Task — engineering-spec@1

GUIDE gains the sandboxed-agent runbook section

doneTASK-IMP-097
module improvement · class product · priority p3 · created 2026-07-17 · shipped 2026-07-17
depends on none · blocks none

TASK-IMP-097: GUIDE gains the sandboxed-agent runbook section

Summary

Three governed runs re-learned the same sandbox facts by hitting them. Write them down where consumers read: a short "Running CyberOS under sandboxed agents" section in the GUIDE source - command caps and background-process death, the local-clone pattern for build/test with a local-ref push back to the mounted repo, manual hook-obligation replay with --no-verify plus recorded evidence, and mount unlink/permission quirks - with a one-line cross-reference from ship-tasks.

Problem

Every fact in that list cost real time to discover and none of it is discoverable today except by reading incident addenda in task folders.

Proposed Solution

One section in tools/install/docs/index.md (which builds into the payload's GUIDE.md), written as a runbook: symptom, cause, working pattern. Explicit that the local-clone push is a local ref move - the no-push policy stays intact. ship-tasks.md gains one cross-reference line pointing constrained environments at the GUIDE section (prose pointer, no version bump - TASK-IMP-099 carries this round's bump). test_full_sdp_payload.sh gains a grep gate that the built GUIDE carries the section.

Alternatives Considered

Success Metrics

Scope

In scope: the GUIDE section, the ship-tasks cross-reference line, the payload grep gate.

Out of scope / Non-Goals

Dependencies

AI Authorship Disclosure

1. Description (normative)

2. Acceptance criteria

3. Edge cases

Audit

TASK-IMP-097-sandbox-runbook-note spec audit - audit_rubric@2.0 (machine floor + judgment)

Machine floor: task-lint clean (the batch's two TRACE-001 misses - uncited suite-landing clauses in 093/094 - were caught by the lint on first run and fixed before audit; the governed loop working as built). Judgment families: metrics carry baseline/target/deadline grounded in run evidence or the recorded 2026-07-16 research; alternatives distinct with real rejection reasons; scope subsections present; ops-verified ACs carry explicit rationales; COND-004 truthful; cone-sharing (install.sh trio, build.sh pair, ship-tasks.md pair) declared in Dependencies per §11a so the batch schedule serializes correctly.

ISSUE ISS-001 (QA-004, wontfix-info): consumer-repo and single-prose-line ACs are ops-verified with recorded-evidence rationales (same accepted pattern as TASK-IMP-086/087/090).

SUMMARY verdict: pass issues_open: 0 issues_human: 0 next_action: ship

§gate-log

Populated during implementation (ship-tasks testing phase).