feat/file-transfer #5

Merged

troed merged 27 commits from feat/file-transfer into main

2026-10-14 16:08:27 +02:00

troed commented

2026-10-14 16:08:17 +02:00

Owner

umwelt_copy_file to enable agents to transfer files between local and remote

troed added 27 commits

2026-10-14 16:08:17 +02:00

docs: spec transferring files between sandboxes c1c8f3b0ed

docs: correct the transfer spec's parent-directory and same-file guards 33ec44f127

docs: add the file-transfer plan and align the spec with it ea22536678

feat(client): add chunked read and write requests 44cd7b9cb9

feat(engine): read file chunks by offset 7adce44fdc

docs: harden the transfer plan's read program (symlink stat, unreadable guard) 07200b8baf

fix(engine): dereference symlink sources and refuse unreadable chunks 1d75a65ca2

feat(engine): write positioned file chunks bbccf16f6f

docs: fix the transfer plan's write-chunk argv assertions and the spec's digest command 83c1c8e091

fix(mcp): attach to the control plane lazily, not before initialize ee4a6ba66c

main() called ensure_sandbox() synchronously before reading stdin.
ensure_sandbox returns immediately when the control plane already
answers, but otherwise it launches run.py and waits up to 60s for a
socket that only appears after container creation. During that window
the MCP `initialize` handshake could not be answered, so a client that
evicts/reboots its MCP servers saw umwelt as unavailable for 5-22s
while every other server reconnected in under 1.5s, and a tool call in
flight across the window was orphaned.

Start with an empty ctx instead. `initialize`, `ping` and `tools/list`
need no control plane, and dispatch()'s tools/call branch already
attaches lazily through attach(), which stores a launched control plane
in ctx["proc"]. Teardown is untouched: main()'s finally block already
reads ctx["proc"], so a lazily launched control plane is still
SIGTERMed on exit and an attached one still is not. A failed lazy
attach keeps reporting through dispatch()'s existing ctx["error"] path,
with no new error shape.

docs: correct the attach-time wording and record the harness MCP-server churn workaround 39bf15e465

docs: correct the transfer plan's requirement for bytes stdin (text mode raises) b707316658

docs: correct the spec's bytes-stdin requirement too 01c9081648

fix(engine): give podman() a real bytes path so chunk writes reach the container 2cc5d393bf

docs: hand over the file-transfer work in progress 2a5467a4be

feat(mcp): stream a copy between two chunk endpoints 59c0cc1121

feat(mcp): add the copy_file tool f6fc123a04

fix(mcp): take the transport's payload directly in copy_file dace414ec5

fix(mcp): rebind the dropped transport and report a failed copy's progress 121a12e2a9

fix(mcp): make copy_file endpoint resolution fail as a tool error e4768293d1

feat(install): allow umwelt_copy_file in the managed policy 183112ab9b

test(integration): copy files between real sandboxes cef2c81b22

test(integration): bring up the control plane and cover the empty/remote cases 4d960c5642

docs: describe file transfer between sandboxes 78a146b94a

Document umwelt_copy_file (Tasks 1-7), which moves one file byte-exactly
between any two sandboxes, local or remote, without the bytes entering
the model's context.

- CONFIGURATION.md: a tool-list bullet plus a new "Transfers" section -
  the endpoint/host shape, the result (including its optional `host`), the
  context-cap exemption, in-place overwrite with the partial-destination
  note (the error reports the offset reached and bytes written, not a
  destination digest), parent-dir creation, preserve_mode on a new
  destination only, the tar idiom, and the stale-agent remedy.
- README.md: the managed-policy sample gains the copy_file allow rule
  (it was out of sync with POLICY_ALLOW after Task 3).
- AGENTS.md, umwelt/plugin/AGENTS.md: one pointer each.
- umwelt/plugin/skills/umwelt/: SKILL.md gains a Where-to-start row and
  the CONFIGURATION.md#transfers anchor; references/remote.md spells out
  that copy_file takes `host` on each endpoint and has no top-level `host`.
- docs/specs: correct the streaming model's round-trip count (75 -> 76;
  the shipped loop makes one extra read for the stat probe).

The installed skill copy is STALE: `install.sh install` was deliberately
not run here (it mutates host state outside this worktree and is the
operator's step). The skill sources edited here are the deliverable;
re-render with `bash umwelt/install.sh install`.

docs: correct the stale-agent symptom and the tree example fence 4d9b6b7aca

Fix-round 1 for the file-transfer docs (78a146b).

- The stale-agent symptom is not always 'read_chunk': a copy whose remote
  endpoint is the destination fails first on its first write with
  'write_chunk' (the front door reads the local source, then writes to the
  remote). Both CONFIGURATION.md and the skill's references/remote.md now
  name both methods; the remedy and the no-PROTOCOL_VERSION-bump note are
  unchanged.
- Quote the tool's actual error line, with the prefix it prepends:
  `umwelt: copy_file: host '<name>' has an older umwelt agent; re-run
  umwelt-ctl remote-install <name>`.
- Retag the tree example fence from `sh` to `text`: it holds pseudo
  tool-call syntax, not shell, and a `sh` tag invites a bad paste.

The installed skill copy is still STALE (install.sh install is the
operator's step and was deliberately not run).

fix(mcp): report both digests and warn when a copy does not verify b3fdfe6eb6

copy_between already had the destination digest in final_write["sha256"]
(the value verified is derived from); it was simply dropped from the
result. A mismatch now carries dst_sha256 and a loud warning, and the
umwelt-copy audit line records verified so a mismatch is not logged as
plain success. Also: the zero-byte progress note no longer claims the
destination was untouched (the first write truncates before it writes),
and the copy_file dispatch comment no longer claims the lazy local attach
is skipped (the local control plane is still needed to read [[remote]]).

test(copy): pin the mismatch digests, the zero-byte note and integration cleanup 3f0092019a

- test_mcp: the mismatch test asserts dst_sha256 and the warning, plus a
  new test pinning the verified-copy shape (no dst_sha256, no warning) and
  one pinning the zero-byte progress note.
- test_run: the idempotency test recorded kw["input"] while the engine
  passes stdin, so the captured input was always None; record stdin.
- test_client: read_chunk/write_chunk join the timeout-default table.
- test_integration: register the agent container with env.names in both
  remote copy tests, make the and-back audit assertion exact, and fix the
  inverted peer labels (the relay leg is not "local").

troed merged commit 1f86c005fa into main

2026-10-14 16:08:27 +02:00

troed deleted branch feat/file-transfer

2026-10-14 16:08:27 +02:00

troed referenced this pull request from a commit

2026-10-14 16:08:27 +02:00

feat/file-transfer (#5)