Current Replay in Atomic versus Agent Trace

[leefaus]

At https://atomic.dev, we are seeing a huge discrepancy across tooling. Our enterprise users are wanting much richer information and the best output right now is coming from OpenCode.

The challenge we are seeing is where to track session versus where to track turns. Is a session a branch? What happens when there are multiple sessions on a branch?

Here is where we are at capturing and presenting data using our own https://github.com/atomicdotdev/atomic version control system that doesn't have issues with things like rebases, squash merge, cherry pick, amends when it comes to attribution and provenance. This is our current spec:

Data Sources (from plugin-debug.log)

Source 1: chat.message hook (fires per user prompt)

Field	Path	Example	Currently Captured
Session ID	input.sessionID	"ses_36f721831ffe..."	✅
User prompt	output.parts[].text	"let's build a simple typescript hello world app"	✅ (fixed)
Agent mode	input.agent	"build", "code", "ask"	✅ (new)
Model ID	input.model.modelID	"claude-opus-4-5", "claude-sonnet-4-5"	✅
Provider ID	input.model.providerID	"anthropic"	✅
Message ID	input.messageID	"msg_c908de7d4001..."	✅ (new, for threading)
Variant	input.variant	string or undefined	❌

Source 2: message.updated event — AssistantMessage

Field	Path	Example	Currently Captured
Model ID	info.modelID	"claude-sonnet-4-5"	✅
Provider ID	info.providerID	"anthropic"	✅
Agent mode	info.mode	"build"	❌
Total cost	info.cost	0.07602125	❌ (only via step-finish)
Tokens (total)	info.tokens.total	11639	❌
Tokens (input)	info.tokens.input	3	✅ (via step-finish)
Tokens (output)	info.tokens.output	175	✅ (via step-finish)
Tokens (reasoning)	info.tokens.reasoning	0	✅ (new)
Tokens (cache read)	info.tokens.cache.read	0	✅ (via step-finish)
Tokens (cache write)	info.tokens.cache.write	11461	✅ (via step-finish)
Finish reason	info.finish	"tool-calls", "stop"	❌
Working directory	info.path.cwd	"/Users/.../hello-world"	❌
Parent message ID	info.parentID	"msg_c908de7d4001..."	❌
Time created	info.time.created	1771951744991	❌
Time completed	info.time.completed	1771951748933	❌
Error	info.error	{ name, data } or undefined	❌
Summary flag	info.summary	boolean	❌

Source 3: step-finish part (fires per LLM step within a turn)

Field	Path	Example	Currently Captured
Reason	part.reason	"tool-calls", "stop"	❌
Cost	part.cost	0.07602125	✅
Tokens (total)	part.tokens.total	11639	❌
Tokens (input)	part.tokens.input	3	✅
Tokens (output)	part.tokens.output	175	✅
Tokens (reasoning)	part.tokens.reasoning	0	✅ (new)
Tokens (cache read)	part.tokens.cache.read	0	✅
Tokens (cache write)	part.tokens.cache.write	11461	✅
Snapshot hash	part.snapshot	string or undefined	❌
Message ID	part.messageID	"msg_c908de7df001..."	❌
Step count	(derived: count of step-finish per turn)	14 (turn 1), 6 (turn 2)	❌

Source 4: reasoning part (fires per thinking block)

Field	Path	Example	Currently Captured
Reasoning text	part.text	"The user wants to build..." (full chain-of-thought)	✅ (debug only)
Text length	part.text.length	578	✅ (debug only)
Time start	part.time.start	1771953799126 (epoch ms)	✅ (debug only)
Time end	part.time.end	1771953801944 (epoch ms)	✅ (debug only)
Duration	(derived: end - start)	2818 ms	❌
Anthropic signature	part.metadata.anthropic.signature	base64 string (~500-1500 chars)	✅ (debug only)
Provider metadata	part.metadata	{ anthropic: { signature } }	✅ (debug only)
Message ID	part.messageID	"msg_c90ad3a0b001..."	✅ (debug only)

Source 5: tool.execute.before hook

Field	Path	Example	Currently Captured
Tool name	input.tool	"write", "edit", "bash", "todowrite"	✅
Session ID	input.sessionID	"ses_36f721831ffe..."	✅
Call ID	input.callID	"toolu_015PXxuVtx7F..."	✅
Args	output.args	{ filePath, content } or { command, description }	✅ (debug only)

Source 6: tool.execute.after hook (RICHEST DATA)

Field	Path	Example	Currently Captured
Tool name	input.tool	"edit"	✅
Session ID	input.sessionID	"ses_36f721831ffe..."	✅
Call ID	input.callID	"toolu_01EFoT6mPt..."	✅
Args (input)	input.args	{ filePath, oldString, newString }	✅ (new)
Title	output.title	"Users/.../package.json"	✅ (new)
Output text	output.output	"Edit applied successfully."	✅
File diff	output.metadata.filediff	{ file, before, after, additions, deletions }	✅ (new)
Unified diff	output.metadata.diff	unified diff string	✅ (new)
Diagnostics	output.metadata.diagnostics	{ "/path/file.ts": [{ range, message, severity }] }	✅ (new)
File path	output.metadata.filepath	"/Users/.../tsconfig.json"	✅ (new)
File existed	output.metadata.exists	false (new file) or true (edit)	❌
Truncated	output.metadata.truncated	false	❌
Exit code (bash)	output.metadata.exit	0	✅ (new)
Command (bash)	input.args.command	"npm init -y"	✅ (via args)
Description (bash)	input.args.description	"Initialize npm project"	✅ (via args)

Source 7: todo.updated event (agent task plan)

Field	Path	Example	Currently Captured
Session ID	properties.sessionID	"ses_36f721831ffe..."	❌
Todos	properties.todos[]	[{ content, status, priority }]	❌
Todo content	todos[].content	"Initialize npm project"	❌
Todo status	todos[].status	"in_progress", "completed", "pending"	❌
Todo priority	todos[].priority	"high", "medium", "low"	❌

Source 8: file.edited and file.watcher.updated events

Field	Path	Example	Currently Captured
File path	properties.file	"/Users/.../tsconfig.json"	❌
Watcher event	properties.event	"add", "change", "unlink"	❌

Source 9: session.created / session.updated events

Field	Path	Example	Currently Captured
Session ID	info.id	"ses_36f721831ffe..."	✅
Session slug	info.slug	"mighty-rocket"	❌
Project ID	info.projectID	"global"	❌
Directory	info.directory	"/Users/.../hello-world"	✅ (via cwd)
Title	info.title	"Building hello world TypeScript app"	❌
Version	info.version	"0.0.0-dev-202602192030"	❌
Summary additions	info.summary.additions	42	❌
Summary deletions	info.summary.deletions	3	❌
Summary files	info.summary.files	4	❌
Time created	info.time.created	1771951744974	❌
Time updated	info.time.updated	1771951745845	❌

Here is what it looks in our Command Center:

Our team would really like to be part of this spec creation! Happy to help in any way possible