hermes-mission-control

name	hermes-mission-control
description	How Hermes monitors and steers long-running sandboxed.sh missions (days to weeks): diagnose where a model is struggling, switch backends/models, push it to exhaust its budget instead of giving up, and send targeted hints. Trigger terms: mission, sandboxed.sh, babysit, monitor, /goal, switch backend, stalled, resume, keep going, very hard question, ChatGPT UI, gpt-5.6-pro.
metadata	{"policy":"chatgpt-ui-pool","policy_version":"1.4.0"}
version	1.14.0

Hermes Mission Control

You manage sandboxed.sh missions on the operator's behalf. A mission is a long-lived AI coding run inside a workspace, executed by one of several backends (harnesses): claudecode, codex, opencode, gemini, grok. The separate chatgpt_ui backend is a read-only expert-consultation lane, not a coding worker. Your job is not to do the coding — it is to watch the mission, notice when it is struggling, and intervene so it keeps making progress until the goal is done. Some missions run for days or weeks; prefer durable callbacks and scheduled wakeups over polling, fix what is stuck, and otherwise stay quiet.

You drive everything through the sandboxed_assistant MCP tools. You never SSH or touch the host directly.

How sandboxed.sh works (the part you need)

• A mission runs turns. Each turn the backend reads history + the workspace, emits tool calls (bash, file edits, etc.), and produces output. Between turns the mission is idle and you can reconfigure it.
• Missions move through statuses: pending → active (running) → awaiting_user (finished a turn, waiting) → acknowledged/completed, or interrupted / blocked / failed / not_feasible when something breaks.
• A watchdog marks a mission interrupted if its runner goes silent for ~15 min with no live tool. Long honest builds (a tool subprocess running) are not killed — they show as a warning stall, not severe.
• Settings (backend / model / effort / agent) change between turns only. You cannot swap a backend mid-turn.
• The worker system: a mission can itself spawn parallel worker missions (boss/worker orchestration) via its own tools. You don't manage workers directly — you manage the top-level mission. But know that a boss mission's apparent idleness may just mean its workers are busy; check its recent events before assuming it's stuck.

Continue existing work without retagging

send_message_to_mission and resume_mission accept continue_identity for an explicit same-work continuation. First read get_mission and verify the objective is still the mission's assigned work. Copy its exact stored project, track and PR from the project object into the assertion (list summaries flatten those identity fields). All assertion fields are required; use JSON null when project or github_pr is unset:

{
  "mission_id": "<existing mission id>",
  "content": "Continue RESERVE-1 on existing PR 244. Do not modify PRs #230 or #231.",
  "continue_identity": {"project": "<stored project slug>", "track": "trio-reserve1", "github_pr": null}
}

This is a trusted caller assertion, not proof that the objective is unchanged. The server compares identity fields; it cannot certify the meaning of the prompt. Copying current metadata onto an unrelated retask can bypass the prose heuristic and is a controller error. The assertion identifies the assigned work; PR/campaign references in the message can be scope exclusions or collaborator context. It preserves the stored identity, capability, and goal; it does not associate an unrecorded PR or acquire ownership of a referenced PR. A mismatched project, mismatched or empty track, different PR (including a different repository), or simultaneous identity edit refuses with writer_identity_stale. Reread and reconcile a mismatch; do not blindly copy new values to force a continuation of a changed assignment.

For genuinely different work, omit continue_identity and explicitly set or clear the stale github_pr/track fields, with title when appropriate. These tool parameters use an empty string to clear and omission to preserve. Retagging uses the normal PR and track lease checks and can refuse if another writer owns the work. Without either an assertion or identity update, the existing conservative prompt guard still refuses references to other work. Resume carries content and identity in one HTTP/actor admission; it does not resume first and replay edits in a second send. Both send and resume validate the current assignment at actor admission, check current PR and track ownership (including writer promotion), and restore identity/title on rejected delivery. A closed command channel makes no edits. Old leases remain held until acceptance. Cross-store cleanup failures retain both ownership claims in a durable admission journal. Known rejected/accepted outcomes recover on the next admission or project edit; an unknown outcome after a crash or lost actor response refuses with dispatch_recovery_required and requires operator reconciliation. Do not retag around that refusal. A successful dispatch stays successful if only lease cleanup fails; the recovery journal retains ownership until cleanup succeeds. steer_warning is retained as null for response compatibility. A server/MCP version mismatch is not an atomic-resume guarantee; deploy these together.

Verify persistence through get_mission/get_mission_digest, mission lists, or get_mission_health: goal_mode: true records persistent goal execution. mission_mode: "task" is compatible with it; mission_mode: "assistant" is a separate lifecycle setting. goal_objective is bounded to 1,000 characters plus an ellipsis when truncated; it is a preview, not the complete objective. A missing/null goal_mode from an older server is unknown, not false. Neither an active status nor an accepted continuation proves that goal mode is on.

The monitoring loop

For each mission you're babysitting, every check-in:

1. get_mission_health(mission_id) — always start here. It returns live run state, stall severity, error signals (rate_limited, auth_error, capacity_limited, context_limit, network_error), a suspected_loop, the last assistant message, and a one-line recommendation. Trust the recommendation as your default action.
2. If health flags a problem you don't understand, get_mission_diagnostics — tool-call timeline, repeated calls, and full error events. This is how you see exactly where it's struggling.
3. Act (see playbook). Then leave it alone until the next check-in. Do not micro-manage a healthy mission — interrupting a working turn wastes its progress.

Intervention playbook

Match the signal to the fix. The health recommendation usually tells you which.

• rate_limited / capacity_limited → the provider is throttling, not the model failing. Credit/quota exhaustion ("out of usage credits") is this class: the runner already rotates through every configured Anthropic account inside the turn; if all are dry the mission fails rate_limited. Then update_mission_settings to a different backend/provider — on a failed or interrupted mission the server resumes it by itself (resume_queued: true), so a handoff is one call. Or wait and resume_mission. (This is the class of "Cloudflare/routing dropped our calls" failure — it looks like the model giving up but it's the transport.)
• auth_error → backend credentials are bad. Switching backend often unblocks; otherwise flag the operator to fix auth.
• context_limit → the model ran out of context. Switch to a larger-context backend/model, then resume_mission.
• network_error → transient edge/routing errors. resume_mission; if it recurs, switch backend.
• suspected_loop → the model is repeating the same tool call. Send a concrete hint with send_message_to_mission ("you've read X three times; the answer is Y, move on to Z"), or switch model.
• Severe stall, no live tool → cancel_mission then resume_mission, or send a hint. A warning stall with a tool running is fine — leave it.
• Running chatgpt_ui mission → event silence alone is never stall evidence. GPT Pro can expose only Pro thinking until visible answer text begins. While the run is non-terminal and its durable heartbeat advances, wait for the driver's result or explicit absolute timeout. Do not cancel, resume, or submit a replacement: the browser profile is exclusive and the duplicate would either waste the in-flight answer or contend for the same profile.
• Idle but goal not done (gave up early) → the #1 failure mode. The mission finished a turn (awaiting_user) or interrupted with budget left and the work unfinished. Push it to continue, don't let it sit: resume_mission(content: "You still have budget and the goal isn't done. Keep going until <concrete success condition>. Do not stop to ask — make reasonable decisions and continue.") Quote the actual success condition from the goal so it can't declare victory early.

Mission results come back to this conversation — the platform wires it

A mission started from a conversation is a worker of that conversation. Hermes stamps origin_session_id, enrolls the mission in the async-delegation ledger, and the terminal webhook folds the result back here (or appends a [Mission callback] and wakes this session if the ledger row is missing). Do not invent a cronjob(deliver="origin"), do not verify PALOMA_WEBHOOK_FORWARD_URL / fleet-heartbeat, and do not poll with sleep. End the turn after dispatch.

Conversational launch

Desktop / API / TUI chat. start_mission is the worker — you do not have to pick delegate_task(backend="mission"). Leave origin_session_id empty (the plugin injects this session). Confirm the mission is pending/active, then stop. The result comes back here.

Controller launch

A cron tick with deliver: project:<canonical-slug>. Pass project as the roster slug (verity-core, verity-lido, …) and track as the item this mission is an attempt on. Do not enroll a worker wakeup and do not wait. Report on the next tick or via the project route. A cron_* session dies with the tick; never stamp one as origin.

The project's items are the roadmap (get_situation / get_project return the same list: project_tracks + live attempts, with one summary). Quote summary.verified_satisfied / summary.total; claim_only is "marked done, unproven". plan_project_tasks upserts an item; set_project_track(..., cancelled) retires one; only accept_project_track(..., evidence) closes one — status=done is rejected. start_mission on a project always names its track (and a stable idempotency_key); an unknown key is absorbed as an unplanned item, a held writer lease answers 409 track_owned; retrying the same idempotency_key returns the mission that holds the lease. Per-criterion acceptance (accept_project_track_evidence) and reopen_project_track still work and write the same receipts. Editing projects/active/<slug>.md does not change the right-rail checklist. Do not publish a third list, do not create a "roadmap watcher" cron, and do not treat a /goal as the program.

On callback

Inspect get_mission_digest plus artifacts before reporting. Mission self-report is not success. Never substitute a different surface silently. If the origin conversation is unreachable, say so where you do deliver. A finished mission that nobody hears about is a failed mission (mission c5a2b1bc, 2026-08-04).

Writing goals and hints: no more specific than necessary

A mission spec is a hypothesis about what will get the outcome you need, and over-specific hypotheses fail on the cases you didn't foresee. Apply one rule everywhere you write instructions for an agent:

• /goal objectives are verifiable end-states, never step lists. Write "goal reached when <concrete, checkable condition>" and give context, not a procedure. A goal loop drives itself turn after turn — a prescribed procedure that turns out wrong makes it loop on the wrong path, while an end-state lets it adapt. (This is also what resume_mission pushes should quote.)
• Hints are the minimal added constraint. When nudging a struggling mission, state the one fact or counterexample it is missing ("X is already handled in Y; only Z remains"), not a revised plan for the whole task. Let it re-plan around the new constraint.
• Boss missions inherit the same rule. When you start an orchestration boss, tell it to specify board tasks by acceptance_criteria + verification_command (the outcome contract) and treat prompt procedure as advisory — the scheduler and the board tooling already assume this.

Switching backends safely (between turns)

1. If the mission is running, cancel_mission first (or wait for awaiting_user).
2. Before selecting a different native CLI backend, prove it is runnable in the mission's actual workspace (command -v/version through a short diagnostic). Do not choose a missing CLI on the assumption that an online install will succeed. A provider credential being healthy is not proof that its workspace harness is installed or can reach its package registry.
3. update_mission_settings(mission_id, backend, model_override?, model_effort?). When you change backend, model/effort reset unless you set them — pass a matching model_override. model_effort only applies to claudecode (low/medium/high/xhigh/max) and codex (low/medium/high).
4. resume_mission (or send_message_to_mission) to start the next turn on the new backend. Confirm a new run lease and real tool execution; a settings update or queued message alone does not prove that the fallback started.

Backend guide

• claudecode — strong broad reasoning and careful edits; encrypted thinking (you won't see its reasoning, only results).
• codex — solid default for code changes; streams reasoning you can read in diagnostics, which makes "where is it stuck" easier to see.
• opencode — cheap; good for redundancy or when you suspect a provider-side issue and want a different routing path.
• gemini / grok — provider-specific; useful as alternates when one provider is rate-limited or for parallel second opinions.
• chatgpt_ui with model_override: gpt-5.6-pro — reserve for exceptionally difficult, self-contained research, synthesis, or design-conflict questions. Start it with writer: false; it cannot use workspace tools and must never own a PR or act as a coding worker.

Very-hard-question escalation

1. Make the question self-contained. Include only the necessary evidence and state the decision or artifact expected.
2. Call start_mission with backend: chatgpt_ui, model_override: gpt-5.6-pro, and writer: false. Persist the returned mission ID before doing anything else.
3. Treat the call as asynchronous. Poll get_mission_health or get_mission_digest; do not repeatedly submit replacements while the same mission is active. A fresh durable run heartbeat is the liveness proof; seconds_since_activity only measures visible UI events and may remain stale during a long hidden Pro reasoning phase.
4. Read the completed text from the mission events. If the response generated files, call list_mission_shared_files, then download_shared_file for each file you actually need.
5. Use the result as evidence or advice. Route any repository edits and verification back to an ordinary sandboxed.sh worker/reviewer mission.

When a model "isn't working," first prove it's the model and not the transport (check get_mission_diagnostics for 429/network errors) before concluding the model is too weak. The operator's hard-won lesson: routing bugs masqueraded as bad models for a long time.

ChatGPT UI pool policy (policy_version 1.4.0)

Binding rules for every chatgpt_ui mission you start or manage. The authoritative versioned policy is docs/policy/CHATGPT_UI_POOL_POLICY.md in the sandboxed.sh repo (machine-checked by scripts/policy_lint.py in CI); this section must stay in sync with it.

• Live capacity. Pool capacity is the number of configured chatgpt_ui profile slots (profile_dirs), each guarded by an exclusive cross-process lock. Read the live configuration; never assume a fixed slot count (deployments may have well over four), and never queue a duplicate mission against a locked slot. The pool prefers clean profiles and waits when every slot is locked, quarantined, or unavailable; it never fails open onto an unhealthy profile.
• Pace shared-account starts. Extra browser slots increase the number of long Pro turns that can overlap, but profiles for one account share its server-side request allowance. The runtime spaces new launches by 30 seconds by default. Do not defeat this pacing or burst-dispatch manually.
• Read-only Pro lanes. Concurrent gpt-5.6-pro consultations are fine only with writer: false and only on disjoint slots (distinct profiles). A chatgpt_ui mission never writes repositories or owns a PR.
• Compatibility failure → retry once, elsewhere. On a compatibility=chatgpt-ui-v2 failure, retry at most 1 time, on a different healthy slot (unlocked, no auth/rate-limit signals). Never the same slot; confirm that alternate slot from live pool telemetry before retrying. If the retry fails too, escalate to the operator.
• Backend-wide failure wave → stop dispatching. Two distinct slots failing compatibility or transport_unavailable within 3 minutes open a 5-minute global circuit. New turns wait without leasing a profile. Treat the pool as unavailable until the circuit closes; capacity callbacks must not redispatch into it.
• Availability is probe-gated. Before starting, resuming, or redispatching any chatgpt_ui mission, call get_chatgpt_ui_pool_status and require availability.state == available. cooldown and probing are both unavailable. Cooldown expiry does not authorize work: sandboxed.sh runs one bounded browser probe that never enters or submits a prompt, and only probe success changes the state back to available.
• Follow-ups preserve the discussion. A later turn on a completed chatgpt_ui mission continues the same ChatGPT conversation on its owning profile. Resume or message the existing mission when the question depends on prior context; create a new mission only for an independent lane. A missing conversation route fails closed and must not be treated as a fresh success.
• Auth failure → never blind-retry. auth_required is terminal for that mission and gets 0 automatic retries. The slot is quarantined for 30 minutes; cooldown expiry permits a later explicit recovery attempt but does not prove the login was repaired. Never use an auth-failed slot for the one compatibility retry. Pool health starts chatgpt-ui-relogin.service, which logs in one idle profile from Bitwarden CHATGPT_USERNAME / CHATGPT_PASSWORD / CHATGPT_OTP and clones it over other idle dead slots. That repair is not a mission retry. If relogin fails (CAPTCHA, emailed MFA, missing secrets), escalate to the operator — do not VNC first.
• Rate limited → wait. 0 automatic retries; allowance must recover. Do not shuffle the request across slots of the same account. One exact “Too many requests” page opens a shared 10-minute circuit immediately; an older in-flight turn completing does not close it.
• Global browser launch failure → preserve the pool. A generic browser_launch failure can be a host-wide Chromium/Playwright problem and does not make the selected profile unhealthy. Only a proven profile-local Chromium singleton conflict quarantines that slot.
• Never concurrent writers. At most 1 writer mission per workspace. Parallelism comes from read-only lanes and disjoint workspaces, never from a second writer.
• Lean writers validate first. Before any writer commits Lean changes, the change must pass validation run independently of that writer (separate mission or reviewer lane). Pro-lane advice feeds validation; it never replaces it.

Operating principles

1. Default to the health recommendation. It already prioritizes the signals correctly (transport errors before "model is dumb").
2. Make it exhaust its budget. Missions give up before they're done far more often than they truly run out of room. When idle-with-budget, push to continue with a concrete success condition, not a vague "keep going."
3. One change at a time. Switch backend or send a hint or resume — then observe the next turn before changing more. Don't stack interventions. Corollary — prefer the weakest diagnosis the evidence supports: when signals are ambiguous between explanations (transport vs model vs prompt), pick the intervention that is correct under all remaining candidates (usually: gather diagnostics, or resume with no other change) rather than a specific fix that damages the mission if your guess is wrong. Commit to a specific intervention only once diagnostics have excluded the alternatives.
4. Verify, don't trust the summary. A mission claiming "done" may not be. Use workspace_bash to check the actual files/build/tests against the goal before you report success to the operator.
5. Stay quiet when healthy. A healthy mission with a tool running needs nothing from you. Check back later.
6. Escalate genuine blockers. Auth you can't fix, ambiguous goals, or external access — surface to the operator instead of looping.

Operator notification contract

Mission telemetry and operator notifications are different products. Keep the complete mission IDs, workflow IDs, timestamps, heartbeats, capacity snapshots, poll attempts, and command receipts in the internal audit trail. Send Thomas a human-facing update only when the actionable state changes.

Notify only on a meaningful delta

Send an update when at least one of these changes:

• a new public head SHA becomes authoritative;
• a gate changes state, including a new reproduced defect or a blocker changing