Handoff: Hermes weekly Codex multi-agent watch

Continuation target

Current state

• Done Local, Battlebox, and XPS13 run Codex CLI 0.146.0.
• Done All three configs passed codex --strict-config doctor --summary --ascii.
• XPS13 uses /home/miro/.codex/config.toml and global instructions at /home/miro/.codex/AGENTS.md.
• The current child-context interface defaults to fork_turns="all" when omitted; later parent turns are not continuously synchronized into an existing child.
• Codex 0.146.0 improved inherited-history pagination, temporary forks, and preservation across forks, but did not document automatic selective context packing.

[agents]
enabled = true
max_concurrent_threads_per_session = 4
default_subagent_model = "gpt-5.6-terra"
default_subagent_reasoning_effort = "medium"

The global policy says to use fork_turns="none" by default, a small recent-turn window such as "3" when necessary, and "all" only exceptionally.

Ordered implementation

1. Inspect the existing Hermes installation, scheduler, notification route, and conventions. Prefer Hermes's native scheduled-run mechanism; use a user-level systemd timer only if Hermes has no scheduler. Do not invent a second orchestration stack.
2. Create a durable state file in the existing Hermes state/config hierarchy. Store the last checked timestamp, installed and latest Codex versions, source fingerprints or relevant release identifiers, and the last known answers to the watch questions below.
3. Run the monitor once immediately, persist the baseline, and deliver its first report through Hermes's normal user-facing channel.
4. Schedule it weekly at a low-noise local time. Record the exact schedule, command/prompt, state path, log path, and how to run or disable it manually.
5. Verify the scheduled unit/job is active and execute one manual dry run through the same entrypoint used by the scheduler.

Sources to check

• Installed version: codex --version
• Published version: npm view @openai/codex version
• Official Codex changelog
• Official subagent documentation
• The current Codex manual/tool schema available to the running agent, especially spawn_agent and fork_turns.
• Official OpenAI sources only for factual claims. Clearly label any inference.

Questions the weekly run must answer

1. Did the stable Codex version change, and do its release notes mention agents, subagents, multi-agent V2, delegation, forks, inherited history, context, thread lineage, or concurrency?
2. Did the default or allowed values of fork_turns change?
3. Is there now a documented global config setting for child-context inheritance?
4. Is selective context packing, summarization, continuous parent-child synchronization, or another context-sharing mechanism now documented?
5. Did proactive delegation behavior, built-in roles, model selection, reasoning effort, concurrency, sandbox inheritance, or permission inheritance change?
6. Do any changes justify reconsidering the current fork_turns="none" operational policy?

Reporting contract

Produce a compact report with: Verdict, Material changes, What remains unchanged, Recommended action, and direct official-source URLs. Distinguish product changes from model-quality impressions. If nothing material changed, send only a one-line heartbeat with the checked version and date. Never claim a schema/default changed based only on release-note wording.

Verification and acceptance

• The initial baseline report is delivered successfully.
• The weekly job is listed as enabled/active by the actual scheduler.
• A manual invocation through the scheduled entrypoint exits successfully and updates the state timestamp.
• A simulated unchanged run produces the compact heartbeat rather than a full repeated report.
• A simulated changed fingerprint produces a delta report without editing ~/.codex/config.toml or ~/.codex/AGENTS.md.
• The handoff back to Miro includes exact commands to inspect, run now, disable, and remove the monitor.

Risks and open questions

• Decide locally Discover Hermes's actual scheduling and notification facilities before choosing cron or systemd.
• The public changelog may lag or omit schema details. Cross-check the current manual/tool schema before declaring defaults changed.
• Do not auto-install a newly published Codex version. Report it and let Miro choose when to update.
• Do not publish credentials, auth files, raw session history, or private prompts in reports or logs.
• If official sources are unavailable, report the source failure and retain the previous baseline; do not infer “no change.”

Suggested skills

• openai-docs for current Codex documentation and manual checks.
• battlebox-ssh only if cross-machine verification is later requested; XPS13 is the primary Hermes host.