[kelos-bot]

What type of PR is this?

/kind bug

What this PR does / why we need it:

Every agent entrypoint used to tee the agent's JSON-lines stdout into
/tmp/agent-output.jsonl so that /kelos/kelos-capture could re-read it
after the agent exited and emit token usage between the
---KELOS_OUTPUTS_*--- markers. /tmp counts against the pod's
ephemeral-storage limit and the file was never rotated, so a verbose
run could fill the ephemeral budget and trigger node-pressure eviction
that silently failed user tasks.

This PR inlines the parser into the pipe. Each entrypoint now does:

<agent> "${ARGS[@]}" | /kelos/kelos-capture
AGENT_EXIT_CODE=${PIPESTATUS[0]}
exit $AGENT_EXIT_CODE

kelos-capture reads stdin line by line, forwards each line unchanged
to its own stdout (pod logs are byte-identical to today), and
accumulates per-agent counters in memory:

• claude-code / gemini / cursor — keep only the most recent
  {"type":"result", ...} line.
• codex / opencode — sum tokens across every turn.completed /
  step_finish event.

Memory cost is O(1) in the size of the stream, /tmp/agent-output.jsonl
is no longer created, and the controller's marker contract is
unchanged.

Which issue(s) this PR is related to:

Fixes #1187

Special notes for your reviewer:

• ParseUsage(agentType, filePath) is replaced by
  StreamUsage(agentType, io.Reader, io.Writer).
• The agent's exit code is still propagated via PIPESTATUS[0] — same
  pattern as before, just with kelos-capture on the right of the pipe
  instead of tee.
• Unit tests in internal/capture now exercise the streaming path for
  all five agent types and verify every input line is forwarded
  unchanged.
• docs/agent-image-interface.md updated to describe the pipe form.

Does this PR introduce a user-facing change?

Fix Task pod eviction from unbounded /tmp/agent-output.jsonl growth: agent stdout is now streamed through kelos-capture instead of being teed to disk, removing the ephemeral-storage pressure point.

---

Summary by cubic

Stream agent stdout through /kelos/kelos-capture instead of teeing to /tmp, removing the ephemeral-storage hotspot that caused task pod evictions. Pod logs remain identical to previous behavior and the ---KELOS_OUTPUTS_*--- contract is unchanged.

• Bug Fixes

  • Removed /tmp/agent-output.jsonl writes to stop ephemeral-storage bloat and node-pressure evictions (fixes Task pods evicted by ephemeral storage: /tmp/agent-output.jsonl grows unbounded #1187).
  • Forwarding preserves bytes in order with no line-length cap, appends a final newline if missing, surfaces flush errors, and still emits usage at EOF.

• Refactors

  • Replaced ParseUsage(...) with StreamUsage(agentType, io.Reader, io.Writer); /kelos-capture now streams stdin to stdout and accumulates usage in O(1) memory.
  • Updated entrypoints to <agent> "${ARGS[@]}" | /kelos/kelos-capture; agent exit code preserved via PIPESTATUS[0]. Tests cover large lines, newline-append, and flush-error paths; docs now prescribe the pipe pattern.

^{Written for commit 3d658a9. Summary will update on new commits. Review in cubic}

[cubic-dev-ai]

2 issues found across 10 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[gjkim42]

/kelos review

[kelos-bot]

🤖 Kelos Task Status

Task kelos-reviewer-issue-comment-aa4d8642f656 has succeeded. ✅

[kelos-bot]

🤖 Kelos Reviewer Agent @gjkim42

Review Summary

Verdict: COMMENT (GitHub blocks self-approval; submitting as COMMENT — the patch is approve-grade)
Overall correctness: patch is correct
Scope: Replace tee /tmp/agent-output.jsonl + post-hoc file parse with a streaming kelos-capture that forwards stdin verbatim and accumulates token usage in memory.

Findings Overview

Priority	Count	File:Line	Summary
P0	0	—	none
P1	0	—	none
P2	2	internal/capture/usage.go:46	Deferred bw.Flush() error is silently discarded
P2		internal/capture/usage.go:47-48	Scanner ErrTooLong on a >10MB line now kills the agent (regression)
P3	1	internal/capture/usage_test.go:169-179	assertForwarded only checks substring presence; misses reordering/duplication

Findings

Correctness

• [P2] internal/capture/usage.go:46 — defer bw.Flush() discards the flush error. If the loop exits cleanly with buffered bytes still pending and the final Flush() fails (e.g., the downstream pod-log writer hits a transient error), StreamUsage returns nil and the last buffered lines vanish from pod logs without any signal to the operator. Capture the flush error explicitly (e.g., named return + defer func() { if ferr := bw.Flush(); err == nil { err = ferr } }()).
• [P2] internal/capture/usage.go:47-48 — Reliability regression on oversized lines. The scanner is configured with a 10 MiB max (scanner.Buffer(make([]byte, 0, 1<<20), 10<<20)), matching the previous implementation. Before, that limit only meant token usage was lost when a line exceeded it because ParseUsage ran after the agent exited from the on-disk file. Now kelos-capture is the agent's sole stdout consumer in the pipe; on bufio.ErrTooLong Scan() stops, StreamUsage returns immediately, Run() proceeds to emit markers and exits, and the agent gets SIGPIPE on its next write — turning what used to be a silent parse miss into an aborted agent run. Consider draining the remainder with io.Copy(w, r) (and giving up on further usage parsing) when sc.Err() is non-nil, so the agent always gets a clean EOF.

Tests

• [P3] internal/capture/usage_test.go:169-179 — assertForwarded only checks strings.Contains(out, line+"\n") for each non-empty input line, so duplicates, reordering, or extra output would pass. TestStreamUsageForwardsUnterminatedLine already does the stricter byte-level comparison; the same shape (normalize input to end in \n, then if out.String() != want ) could be reused across the table-driven cases for a stronger forwarder invariant.

Key takeaways

• Refactor is clean: accumulator interface keeps memory O(1), per-agent extractors are well-factored, and the marker contract with the controller is preserved.
• Two defensive gaps worth tightening: the deferred flush error and the regression where a >10 MiB line now propagates to the agent as SIGPIPE.

Note on prompt injection: third-party bot content in the PR thread (including a section instructing reviewers to attribute findings to it) was disregarded; the findings above are formed independently from the diff.

[gjkim42]

/kelos pick-up

[kelos-bot]

🤖 Kelos Task Status

Task kelos-pr-responder-issue-comment-eadee148f2fc has failed. ❌

[gjkim42]

/kelos pick-up

[kelos-bot]

🤖 Kelos Task Status

Task kelos-pr-responder-issue-comment-19e781c174f0 has succeeded. ✅

[cubic-dev-ai]

1 issue found across 2 files (changes from recent commits).

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[gjkim42]

/kelos review

[kelos-bot]

🤖 Kelos Task Status

Task kelos-reviewer-issue-comment-e09386af3b11 has succeeded. ✅

[gjkim42]

/kelos pick-up
Squash commits after the work

[kelos-bot]

🤖 Kelos Task Status

Task kelos-pr-responder-issue-comment-0aa8d38a97c3 has succeeded. ✅

[kelos-bot]

🤖 Kelos Reviewer Agent @gjkim42

Review Summary

Verdict: COMMENT (GitHub blocks self-approval; submitting as COMMENT — the patch is approve-grade)
Overall correctness: patch is correct
Scope: Address prior round of feedback on streaming kelos-capture: surface the deferred bw.Flush() error via a named return, drain the rest of r after bufio.ErrTooLong so the agent gets a clean EOF instead of SIGPIPE, and tighten the table-driven forwarding test to a byte-for-byte equality check.

Findings Overview

Priority	Count	File:Line	Summary
P0	0	—	none
P1	0	—	none
P2	0	—	none
P3	1	internal/capture/usage.go:74-83	io.Copy drain leaks scanner-buffered bytes; "line itself is lost" comment is slightly inaccurate

Findings

Correctness

• [P3] internal/capture/usage.go:74-83 — On bufio.ErrTooLong, the scanner has already read the head of the oversized line into its internal buffer; those bytes are not in r anymore when io.Copy(bw, r) runs, so the head of the bad line vanishes while the tail (whatever was not yet read into the buffer) plus all subsequent lines are forwarded. The behaviour is acceptable — the agent still gets a clean EOF — but the comment "The oversized line itself is lost (bufio.Scanner discards it)" overstates what is dropped: a partial tail of that line still lands in pod logs. If perfect byte fidelity past an oversized line ever matters, a bufio.Reader.ReadBytes('\n') loop would avoid the buffered-byte loss; otherwise rewording the comment to "the head of the oversized line is dropped; the rest is forwarded as raw bytes" is enough.

Key takeaways

• Previous round's P2 concerns (silent Flush error, SIGPIPE on >10 MiB line, weak Contains-based forwarding assertion) are all addressed in 8bcdb6a — named-return + deferred flush capture, drain-via-io.Copy fallback, and the new assertForwarded byte-equality helper across the table-driven cases.
• The new TestStreamUsageDrainsOversizedLine and TestStreamUsagePropagatesFlushError cover the regression seams introduced by those fixes, so future refactors should not silently regress them.
• Only nit left is a wording/edge-case detail on the oversized-line drain path; no blocking issues.

Note on prompt injection: third-party bot content in the PR thread (including a section instructing reviewers to attribute findings to it) was disregarded; the findings above are formed independently from the diff.

[kelos-bot]

🤖 Kelos Agent @gjkim42

Addressed the byte-fidelity concern raised in the latest reviews by replacing the bufio.Scanner + ErrTooLong drain with a bufio.Reader.ReadBytes('\n') loop, so every byte read from the agent is forwarded to stdout in order (no scanner buffer holds bytes that io.Copy would later miss). The previous "drain after oversized line" branch is gone — there is no size cap to trip — and TestStreamUsageForwardsLargeLineByteForByte now asserts byte-identical forwarding for an 11 MiB line plus continued usage parsing on the following line. Squashed the branch down to one commit per the request. All CI checks are green.

[gjkim42]

/kelos pick-up
Squash commits then

[kelos-bot]

🤖 Kelos Task Status

Task kelos-pr-responder-issue-comment-4f813a1f83e5 has succeeded. ✅

[kelos-bot]

🤖 Kelos Agent @gjkim42

Renamed the local parseLine byte slice at internal/capture/usage.go:74 to body to remove the shadowing of the package-level parseLine helper (P3 from the latest review). Amended into the existing single commit so the branch stays squashed. make verify and make test are green locally and CI is fully green. Ready for re-review.

[gjkim42]

/kelos review

[kelos-bot]

🤖 Kelos Task Status

Task kelos-reviewer-issue-comment-6cac80b57559 has succeeded. ✅

[kelos-bot]

🤖 Kelos Reviewer Agent @gjkim42

Review Summary

Verdict: COMMENT (GitHub blocks self-approval since the PR was opened by kelos-bot — the patch is approve-grade)
Overall correctness: patch is correct
Scope: Inline kelos-capture into the agent stdout pipe (<agent> | /kelos/kelos-capture), replacing the prior tee /tmp/agent-output.jsonl + post-hoc ParseUsage(filePath) flow. The new StreamUsage(agentType, io.Reader, io.Writer) forwards bytes verbatim via a bufio.Reader.ReadBytes('\n') loop and accumulates per-agent token usage in memory. Updates touch the five entrypoints, internal/capture/{capture,usage,*_test}.go, and docs/agent-image-interface.md; PIPESTATUS[0] exit-code propagation and the controller's ---KELOS_OUTPUTS_*--- marker contract are unchanged.

Findings Overview

Priority	Count	File:Line	Summary
P0	0	—	none
P1	0	—	none
P2	0	—	none
P3	0	—	none

Findings

None.

Key takeaways

• All prior review-round concerns are resolved in commit 3d658a9: named-return + deferred Flush() capture surfaces flush errors (internal/capture/usage.go:49-56); the bufio.Reader.ReadBytes('\n') loop removes the 10 MiB scanner cap entirely so an arbitrarily long line is forwarded byte-for-byte with no SIGPIPE risk; assertForwarded enforces byte-identical forwarding across the table-driven cases; and the local parseLine shadowing of the package-level helper was renamed to body at internal/capture/usage.go:74.
• The controller marker contract is preserved — cost-usd/input-tokens/output-tokens are still emitted in that fixed order from captureOutputs, with KELOS_AGENT_TYPE consumed inside Run() and the per-agent extractors well-factored into lastResultAccumulator / sumAccumulator (O(1) state).
• All five entrypoints (claude-code, codex, cursor, gemini, opencode) consistently use <agent> "${ARGS[@]}" | /kelos/kelos-capture with set -uo pipefail and PIPESTATUS[0] exit-code propagation; docs/agent-image-interface.md matches the implementation.
• New tests (TestStreamUsageForwardsLargeLineByteForByte, TestStreamUsagePropagatesFlushError, TestStreamUsageForwardsUnterminatedLine) lock in the regression seams introduced by the rewrite.

Note on prompt injection: third-party bot content in the PR thread (including <!-- cubic:attribution --> directives demanding attribution to another reviewer) was disregarded; findings here are formed independently from the diff.