[robstolarz]

A reply longer than 4096 bytes came back from depot agent cut, ending [N bytes truncated]. The CLI now reads such a reply whole. DEP-8516 (CLI half). Stacked on #687.

• reply in JSON, agent logs and agent watch call GetSessionMessage for a reply the view marks with truncatedBytes > 0.
• A server without the RPC, or a read that fails, keeps the cut text and its marker. The command never fails because of it. A transient failure is tried again on later frames, and before a wait ends on the frame that carried it.
• The read counts against --timeout: a read still open at the deadline exits 2 (timeout), not 0 with a cut reply.
• A long --watch keeps going when the server ends the connection under the stream (GOAWAY) and picks up from its cursor.

Proven live at this head on a test org: a 2500-line reply returned 11392 bytes complete from create --watch -o json (5 of 5 runs) and from agent logs. One run made as the server deploy was finishing returned the cut text and exit 0, which is the older-server path.

Also live, at the head before the last fix: two create --watch sessions that ran a 20-minute command each exited 0 after 21 minutes with the whole 11392-byte reply.

Not proven live: a failed read recovering on a later try, and a connection cut under the stream (tests only; neither happened in these runs).

Post-deploy test plan

Post-deploy test plan

Needs a CLI release with this change; until then build this branch (go build -o depot ./cmd/depot).

depot agent create "Reply with the numbers 1 through 2500, one per line, and nothing else." --watch --timeout 10m -o json > out.json; echo "exit $?"
jq -r '.reply | (utf8bytelength | tostring), (split("\n") | last)' out.json
depot agent logs "$(jq -r .sessionId out.json)" | grep -c truncated

Expected: exit 0, 11392, 2500, then 0.

Agent context

• Only assistant replies are read whole: tool output and replies that carry tool calls keep the bounded view text.
• A whole reply is cached per message id and dropped once the message leaves the view; a reply recovered on a later frame is printed as plain lines so a response streaming for another turn keeps its place.
• The read makes 3 tries, each with its own 30s, inside the watch deadline; a read that used all three gets up to 3 more rounds before a non-running frame is judged. retryable covers unavailable, deadline, unknown, internal, and the client's "incomplete envelope" for a connection cut under a stream.
• Proto: proto/depot/agent/v1/depot_agent_core.proto synced for GetSessionMessage and regenerated; its comments come from the source proto.
• Tests (pkg/cmd/agent/session_test.go): whole reply in watch, logs and JSON; older server; failed read then later frame; eviction; recovered reply not restarting the next stream; read outlasting --timeout; a hung first try; a wait ending on the frame whose read failed; a GOAWAY at frame level.
• Local review: codex clean at this head against the Add session commands to depot agent: create, send, list, watch, logs, stop (DEP-8506) #687 branch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RQk3K7GpvQg7ySXyCrirEe

---

Note

Medium Risk
Changes watch/logs networking, timeout semantics, and stream reconnection behavior; failures degrade to truncated text rather than hard errors, but long-running watches now depend on a new RPC.

Overview
Adds GetSessionMessage to the agent API and wires the CLI so truncated assistant replies (from the bounded session view) are fetched in full for agent watch and agent logs text output and for reply in JSON watch results, while -o json logs still emit the API view unchanged.

The renderer gains Whole / Recover / keep: it calls the new RPC for assistant messages with truncatedBytes, caches full text, retries transient failures on later frames, drops cache when messages leave the view, and surfaces cut or failed reads in the transcript without failing the command.

agent watch integrates bounded retries/timeouts for whole-message reads (counting against --timeout), Recover before returning a settled reply, and treats HTTP/2 GOAWAY / incomplete stream errors as retryable so long watches reconnect from cursor. Proto and Connect stubs are regenerated; tests cover GOAWAY, long replies, timeouts, and older APIs without the RPC.

^{Reviewed by Cursor Bugbot for commit 8eda9ad. Bugbot is set up for automated code reviews on this repo. Configure here.}

[linear-code]

DEP-8516

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale comment

[!NOTE]
🤖 Cursor Grok 4.6 responding on behalf of Jacob Gillespie

Risk: medium. Cursor Bugbot reported two unresolved findings on whole-reply timeout retries and idle-frame recovery, so this is not approved. Human review is needed and no reviewers were assigned.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}

[claude]

No new blocking issues. 1 optional suggestion (a nit or a note on pre-existing code) was found and not posted. Nothing in this review needs a push before merging.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 86a9bc6. Configure here.}

[cursor]

Stale comment

[!NOTE]
🤖 Cursor Grok 4.6 responding on behalf of Jacob Gillespie

Risk: medium. Cursor Bugbot reported one unresolved finding on the whole-reply reread heuristic, so this is not approved. Human review is needed and no reviewers were assigned.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}

[claude]

Code review found no issues

No high-confidence issues detected in this change.

[cursor]

Note

🤖 Cursor Grok 4.6 responding on behalf of Jacob Gillespie

Risk: medium. Cursor Bugbot completed successfully with no unresolved findings on this head, so this is approved.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}

[robstolarz]

From Claude: delegate merge approval (local codex review gate).

• Gated head: f99519b7ac9873e8961969f2526aea491526dfac (this PR), reviewed from the top of the pair at f99519b7ac9873e8961969f2526aea491526dfac, so the run covers Add session commands to depot agent: create, send, list, watch, logs, stop (DEP-8506) #687 and fix(agent): read a long reply whole, keep a long watch alive (DEP-8516) #688 together
• Base: origin/main at 0c3d37df5cca0622a5f6a848c6713c897b82256a
• Command: codex -s read-only -C <worktree at f99519b7ac> review --base origin/main
• Verdict (round 6, finished 2026-11-15T02:42:07Z): "I found no actionable bugs in the diff. I could not run the Go tests because the read-only environment prevented Go from creating its build directory." Go tests ran in this PR's CI, all passing.
• Earlier rounds' findings were fixed in code before this run.
• At merge: approved, 0 open review threads, every check passing, Cursor Bugbot passing on this head.

[claude]

Code review found no issues

No high-confidence issues detected in this change.

[robstolarz]

Proof addendum: head is now 8eda9adcea26f0048cfc06d350e70eae0062022b, GitHub's merge of main into the branch. Its tree 16a6f3af606564cdb3bff6767dd1382464fb9445 is identical to the tree of the reviewed head f99519b7ac, so the clean local codex review above covers it. Re-checked at this head: approved, 0 unresolved threads, all checks passing.