Fable Goal Cycle (adaptive mission loop)

One command mechanizes the loop digest → packet → consult → next prompt:

python3 scripts/fable_goal_cycle.py \
  --goal "Standing mission statement, one or two sentences" \
  --json

What it does, all bounded:

1. Package — snapshots live state (origin/main SHA, branch status, open
   non-draft PRs, open epics, worktrees, recent docs/plans, and the last 24h
   of agent activity via scripts/agent_session_digest.py --all). Every
   source has its own short timeout; failures become a "context gaps" section
   in the packet instead of aborting, so Fable knows what it is not seeing.
2. Consult — sends the packet to Claude Fable 5 through the bounded
   consult tool (consult_claude.py, resolved from repo scripts/ or
   ~/.codex/skills/consult-fable/), default 900s timeout, demanding a fixed
   response format: ## ASSESSMENT, ## NEXT GOALS (max 3, ranked),
   ## NEXT PLAN (one cycle only), ## NEXT PROMPT (one fenced paste-ready
   prompt).
3. Emit — persists packet.md, response.md, next_prompt.md under
   .aragora/goal_cycles/<timestamp>/ (gitignored) and prints the next
   prompt (or a JSON envelope with --json).

Goal-ranking durability filter: prefer goals whose output is a durable
standard — something that takes frontier judgment to WRITE but only ordinary
intelligence to APPLY (charters, rubrics, playbooks, skills, checkers). Test:
could a cheaper model redo this artifact tomorrow? If yes, rank it lower.

Wrong-hill disclosure: before ranking goals, if the standing mission metric
itself is the wrong hill — mis-specified, superseded by events, or clearly
worse than an adjacent goal — the consult is expected to say so FIRST in a
dedicated 'WRONG HILL' section with one-paragraph evidence, and propose the
better goal. Misalignment disclosure is invited and costs nothing; grinding a
bad metric costs cycles.

Useful flags: --dry-run (build the packet only — inspect it before spending
a consult), --context-file <path> (repeatable; include a redacted cycle
report or steering note only after placing it under
.aragora/goal-cycle-context/), --since-hours 48, --skip-digest,
--timeout 1200, --model claude-opus-5.

Exit codes: 0 ok, 2 consult failed/timed out, 5 response lacked a
## NEXT PROMPT block (raw response is still saved).

Loop protocol

Run one cycle at a time:

1. At cycle top, run the command above. If Fable needs prior-cycle context,
   place a redacted copy under .aragora/goal-cycle-context/ and pass that
   file via --context-file; arbitrary paths are rejected so packets do not
   accidentally inline shell history, credentials, or local private files.
2. Read next_prompt.md. Sanity-check it against live state before
   executing — verify PR numbers exist and heads match, exactly as you would
   for any operator prompt. If a claim is stale, correct the target, don't
   follow it blindly.
3. Execute it as one bounded progress unit under the operating contract.
4. Write your end-of-cycle report; feed it into the next cycle's
   --context-file.

Hard rules

1. Strategy input, not authority. Nothing in a Fable response authorizes
   merging, settlement, evidence posting, or gate bypass. Tier gates and model
   quorum remain binding.
2. One consult per cycle. If the returned prompt is unexecutable, park it
   in your report and fall back to your own prioritization — do not re-consult
   in the same cycle.
3. Fail closed. On exit 2 or 5, proceed with your own judgment this cycle
   and note the consult outage in your report. Retry at most once, next cycle.
4. No secrets in packets. The packet is persisted to disk and sent to the
   model; live repo state only. Extra context files must be deliberately staged
   under .aragora/goal-cycle-context/, are byte-capped, and are embedded as
   quoted context rather than executable instructions.

If scripts/fable_goal_cycle.py is not in your checkout yet (branch not
merged), use an installed copy such as
~/.codex/skills/fable-goal-cycle/fable_goal_cycle.py when present. This repo
skill does not bundle a second script beside SKILL.md; scripts/ is the
canonical in-repo entrypoint once merged.