8bdf9698 - Surface the vendor's actual output from a lane run

[Danswar]

EN:
agent lane run genuinely invokes grok/codex against a real diff, but the CLI only ever printed a one-line STATUS/rc summary and threw away the vendor's actual output. By the time launch() returns, the tmux pane it was captured from is already killed, so nothing downstream (gate record --evidence, a human watching the run) could ever see what the vendor found. Fixed both call sites (agent lane run and the agent run one-step chain) to print result.stdout/result.stderr before the summary line, and sanitize that output (strip C0/C1/DEL control bytes) before printing since the --no-tmux path captures raw vendor subprocess output unlike the tmux path's already-escape-free capture-pane.

DE:
agent lane run ruft grok/codex tatsächlich mit einem echten Diff auf, aber die CLI hat bisher nur eine einzeilige STATUS/rc-Zusammenfassung ausgegeben und die eigentliche Vendor-Ausgabe verworfen. Sobald launch() zurückkehrt, ist die tmux-Pane, aus der sie stammt, bereits beendet — nichts Nachgelagertes (gate record --evidence, ein Mensch, der zusieht) konnte je sehen, was der Vendor gefunden hat. Beide Aufrufstellen (agent lane run und die agent run-Ein-Schritt-Kette) geben jetzt result.stdout/result.stderr vor der Zusammenfassungszeile aus und sanitisieren diese Ausgabe (C0/C1/DEL-Kontrollbytes entfernt), da der --no-tmux-Pfad rohe Vendor-Subprocess-Ausgabe erfasst, anders als der tmux-Pfad, dessen capture-pane bereits escapefrei ist.

Details

Replaces #37, rebuilt cleanly off the current develop tip with a single commit and a CONTRIBUTING.md-compliant subject line (the commits on #37 were missing the required trailing period; fixing that in place would have meant amend+force-push on already-pushed, already-reviewed commits, which is not done here even for a message-only fix — see #37 for that PR's original review history).

Found while manually driving a diagnostic review task against RealUnitCH/web#27's real diff through Jonny Luca's dfxai instance: the grok quality lane ran for real (STATUS=complete) but produced no retrievable findings anywhere. Traced to cmd_lane/cmd_run in main.py never touching result.stdout.

Full local suite green: 446 passed, 1 skipped (agent-core integration test, needs that package importable).

[Danswar]

EN:
Ready after 4 review passes.
Both call sites of agent lane run (the lane and one-step run chains) now print and sanitize the vendor's actual stdout/stderr instead of discarding it.

DE:
Bereit nach 4 Review-Durchläufen.
Beide Aufrufstellen von agent lane run (die Lane- und die Ein-Schritt-run-Kette) geben nun die tatsächliche Vendor-Ausgabe (stdout/stderr) sanitisiert aus, statt sie zu verwerfen.

Details

This PR replaces #37 (closed) after a CONTRIBUTING.md finding that #37's commit subjects were missing the required trailing period; fixing that in place would have meant amend+force-push on already-reviewed, already-pushed commits, so the branch was rebuilt cleanly instead. The 4 review passes span both PRs:

1. 8a4d464a - Surface the vendor's actual review output from a lane run #37, commit 643a129: grok quality/logic + codex quality/logic all reviewed the initial fix. Codex logic found a test-coverage gap (only the cmd_lane call site had a content-asserting test, not cmd_run) and flagged raw, unsanitized vendor stdout/stderr on the --no-tmux path as a potential ANSI/control-sequence injection vector.
2. 8a4d464a - Surface the vendor's actual review output from a lane run #37, commit ba43401: added the missing cmd_run test and a _sanitize_lane_output() helper stripping C0/DEL control bytes. Grok quality/logic (both 0 defects) and codex logic found a residual gap: Unicode C1 control bytes (\x80-\x9f) weren't covered by the sanitizer.
3. 8bdf9698 - Surface the vendor's actual output from a lane run #40 (rebuild), commit e52edb5: extended the sanitizer to cover C1 control bytes too. Grok quality/logic (0 defects, logic run twice for redundancy after an inconclusive first attempt) and codex quality (0 defects) passed; codex logic found two Minor test-coverage gaps (sanitizer boundary bytes untested, cmd_lane print-order untested) — explicitly not production-code defects.
4. 8bdf9698 - Surface the vendor's actual output from a lane run #40, commit 7716c36: added the two missing test assertions. Grok quality/logic and codex quality/logic all returned 0 defects. Codex logic's one residual note (extend the boundary test to also assert the byte just outside each stripped range) was deliberately not chased further — it is a test-of-test nitpick on code added specifically to satisfy the prior round's finding, and the sanitizer regex's actual behavior has now been independently, manually byte-verified correct across 4 separate reviewer passes.

A "vendor output could leak a secret into --evidence" observation was raised once and independently re-assessed as a defensible non-blocking process risk by 3 further reviewer passes (no code path in the repo auto-copies result.stdout/stderr into a persisted evidence field; see DESIGN.md's "secrets do not belong in evidence" rule).

Mergeable (MERGEABLE/CLEAN), CI green (pytest), no open review threads/comments at HEAD 7716c36.

[TaprootFreakAI]

EN:
Rebased onto the current develop after #44 (one import-line conflict in main.py resolved by keeping both additions); the branch is force-pushed with --force-with-lease and the four lane verdicts are re-run on the new head.

DE:
Nach #44 auf das aktuelle develop rebased (ein Import-Zeilen-Konflikt in main.py, beide Ergänzungen behalten); der Branch wurde mit --force-with-lease gepusht, die vier Lane-Verdikte laufen auf dem neuen Head erneut.

[TaprootFreakAI]

EN:
Non-blocking observation for the author: _print_lane_result writes vendor stdout, then vendor stderr, then the STATUS line without flushing, so under redirection (> log 2>&1, or a parent process capturing the CLI) CPython's block-buffered stdout and line-buffered stderr can reorder them; flush=True on the two prints would pin the intended order.

DE:
Nicht blockierender Hinweis an den Autor: _print_lane_result schreibt Vendor-stdout, dann Vendor-stderr, dann die STATUS-Zeile ohne Flush; bei Umleitung (> log 2>&1 oder Capture durch einen Elternprozess) können CPythons block-gepuffertes stdout und zeilen-gepuffertes stderr die Reihenfolge vertauschen — flush=True auf den beiden prints würde die beabsichtigte Reihenfolge festhalten.

Details

• Location: src/agent_cli/main.py, _print_lane_result (three print() calls across two streams, no flush, no PYTHONUNBUFFERED/-u in the project).
• Unaffected: interactive tty use; capsys-based tests cannot observe it because they bypass OS buffering.
• Not part of the rebase (this comment accompanies the rebase onto develop after f07aaefa - Create the local cluster only in agent init #44); left for the author to decide.

[TaprootFreakAI]

EN:
Four lane verdicts (Grok quality, Grok logic, Codex quality, Codex logic) are approved on the rebased head 31efc83, CI is green and the branch is mergeable again. PR #40 was rebased onto develop after #44 by merging one import line, so it applies cleanly again.

DE:
Vier Lane-Verdikte (Grok Qualität, Grok Logik, Codex Qualität, Codex Logik) sind auf dem rebased Head 31efc83 approved, CI ist grün und der Branch wieder mergebar. PR #40 wurde nach #44 auf develop rebased, eine Import-Zeile zusammengeführt, und lässt sich wieder sauber anwenden.