fix(artifact): bound image saves and recover failed publication

[morluto]

Closes #88

Apply one deadline to nonblocking reads from both Docker-save pipes, stream gzip output, and retain a bounded 16 KiB stderr tail. On failure, terminate the owned process group, close the parent pipe ends, and reap the save process. File and byte uploads now reuse the existing attempt-isolated locator scheme so a failed document insert cannot block the next publication.

Reproduced failure	Before	After
Stalled save with saturated stderr; 0.2 s deadline	Still blocked at 2 s; external watchdog required	Fails in 0.206 s
Retry after document insert fails	Three retries collide with orphan object	Next retry publishes version 1
Temporary archive after failed save	Cleanup unreachable while blocked	Removed
Retained stderr	Unbounded	Last 16 KiB; no stderr staging file
Detached child retains stdout/stderr	Hang or cleanup failure	Returns in 1.501 s at a 1.5 s deadline; archive removed

Measured with local subprocesses and real SQLite/filesystem stores. Tests cover stalled stdout, saturated stderr, surviving descendants, compression errors, valid 16 MiB gzip output, concurrent writes, and both file-upload APIs.

Only newly published object locators change. Existing saved locators still load. Failed publication can leave an isolated object; this change makes retry safe rather than adding garbage collection.

Sandbox file-method dispatch preserves providers implementing either the object-store method names or the deprecated S3 names, including inherited overrides and super() delegation. Regression tests cover artifact staging through an older provider and both keyword spellings.

Validation: 6,515 unit/protocol tests passed, 13 capability skips; A2A, sandbox, and protocol tests also pass after merging latest main. Plugin API check passed. Docker/cloud integration runs in CI.

Confidence Score: 5/5

The PR appears safe to merge based on the reviewed changes.

What we checked:

• Failed saves leave no archive: The save helper removes its output on failure, and the caller also removes its temporary file.
• Upload fallback rereads the source: Each upload destination opens a fresh source, so a fallback starts reading at the beginning.

Summary

Docker image saves now stream under one deadline, and file uploads use a separate object location for each attempt so failed publications can be retried. Sandbox providers can keep using either the older or newer file-copy method names.

• Docker image saves stream both pipes under one deadline.
• Each file upload gets its own object location.
• Sandbox providers can use either file-copy method name.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  G[Staging grant] --> H{Staging path and port known?}
  H -- Yes --> L[Try agent server over loopback]
  L -- Connection refused --> P[Try grant URL]
  H -- No --> P
  L -- Server responds --> R[Use response]
  P --> R

Loading

_{Reviews (5) · Last reviewed commit: "Merge remote-tracking branch 'origin/mai..." · Reviewed by Greptile}

[morluto]

Applied the shared sandbox-provider compatibility fix in c14168f, included in head ac4c1ed, and merged latest main. Either PR can merge first without dropping providers that override the deprecated file-transfer methods. The regression covers actual artifact staging, inherited overrides, VM super() delegation, and both keyword spellings.

Validation: 6,515 unit/protocol tests passed with 13 capability skips; after merging latest main, all 901 A2A/sandbox/protocol tests passed. Plugin API check passed. All existing inline review threads are resolved.

[earakely-scale]

🤖 Automated Claude routine — Env Pod PR Context Review. This is a bot, not Edgar Arakelyan; not a human review.

Context from outside this diff: open PR #85 carries the same two changes, byte for byte, and neither PR mentions the other. Both branch from main, so whichever merges first leaves the other conflicting on both files.

• tst/unit/providers/sandbox_providers/sandbox_method_compatibility_test.py is an added file in both. The two copies are identical — sha256 a73cd5f760cb6a88b1633ab6fa2ae3d7e094df8a38ae9cf63b5d2799ce6efafa at this PR's head ac4c1ed and at fix(lifecycle): reject failed image loads and drain gateway work #85's head 8e6adea. Two add/add of the same path is an unavoidable conflict, not a mergeable one.
• In src/agent_env/providers/sandbox_providers/sandbox.py, the _object_file_method / _legacy_file_method helpers plus Sandbox.__init_subclass__ are the same 55-line block in both (here at :137 and :168; fix(lifecycle): reject failed image loads and drain gateway work #85 at :139 and :170, the offset just being its extra import hunk at @@ -47,6 +47,8 @@). The two PRs' remaining sandbox.py hunks differ and don't overlap: this one's are at @@ -182 @@ and @@ -341 @@, fix(lifecycle): reject failed image loads and drain gateway work #85 adds @@ -307 @@ and @@ -422 @@.

Both are currently blocked, and #85 already has a request on it to resolve conflicts, so this may already be in hand. If not, the usual seam is to land the shared compatibility shim plus its test once — in whichever of the two goes first — and rebase the other onto it, rather than resolving the same block twice.

---

Generated by Claude Code