Request access to this agent session
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository holds a coding-agent trace. Agent traces can contain source code, local file paths, command output and other context from the machine that produced them. Access is granted per user so the owner can see who has read the session.
Log in or Sign Up to review the conditions and access this dataset content.
Agent Manager session — designing session sharing (gated)
Access: gated. The repo and this card are publicly listed, but the files — including
<session-uuid>.jsonl— require per-user approval. This is the middle ground for a teammate when you do not want a fully public trace and do not have an organization to put it in: a user-owned private dataset cannot have collaborators at all, so gating is the only way to grant one specific person access under your own namespace.Trade-off: the repo name, this card and the metadata are visible to everyone. Only the trace contents are protected.
A single real Agent Manager session, exported as a shareable trace. This is the working session in which the session-sharing design itself was researched and written — surveying how five coding-agent harnesses persist sessions, testing what the Hub's trace viewer can do, and producing a design doc.
View it
Open Data Studio and click the single row: the Hub renders the full session — prompts,
assistant turns, tool calls, token counts, thinking blocks. The repo holds one session, so
it is always row=0, which makes it directly embeddable:
<iframe src="https://huggingface.co/datasets/thomwolf/am-session-sharing-design-gated/embed/viewer/default/train?row=0"
frameborder="0" width="100%" height="700px"></iframe>
What's in here
| Path | Contents |
|---|---|
<session-uuid>.jsonl |
the session, verbatim Claude Code JSONL (a Hub-native trace format) |
meta/manifest.json |
provenance: harness + version, session ids, stats, content hash, lineage |
meta/briefing.md |
mechanically generated handoff summary — what was asked, tools used, files written |
meta/redaction.json |
which redaction rules ran and what they caught |
Session at a glance
- Harness: Claude Code 3.3a ·
claude-opus-5[1m] - 7 user prompts · 38 assistant turns · 40 tool calls
- 95,795 input / 39,337 output tokens (2.68M cache read)
- Tools:
Bash×23,Read×6,WebSearch×4,WebFetch×4,Skill,ToolSearch,Write
Provenance and redaction
Exported by the phase-1 prototype described in the design doc: locate → redact → assemble
→ publish. Before upload the trace went through redaction ruleset v1 — token-shaped
patterns (hf_, sk-ant-, gh?_, AWS, Google, JWT, private keys), value matching
against the Space's actual secret-bearing environment variables, and an email rule.
Result: no secret matches, 7 email addresses redacted (shown as «redacted:email»),
and 8 file-history-snapshot/delta lines dropped — those embed full file contents and
are the single worst leak vector in a Claude transcript.
The source code visible in this trace belongs to Agent Manager, which is published as a public template Space; the files quoted here were verified byte-identical to that public copy before publishing.
Traces can still contain local paths, command output and private context. Review before publishing your own.
Forking it
The transcript filename is a Claude Code session id, so a recipient can continue this
conversation: place <session-uuid>.jsonl at <CLAUDE_CONFIG_DIR>/projects/<cwd-slug>/<new-uuid>.jsonl
and run claude --resume <new-uuid>. Always copy — --resume appends to the
transcript, so resuming a shared artifact in place would mutate the thing that was meant
to be frozen. meta/manifest.json records trace.nativePath for exactly this.
- Downloads last month
- 9