- Backup JSON and transcript artifacts must remain UTF-8 writable when chat content contains malformed Unicode such as lone surrogates.
- Keep generated summaries bounded by configured model and token limits.
+- Compacted summaries must be resumable task state: preserve the latest request, authorization boundaries, decisions, evidence, modified artifacts, pending jobs and their IDs, the next executable step, blockers, and checks not run.
- Preserve loaded skill names from `skill_instructions` metadata without copying full skill bodies into compacted summaries.
+- Preserve only secret references such as names, aliases, purposes, or storage locations; never preserve secret values.
+- Clear the cached context window after replacing history so stale transcript content is not persisted as active resumable state; the cache rebuilds on the next model turn.
- After replacing local history, clear the active Responses provider continuation while preserving stored response IDs for cleanup.
- Do not discard original context data unless the compaction flow explicitly owns that behavior.
diff --git a/plugins/_chat_compaction/helpers/compactor.py b/plugins/_chat_compaction/helpers/compactor.py
index f0a56c225..d8e8b1a82 100644
--- a/plugins/_chat_compaction/helpers/compactor.py
+++ b/plugins/_chat_compaction/helpers/compactor.py
@@ -146,6 +146,7 @@ async def run_compaction(
agent.history = History(agent=agent)
agent.history.add_message(ai=True, content=compacted_content)
clear_responses_provider_state(agent)
+ agent.data.pop(Agent.DATA_NAME_CTX_WINDOW, None)
# Clear subordinate chain
agent.data.pop(Agent.DATA_NAME_SUBORDINATE, None)
diff --git a/plugins/_chat_compaction/prompts/compact.sys.md b/plugins/_chat_compaction/prompts/compact.sys.md
index 49a2593be..62db6575c 100644
--- a/plugins/_chat_compaction/prompts/compact.sys.md
+++ b/plugins/_chat_compaction/prompts/compact.sys.md
@@ -1,18 +1,34 @@
-You are a conversation compactor. Produce the most concise summary possible while preserving critical information.
+You are a conversation compactor. Preserve the minimum state another agent needs to resume the task correctly without rereading the original conversation.
Rules:
-- Extract only: key decisions, final outcomes, actionable facts, unresolved items
-- Discard: intermediate reasoning, failed attempts, redundant exchanges, pleasantries
-- Use terse bullet points, not prose
-- Collapse related items into single lines
-- Keep exact values: file paths, config values, code identifiers, credentials, URLs
-- Preserve loaded skill names from skill_instructions metadata, but do not copy full skill bodies
-- Omit anything that can be re-derived from context
-- Group by topic, not chronology
-- No meta-commentary about the summarization
-- Target: 10-20% of original length
-
-Output format:
-- Markdown with short section headers
-- Bullet lists, no paragraphs
-- Code/paths in backticks inline, not fenced blocks unless multi-line
+- Capture the latest active user request, including anything that superseded an earlier request.
+- Preserve only explicit or clearly implied authorization and prohibitions; never broaden scope.
+- Separate decisions from assumptions and mark unverified assumptions.
+- Record completed work only with available evidence such as results, commands, tests, or artifact paths.
+- Preserve exact file paths, URLs, config values, code identifiers, job IDs, context IDs, and verification status.
+- Record pending workers or jobs with their IDs, state, and the next executable step.
+- Record blockers and every relevant check that was not run.
+- Preserve loaded skill names from `skill_instructions` metadata, but never copy skill bodies.
+- Never include passwords, API keys, tokens, credentials, private keys, session secrets, or other secret values. Preserve only a secret's name, purpose, storage location, or reference alias when needed.
+- Discard intermediate reasoning, redundant exchanges, pleasantries, and facts that can be safely re-derived.
+- Use terse bullets, no prose or meta-commentary. Target 10-20% of the original length when possible.
+
+Use exactly these sections in this order. Include `- None recorded.` for an empty section.
+
+## Current objective and latest user request
+
+## Authorized scope and prohibited actions
+
+## Decisions and assumptions
+
+## Completed work with evidence
+
+## Modified files and artifacts
+
+## Pending jobs and next executable step
+
+## Blockers and checks not run
+
+## Loaded skill names
+
+## Secret references
diff --git a/tests/test_chat_compaction.py b/tests/test_chat_compaction.py
index f9678169f..9bd678cd3 100644
--- a/tests/test_chat_compaction.py
+++ b/tests/test_chat_compaction.py
@@ -72,6 +72,10 @@ class _CompactionAgent:
def __init__(self):
self.history = _CompactionHistory()
self.data = {
+ "ctx_window": {
+ "text": "pre-compaction transcript with secret values",
+ "tokens": 42,
+ },
"responses_state": {
"response_id": "resp_current",
"previous_response_id": "resp_previous",
@@ -86,6 +90,31 @@ class _CompactionAgent:
self.data[key] = value
+def test_compaction_prompt_is_resumable_task_state_without_secret_values():
+ prompt = (
+ PROJECT_ROOT / "plugins" / "_chat_compaction" / "prompts" / "compact.sys.md"
+ ).read_text(encoding="utf-8")
+ headings = [
+ "## Current objective and latest user request",
+ "## Authorized scope and prohibited actions",
+ "## Decisions and assumptions",
+ "## Completed work with evidence",
+ "## Modified files and artifacts",
+ "## Pending jobs and next executable step",
+ "## Blockers and checks not run",
+ "## Loaded skill names",
+ "## Secret references",
+ ]
+
+ positions = [prompt.index(heading) for heading in headings]
+ assert positions == sorted(positions)
+ assert "Never include passwords, API keys, tokens, credentials" in prompt
+ assert "Preserve only a secret's name, purpose, storage location" in prompt
+ assert "Keep exact values: file paths, config values, code identifiers, credentials" not in prompt
+ assert "next executable step" in prompt
+ assert "job IDs" in prompt
+
+
def test_pre_compaction_backup_sanitizes_surrogate_text(tmp_path, monkeypatch):
monkeypatch.setattr(
compactor,
@@ -193,3 +222,4 @@ async def test_manual_compaction_clears_active_responses_state(monkeypatch):
assert "response_id" not in state
assert "previous_response_id" not in state
assert state["response_ids"] == ["resp_previous", "resp_current"]
+ assert "ctx_window" not in agent.data