research-cycle

name	research-cycle
description	Drive the new 6-stage research pipeline (init→survey→axiom→propose→dispatch→loop→review) via Hermes cron. Replaces deprecated auto-research and research-host cron drivers. Reads ledgers/<tag>/TODO.md, picks next stage, invokes claude -p with the appropriate /research-* command. Halts on SOTA/budget/BLOCK/human-flag.
version	1.0.0
author	Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["research","cron","autonomous","hermes-native"],"related_skills":["claude-code-router","codex-academic","gemini-academic"]}}

research-cycle — Hermes-native research orchestrator

Replaces the deprecated:

• ~/.hermes/scripts/deprecated/auto-research.py (drove /exp-run, deprecated)
• ~/.hermes/scripts/deprecated/research-host-cron.py (drove /research-host, deprecated)

Drives the NEW 7-command pipeline (commands at ~/.claude/commands/research-*.md):

research-init <tag>           ← Stage 0: human-only, sets up sprint
research-survey               ← Stage 1: cron picks up here
research-axiom                ← Stage 2
research-propose              ← Stage 3
research-dispatch             ← Stage 4 (loop start)
research-loop "<idea>"        ← Stage 5
research-review               ← Stage 6 (loop end)
                              ← back to dispatch for next cycle

Procedure

Step 1 — pre-checks (Python; the script does this)

python3 ~/.hermes/scripts/research-cycle.py --dry-run

Returns next stage + halt status. If halted (SOTA / budget / BLOCK / human flag), stop here.

Step 2 — actual cron tick

python3 ~/.hermes/scripts/research-cycle.py

The script:

1. Finds latest active sprint (~/research/rgmare-lite/ledgers/*/sprint.yaml mtime)
2. Reads TODO.md to determine next stage ([ ] items, then loop phase detection)
3. Halt-checks: human flag (~/.hermes/shared/.research-halt), SOTA, budget, last-BLOCK
4. Invokes claude -p --model claude-sonnet-4-6 "/research-<stage> [args]" with stdin=DEVNULL
5. For /research-loop, extracts the rank-1 idea from latest dispatch/cycle-N/ranked.jsonl
6. For /research-review, extracts cycle number from latest dispatch
7. Captures stdout/stderr, parses verdict, logs to ~/.hermes/shared/research-cycle-log.jsonl
8. Writes status JSON to ~/.hermes/shared/.last-research-cycle-status

Step 3 — Hermes wakes summary agent (only if exit_code == 0)

The script emits {"wakeAgent": <bool>, "stage": ..., "verdict": ...} on stdout. Hermes uses this to decide whether to spawn a summary agent.

Halt conditions (priority order)

Order	Condition	Action
1	~/.hermes/shared/.research-halt exists	yield; user must rm it to resume
2	No active sprint (ledgers/*/sprint.yaml empty)	yield; user must run /research-init <tag>
3	results.tsv has keep row meeting target_value from sprint.yaml	yield; SOTA reached
4	Cycle count ≥ budget_cycles from sprint.yaml	yield; budget exhausted
5	Last research-review verdict was BLOCK and not yet resolved	yield; user must inspect
6	claude -p timed out (>90 min default)	log error; do not retry this tick

Permission boundary for claude-code

When the cron driver invokes claude -p, it uses:

working_dir: ~/research/rgmare-lite
permission_mode: dangerously-skip-permissions   # cron, no human-in-loop
extra_constraints:
  - "Operate ONLY within ~/research/rgmare-lite/ — do not touch other workspaces"
  - "If the next /research-* command's HARD GATE fails, exit 1; do NOT improvise"
  - "Respect the constitution at contracts/constitution.v0.1.0.yaml"
  - "Never ask 'should I continue' — the human is asleep; loop until halt condition fires"

Cron registration (NOT auto-applied — user must approve)

To register, the user adds a job entry to ~/.hermes/cron/jobs.json. This skill does NOT auto-edit jobs.json. Suggested entry:

{
  "id": "<auto-generated>",
  "name": "Research Cycle",
  "prompt": "The script already drove one stage of the research pipeline. Output above is the result. Summarize what was done in 2-3 sentences. If verdict=BLOCK, surface it. Save sprint progress to memory.",
  "skills": ["research-cycle"],
  "skill": "research-cycle",
  "model": null,
  "provider": null,
  "base_url": null,
  "script": "research-cycle.py",
  "schedule": {
    "kind": "interval",
    "minutes": 90,
    "display": "every 90m"
  },
  "schedule_display": "every 90m",
  "repeat": {"times": null, "completed": 0},
  "enabled": false,
  "state": "scheduled",
  "deliver": "local"
}

Recommended cadence: 90 minutes (gives /research-loop time to complete + reviewer time, but quickly enough that a missed cycle isn't catastrophic).

Set enabled: true only after user confirms a sprint is active.

Status outputs

• ~/.hermes/shared/.last-research-cycle-status — single JSON snapshot of last tick
• ~/.hermes/shared/research-cycle-log.jsonl — append-only log per tick
• ~/.hermes/shared/.research-halt (file presence) — human halt flag

Anti-patterns

• ❌ Do NOT enable this cron without an active sprint (will fire every 90m doing nothing — noisy)
• ❌ Do NOT remove the halt flag check — that's the user's emergency brake
• ❌ Do NOT increase CLAUDE_TIMEOUT beyond 90 min without reason — research-loop is bounded by autoresearch's 5-min training; longer means something hung
• ❌ Do NOT auto-resolve a BLOCK verdict — human decides whether the BLOCK is fixable or the sprint is dead
• ❌ Do NOT touch auto-research.py or research-host-cron.py — they're in deprecated/ for archive purposes only