Relationships
#4752 Swampify: a probe users run and share privately so we can add other agents, and the skill reads windows up to all history
Opened by skunk-ape · 11/14/2026· Shipped 11/14/2026
Part of Swampify onboarding, from an external tester's first run.
Repo: swamp-extensions (@swamp/agent-sessions, its agent-sessions and swampify skills).
Goal
Two changes:
- A
probemethod that a user with other agents' history can run, read and share privately, so we learn enough to add those agents. - The swampify skill handles windows longer than Claude Code's default retention, including all history.
1. probe
@swamp/agent-sessions reads Claude Code and Codex only. The tester also uses pi, Amp, Codeium and Windsurf. We need to know where each keeps its history and what it looks like, without their prompts, code or secrets leaving their machine.
Research first
Before writing the probe, find out for each of pi, Amp, Codeium and Windsurf where it keeps session history and in what format, from its docs and, where it is open source, its source. Pi is open source and likely the easiest. Amp syncs threads to its servers, so check whether a full local copy exists. Windsurf and Codeium may share one format, possibly binary. Record what you find in the README's agent support section. It sets the probe's built-in roots.
What it does
swamp behavior run <b> probe # built-in roots
swamp behavior run <b> probe --input root=<path> # any other rootFor each root it finds, it writes one agent-sessions-probe data item and a readable Markdown file:
- Layout. File counts, types and the directory pattern, with ids and dates replaced (
<uuid>/chats/<date>.jsonl); date range and total size. - Format. JSONL, JSON, SQLite (tables, columns and types), Markdown or binary. For JSON, every key path with its types and how often it appears. For SQLite, the same per column.
- Values. Numbers as type and range. Strings as length bucket and shape (timestamp, uuid, path, URL, free text). A short token with no spaces that appears in several sessions (an event type, a role, a model name) is kept verbatim.
- Samples. Up to 3 records per record type, from different sessions, and one whole turn in file order (user message, tool call, tool result, reply). Every sample goes through the extension's redaction (paths, home, secret shapes, tokens, emails, heredoc bodies), and every free-text string is cut to about its first 60 characters, with its length noted. Keys, numbers, timestamps and short tokens stay whole.
- Versions. The probe's version, and each agent's version when its files record one.
- Size. The shareable file is capped at about 100 KB; samples are dropped before structure.
- Binary formats. Layout and size only, and say that samples need the format first.
probe reads only. It calls no model and sends nothing anywhere.
Review and sharing
A section in the agent-sessions skill, and a pointer from swampify when the user mentions an agent it doesn't support:
- Run
probe, then go through the Markdown file with the user: structure first, then samples, pointing out any surviving string longer than a few words. The user confirms, or edits the file, before it is final. Their edited file is the one that counts. - Say plainly: share the file privately with the swamp team, directly. Don't attach it to a Lab issue or post it in a public thread; issues are public.
2. Longer windows in the swampify skill
The start page will offer 180 and 365 days and all history (swamp issue, filed with this one).
- Read the kick-off's window line, including all history, which means calling
assesswith nosince. Agree the exact line with the swamp issue and pin it with a test. - The thin-set step offers to widen "up to Claude Code's retention" (30 days when unset). Widen to what is actually on disk instead: the oldest transcript under the roots in scope.
- Before
understood, the session count already shows and is asked about; keep that for long windows.
Out of scope
- Parsers for pi, Amp, Codeium or Windsurf. Each comes after its probe reports.
- The start page (swamp).
Done when
- The README records where each of the four agents keeps history and in what format, with sources.
probehas unit tests over fixtures seeded with prompts, secrets, emails and absolute paths, in JSONL, JSON and SQLite: none survive into the item or the Markdown, free text is cut, short tokens survive, and the size cap holds.- A property test shows that no input string longer than the cut appears whole in the output.
- Routing evals: an unsupported agent leads to
probeand the review; the all-history kick-off callsassesswith nosince. - Both skills pass
tessl reviewat 90% or above, and their trigger and routing evals still pass. - A manual run of
probeon the implementer's own machine produces a file they would be comfortable sending. No probe output goes in the PR.
Shipped
Click a lifecycle step above to view its details.
system commented 11/14/2026, 4:47:48 PM
Classified automatically when this issue was filed.
- Source: Extensions
If you feel this classification is incorrect, add a ripple to tell us so.
Sign in to post a ripple.