aeon.yml

# Aeon configuration
#
# Enable skills by setting enabled: true. Multiple skills at the same time run in parallel.
# Set var: "value" to pass a default parameter to the skill.
# Set model: "claude-haiku-5-5" per skill to override the default model (cost optimization).
# Set harness: "<name>" per skill (claude, grok, codex, pi, vibe, kimi, fx, cursor,
# hermes) to run it on a different coding-agent CLI than the default Claude Code
# harness (see the top-level `harness:` key below).

skills:
  # --- Morning aggregation (7 AM UTC) ---


  # --- Mid-morning (9 AM UTC) ---
  miroshark-matchday: { enabled: false, schedule: "0 9 * * 5", var: "" } # Friday — pays $1 USDC each (x402, Finance District wallet, $4/run cap) for one MiroShark matchday sim per league (PL, Serie A, La Liga); delivers share links + reports, renders videos on local runs. var: league subset (pl,seriea,laliga) or one custom scenario
  pr-triage: { enabled: false, schedule: "30 9 * * *", var: "" } # first-touch external-PR triage (verdict + label + welcoming comment); var: owner/repo or owner/repo#N for one-shot
  pr-review: { enabled: false, schedule: "0 9 * * *", var: "" } # var: empty/owner/repo=per-PR review (posts comments) | --survey [dry-run] [owner/repo]=risk-tiered PR triage digest (folds in pr-merge)
  auto-merge: { enabled: false, schedule: "0 14 * * *" } # merge green PRs daily at 2 PM UTC; max 3 per run
  github-monitor: { enabled: false, schedule: "0 9 * * *", var: "" } # watch your repos. var: empty=combined monitor (PRs/issues/releases) | issues | releases | prs | owner/repo (folds in github-issues/github-releases/pr-tracker)
  github-trending: { enabled: false, schedule: "0 9 * * *", var: "" } # trending repos + Hugging Face Hub. var: empty=GitHub | <lang> | hf | hf:{models,datasets,spaces}

  # --- Midday (12 PM UTC) ---
  token-movers: { enabled: false, schedule: "0 12 * * *", var: "" } # market-movers scan + single-token deep report. var: empty=movers (CoinGecko) | geckoterminal[:chain]=on-chain runners | category:<slug> | <address|symbol|token:sym>=single-token report. Optional COINGECKO/ALCHEMY/XAI/BASE_RPC keys.
  onchain-monitor: { enabled: false, schedule: "0 12 * * *" }
  distribute-tokens: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand contributor rewards. var: empty/<label>=distribute (send) | plan[:week]=compute reward plan | all[:week]=plan+send | dry-run (folds in contributor-reward). BANKR_API_KEY for send.
  defi-overview: { enabled: false, schedule: "0 12 * * *", var: "" } # one-pass DeFi read: tracked-protocol positions + macro context (regime verdict, movers, yields, fees, F&G, prediction markets; refreshes memory/topics/market-context.md). var: empty=full | positions[:label] | macro | <chain/protocol focus>. Optional COINGECKO_API_KEY.
  compute-resell: { enabled: false, schedule: "15 6,14,18 * * *", var: "" } # resell free/low-cost provider compute on the Surplus Intelligence marketplace, one reactive engine per enabled provider (bankr / aws / vertex). Reads the live order book + your cost + usage, auto-lists your models, prices each offer to win routing within floor/cap guardrails. Enable a provider by setting its wallet + credential: bankr = SURPLUS_SELLER_KEY + BANKR_LLM_KEY; aws = SURPLUS_SELLER_AWS_KEY + BEDROCK_API_KEY + BEDROCK_BASE_URL var; vertex = SURPLUS_SELLER_VERTEX_KEY + VERTEX_SERVICE_ACCOUNT_JSON + VERTEX_PROJECT_ID/VERTEX_LOCATION vars. var: empty=all enabled providers reactive | <provider> | [<provider>:]{monitor|reprice|pause|resume}. Optional per-provider *_COMPUTE_RESELL_CONFIG.

  # --- Hound onchain investigation (on-demand; var = address / tx hash) ---
  investigation-report: { enabled: false, schedule: "workflow_dispatch", var: "" } # var: <token-address> [--checks=rug,contract,deployer,holders,honeypot,lp] [--depth=quick|deep] — composite investigation (folds in rug-scan/contract-audit/deployer-trace/holder-concentration/honeypot-check/lp-lock). read-only
  tx-explain: { enabled: false, schedule: "workflow_dispatch", var: "" } # var: tx hash — plain-English decode
  deploy-uni-hook: { enabled: false, schedule: "workflow_dispatch", harness: "claude", model: "claude-opus-5-5", var: "" } # generate+audit+simulate+deploy a Uniswap v4 hook + test pool on any v4 chain. var: [arm:][template:dynamic|noop|skim] [chain:base-sepolia] <brief>. Dry-run by default; arm: to broadcast (needs HOOK_DEPLOYER_PRIVATE_KEY burner). Testnet default; mainnet behind arm:+explicit chain:+HOOK_MAINNET_OK=1. Opus-pinned: freeform hook-gen is hard.
  arc-studio: { enabled: false, schedule: "workflow_dispatch", var: "" } # Drive Arc Studio on Arc testnet. var: empty/poll=attach an in-flight turn, else silent | <prompt>=one detached turn | answer:<json array or sentence>=reply to needs_input. Needs ARC_STUDIO_TOKEN. One turn at a time. On-demand until you set a cron to poll.

  monitor-polymarket: { enabled: false, schedule: "30 12 * * *", var: "" } # prediction-market monitor (Polymarket + Kalshi) — 24h moves, volume, comments, conviction alerts. var: empty=both | polymarket | kalshi | polymarket:<slug> | kalshi:<ticker>
  token-pick: { enabled: false, schedule: "0 12 * * *" }
  price-alert: { enabled: false, schedule: "*/30 * * * *", var: "" } # every 30 minutes — fire on tracked-token ATH, ±20% 1h move, or operator-set target crossing (var=target_price,target_price2,...). Silent on normal runs. Set var=dry-run to skip notify.
  narrative-tracker: { enabled: false, schedule: "30 13 * * *" } # track rising/peaking/fading narratives
  unlock-monitor: { enabled: false, schedule: "0 10 * * 1" } # weekly token unlock/vesting tracker

  # --- Afternoon (2 PM UTC) ---
  article: { enabled: false, schedule: "0 14 * * *", var: "" } # long-form article. var: empty=standard | repo:<owner/repo> | lens:<topic>; append --visual for a Replicate hero image (angles fold in repo-article/project-lens/technical-explainer)
  digest: { enabled: false, schedule: "0 14 * * *", var: "" } # topic digest across web + X + RSS. var: empty=web | <topic> | rss[:<topic>] | "<topic> +rss" (RSS feeds from memory/feeds.yml)
  you-web-search: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — high-quality web search via You.com API. var: search query/topic; runs keyless with no setup (100/day per IP), YDC_API_KEY unlocks keyed limits + livecrawl via YOUCOM_LIVECRAWL
  last30: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — cross-platform 30-day social research
  search-skill: { enabled: false, schedule: "0 14 * * *" }

  # --- Repo intelligence (3 PM UTC) ---

  # --- Late afternoon (4 PM UTC) ---
  changelog: { enabled: false, schedule: "0 16 * * *", var: "" } # changelog from recent commits/PRs. var: empty/owner/repo=in-repo | push-to:owner/website-repo=cross-repo changelog PR (needs GH_GLOBAL)
  vuln-scanner: { enabled: false, schedule: "0 16 * * 6", var: "" } # security scan + disclose. var: empty/owner/repo=scan | resubmit[:owner/repo]=PVR re-submit | disclose=queue armed emails (folds in pvr-watchlist/disclosure-emailer). RESEND_* for email.
  sc-audit: { enabled: false, schedule: "workflow_dispatch", model: "claude-opus-5-5", var: "" } # deep smart-contract audit (Slither best-effort + agentic invariant/access-control/oracle pass + optional Echidna/Medusa fuzz) of a Solidity repo OR live onchain contract. var: empty=auto-select fresh Solidity feed | owner/repo | <chain>:0x<addr> (onchain, operator-gated, never auto-filed) | fixture:<name>. Contract arm split out of vuln-scanner; routes findings through its shared disclosure machinery. ETHERSCAN/BLOCKSCOUT/RESEND_* optional. Opus-pinned: invariant reasoning is hard.
  deploy-prototype: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — build app and deploy to Vercel
  base-mcp: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — Base Account access via the Base MCP server (mcp.base.org)
  robinhood-mcp: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — Robinhood Agentic Trading via the Robinhood MCP (agent.robinhood.com, OAuth Connect in the dashboard). var: empty=portfolio report | orders[:N] | trade:<explicit one-order instruction> (only writing branch; refuses ambiguous orders)
  glim-mcp: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — live-data research via the glim.sh MCP (web/X/Reddit/GitHub/Amazon/YouTube; OAuth Connect in the dashboard, pay-per-call from glim balance). var: research question, append --deep for a wider sweep
  executor-mcp: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — run a task through the Executor Cloud tool catalog (executor.sh/mcp; one MCP endpoint proxying the operator's connected integrations, per-tool allow/approve/block policies; OAuth Connect in the dashboard). var: the task
  finance-district-mcp: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — multichain non-custodial agent wallet via the Finance District MCP (keys in a TEE, server-side spend caps; OAuth Connect in the dashboard). var: empty=daily wallet brief | best USDC yield on Base | swap 5 USDC to ETH on Base | pay <x402-url> for <data>
  taskmarket-delegate: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — browse/create/submit work on the TaskMarket agent-worker market (api.taskmarket.dev). var: browse[:query] | create "<title>" "<desc>" [reward] | submit <taskId> <message> [url]. Write actions need TASKMARKET_API_KEY + explicit operator authorization.
  autoresearch: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — set var to skill name to evolve
  create-skill: { enabled: false, schedule: "workflow_dispatch", var: "" } # one-shot skill creator — set var to a description prompt
  install-skill: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — install a community skill pack and open a PR; var: owner/repo [skill...]
  pack-submit: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — publish a local skill as a community pack + submit registry PR; var: skill-slug [--repo o/r] [--no-register] [--dry-run]

  # --- X / social (5 PM UTC) ---
  fetch-tweets: { enabled: false, schedule: "0 17 * * *", var: "" } # curate X/Twitter. var: keyword:<q> | topic:<t> | account:<handle> (empty=digest tracked accounts) | list:<ids>[|topic] | agent-buzz[:<t>]. Optional XAI_API_KEY.
  reply-maker: { enabled: false, schedule: "0 17 * * *", var: "" } # draft X replies. var: empty=discover reply-worthy tweets | @handle | <listID> | <topic> | from-logs (reply to flagged log opportunities). Optional XAI_API_KEY.

  # --- Content distribution (runs 30 min after article-generating skills) ---

  # --- Evening / meta (6 PM UTC) ---
  strategy-builder: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — draft STRATEGY.md (north-star / priorities / audience / constraints) from a brief + repo README + memory. var: bare goal text, or structured brief "repo=owner/repo | links=url1,url2 | goal=free text". Powers the dashboard Strategy → Build my strategy button. No API key needed.
  skill-health: { enabled: false, schedule: "0 18 * * *", var: "" } # self-healing health audit + analytics. var: empty=health (files/resolves issues) | <skill>=single | analytics[:hours]=fleet run/pass-rate metrics (folds in skill-analytics)
  aeon-doctor: { enabled: false, schedule: "0 15 * * 1", var: "" } # weekly Mon 15:00 UTC — static config-correctness linter (silent-failure class: unquoted schedules, dup keys, unconfigured-on-disk, requires-as-list, mode typo, mcp ${VAR}, multi-line entries). read-only; notifies only on problems. var: empty=whole config | <skill>=one skill

  self-improve: { enabled: false, schedule: "0 18 1/2 * *", var: "" } # every other day — improve the agent, or audit it. var: empty/improve[:focus]=propose+apply one fix | audit[:focus]=review what ran/failed (folds in self-review)

  action-converter: { enabled: false, schedule: "0 18 * * *" }
  skill-repair: { enabled: false, schedule: "reactive", var: "" } # reactive — auto-fix failing skills

  # --- Cost tracking (Monday 7 AM UTC, weekly) ---

  # --- Weekly (Monday only) ---
  fork-fleet: { enabled: false, schedule: "0 10 * * 1", var: "" } # fork divergence: code (unique commits/skills) + config (aeon.yml vs upstream defaults; folds in fork-digest). var: empty=both | code | config | fork=owner/repo | repo=owner/repo
  aeon-update: { enabled: false, schedule: "0 11 * * 1", var: "" } # weekly Mon 11:00 UTC — pull upstream framework changes (canon) into this instance as a PR; 3-way merge, never clobbers operator config. var: empty=sync | report (dry) | repo=owner/name | reset=fork-point
  spawn-instance: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — var: "name: purpose"
  fleet-control: { enabled: false, schedule: "0 9,15 * * *", var: "" } # managed child instances. var: empty=health check | status | dispatch <inst|*> <skill> [var=…] | scorecard (fleet-wide runs/tokens/cost/reliability; folds in fleet-scorecard)
  shiplog: { enabled: false, schedule: "0 9 * * 1" } # everything shipped SINCE THE LAST RUN (memory/state/shiplog-last.json) — cross-repo PRs/commits, security fixes merged into others' repos, star deltas, X/ecosystem traction → digest article + ready-to-post shiplog in the operator's voice. Cadence-agnostic: this schedule just sets frequency (daily/weekly/on-demand). Config: memory/products.md + memory/watched-repos.md. Optional XAI_API_KEY (X traction) + GH_GLOBAL (cross-repo reach). var: since:DATE | days:N | dry-run | owner/repo | <theme>.
  operator-scorecard: { enabled: false, schedule: "30 10 * * 1", var: "" } # Monday — weekly plain-language scorecard: agent health + community growth + economic activity (worst-of-three OK/WATCH/DEGRADED verdict). var: empty=scorecard (7d) | dry-run=no notify | N=window N hours | ops[ DATE]=end-of-day ops recap (shipped/failed/follow-up) | push[ owner/repo]=diff-deep-dive push recap (folds in ops-recap + push-recap)

  # --- Paid ads (AdManage.ai) ---
  schedule-ads: { enabled: false, schedule: "0 8 * * *", var: "" } # daily 08:00 UTC — paid ads via AdManage.ai. var: empty=schedule ad launches across Meta/TikTok/Snapchat/Pinterest/LinkedIn from config.yaml (PAUSED by default) | create=provision Meta campaigns/ad sets on-demand from config.create.yaml (PAUSED, IDs written back to state so the schedule branch can launch into them; folds in create-campaign)

  # --- Utilities ---
  auto-workflow: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand aeon.yml builder. var: <url[,url]>=analyze→enablement plan | enable:<slugs>=flip enabled:true (validate, commit, PR); enable:dry-run:<slugs> (folds in skill-enabler)

  # --- Ported from derivative instances (PR #219) ---
  # Markets & crypto
  pm-manipulation: { enabled: false, schedule: "15 13 * * *" } # daily — flag suspected PM manipulation from 3-day price/volume/comment + local-press anomalies

  # Research & content
  narrative-convergence: { enabled: false, schedule: "0 13 * * *", var: "" } # daily 13:00 UTC — cross-skill convergence detector: entities/themes surfaced by 3+ independent skill categories in 48h = high-confidence write opportunities. Category map in memory/topics/signal-categories.md. Pass var to filter.

  # Security
  vuln-tracker: { enabled: false, schedule: "30 16 * * *", var: "" } # daily lifecycle poll: vuln-scanner PR/advisory status + PVR triage + disclosure-queue aging (folds in pvr-triage/disclosure-tracker). var: empty=full | pvr | queue | prs | <GHSA-id>

  # Ideas & productivity
  idea-pipeline: { enabled: false, schedule: "0 18 * * 1" } # Monday — execution-gap audit: top 3 backlog ideas to build this week

  # Self-maintenance / meta

  # --- Ported from aeon-aaron ---
  # Infra & health
  # Markets & crypto
  picks-tracker: { enabled: false, schedule: "0 9 * * 0" } # Sunday 09:00 UTC — 30-day retrospective scoring token + prediction-market picks from logs against current prices; W/H/L + hit rate; honest record-keeping. Pairs with token-pick / monitor-polymarket.
  # Content pipeline
  soul-builder: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — build a SOUL from any mix of sources (X handle, full name, links), drafting soul/SOUL.md + STYLE.md + examples so every content skill speaks in that voice. var: bare X handle, or structured brief "x=handle | name=Full Name | links=url1,url2". Powers the dashboard Soul → Build my soul button. XAI_API_KEY optional (name/links use built-in WebSearch/WebFetch).
  # Social
  mention-radar: { enabled: false, schedule: "25 7 2/2 * *", var: "" } # every other day — external web/social mentions of the operator's projects (discovery/confusion/friction/competitor taxonomy + engagement opportunities). var: comma-separated project names; falls back to MEMORY.md / memory/topics/projects.md
  # Sports

  # --- Ported from derivative instances (2026-06-29) ---
  # Security (pairs with vuln-scanner ↑ channel 5d)
  inbox-triage: { enabled: false, schedule: "30 11 * * *" } # daily 11:30 — GitHub notification-inbox triage (maintainer replies to disclosures/advisories/mentions)

  # Dev & repo
  feature: { enabled: false, schedule: "0 10 * * 6", var: "" } # build a feature + open PR. var: watched (default) | external:<owner/repo> | dormant[:<owner/repo>], append --fix-issues (folds in external-feature/repo-revive). GH_GLOBAL for external.
  feedback-builder: { enabled: false, schedule: "0 */6 * * *", var: "" } # read agent reports from a service's /feedback endpoint, triage + cluster, build the top accepted one as a PR. Sources in memory/feedback-sources.md or var "<feedback-url> <owner/repo>"; dry-run: prefix; --max N. Optional FEEDBACK_TOKEN.

  # Research & content
  write-tweet: { enabled: false, schedule: "0 21 * * *", var: "" } # tweet studio. var: empty=drafts (10×5 tiers) | thread [topic] | remix [window]. Optional XAI_API_KEY (required for remix).
  bd-radar: { enabled: false, schedule: "20 7 * * *", var: "" } # daily 07:20 — BD radar: who's building/forking/integrating/mentioning your products → ranked leads. Products from memory/products.md. Quiet unless high-intent.

  # --- Ported from derivative instances (2026-08-08) ---
  # Security (pairs with vuln-scanner)
  hunter-22: { enabled: false, schedule: "0 10 * * *", var: "" } # daily bounty scan against ClawHunter's free discovery API (docs/ClawHunter-API.md) — capability-matched, triaged for real work vs content-farming, dedup'd against memory/topics/hunter-22-seen.json. Discovery only, no wallet, never claims/submits; audit-shaped matches with a linked GitHub repo get a one-tap "run vuln-scanner on it" button in the notification.

  # Product / ops (config: memory/products.md)

  # Self-maintenance / meta
  memory-flush: { enabled: false, schedule: "30 23 * * *" } # daily — promote important recent log entries into MEMORY.md

  # Utilities / on-demand

  send-email: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand — compose + send a 1:1 email to a named recipient via Resend (general sibling of disclosure-emailer). var: "to=… | subject=… | about=…"
  idea-forge: { enabled: false, schedule: "0 9 * * 1", var: "" } # weekly business-idea engine. var: empty/generate=timing-scored wedges | validate[:theme]=viability screen | memo[:theme]=2 evidence-backed memos
  seo-audit: { enabled: false, schedule: "0 13 * * *", var: "" } # daily on-page + technical SEO audit of every page on a site. var: comma-separated origins (empty = no target, no-op). PAGESPEED_API_KEY folds in Core Web Vitals
  posthog-errors: { enabled: false, schedule: "0 15 * * 1", var: "" } # weekly cross-project PostHog error digest. Needs PostHog MCP connected in the dashboard. var: empty=all projects/7d | Nd window | project-name substrings
  rightstack: { enabled: false, schedule: "workflow_dispatch", var: "" } # on-demand Web3 stack advisor. var: <build goal> | recommend:<goal> | workflow:<id> | compare:<a>|<b> | explain:<tool> | migrate:<package>. Read-only; pinned to rightstack@0.3.2 schema v1 JSON.

  # On-demand (run from the dashboard / `./aeon skills run`; create-prove is also dispatched by the dev-loop chain)
  competitor-monitor: { enabled: false, schedule: "workflow_dispatch", var: "" } # snapshot competitor pages (pricing, headings, CTAs, new/removed pages) and report what changed. var: comma-separated URLs
  cortx-reliability: { enabled: false, schedule: "workflow_dispatch", var: "" } # check an x402 endpoint's paid delivery rate, incidents, and latency before spending on it. var: endpoint URL (required)
  create-prove: { enabled: false, schedule: "workflow_dispatch", var: "" } # run a changed skill for real at a PR head and attach SHA-bound proof. var: owner/repo#pr@<40-char sha>
  higgsfield: { enabled: false, schedule: "workflow_dispatch", var: "" } # generate images/video via the Higgsfield MCP. var: the generation request (required)
  remotion: { enabled: false, schedule: "workflow_dispatch", var: "" } # render a short (~10s) Remotion video from a brief. var: <brief> [--orientation ...] [--theme ...]; empty = no-op
  skill-article: { enabled: false, schedule: "workflow_dispatch", var: "" } # receipts-first launch article for any skill in this instance. var: <skill-name> [--brand <handle>] [--banner]
  spend-watch: { enabled: false, schedule: "workflow_dispatch", var: "" } # cloud-cost analyst across Neon, Vercel, Railway, GitHub Actions. var: scope selector + optional arm flag
  submit-hook: { enabled: false, schedule: "workflow_dispatch", var: "" } # submit a deployed Uniswap v4 hook to the univ4-hooks registry. var: hook address; empty = latest live deploy
  video-script: { enabled: false, schedule: "workflow_dispatch", var: "" } # receipts-first, recording-ready video script. var: <source> [--minutes N]
  weekly-aeoncard: { enabled: false, schedule: "workflow_dispatch", var: "" } # weekly token-consumption recap card (SVG) from memory/token-usage.csv. var: window + delivery control

  # --- Fallback (must be last) ---
  heartbeat: { enabled: true, schedule: "0 8 * * *", var: "" } # ambient fleet-health check (default) — surfaces anything worth attention. var: empty=ambient check | <area>=ambient focused on that area | brief[:<area>]=priority brief (3 things to focus on, why now, what moved; folds in priority-brief)

# Reactive triggers — event-driven skill dispatch based on cron-state.json conditions.
# Skills with schedule: "reactive" are never cron-matched. Instead, the scheduler
# evaluates their trigger rules after processing cron skills.
#
# Format: trigger: { on: <source-skill>, when: "<condition>" }
# Conditions (evaluated by scripts/reactive_when.sh):
#   consecutive_failures >= N    - source skill failed N+ times in a row
#   success_rate <op> X          - success rate vs threshold; op is < <= > >= and X
#                                  is a 0.0-1.0 float. Only checked once the source
#                                  skill has run at least once (a never-run skill is
#                                  0.0 and would otherwise false-fire `< X`).
#   last_status = <value>        - source skill's last run ended <value> (e.g.
#                                  success), for chaining
#
# Multiple triggers: trigger: [{ on: a, when: "..." }, { on: b, when: "..." }]
# Any matching trigger fires the skill (OR logic).
reactive:
  # skill-repair:
  #   trigger:
  #     - { on: "*", when: "consecutive_failures >= 3" }
  # autoresearch:
  #   trigger:
  #     - { on: skill-health, when: "last_status = success" }

# Chains — multi-step skill pipelines with parallel execution and output passing.
# Steps run sequentially; use parallel: [...] for concurrent execution within a step.
# Skills with consume: [...] receive prior step outputs injected into their context.
#
# Format:
#   chains:
#     chain-name:
#       schedule: "cron expression"
#       on_error: fail-fast          # fail-fast (default) | continue
#       max_dispatches: 10           # optional total-dispatch cap for the chain
#       steps:
#         # Multi-key steps use flow-brace form so they stay valid YAML when uncommented.
#         - { parallel: [skill-a, skill-b] }
#         - { skill: skill-c, consume: [skill-a, skill-b] }
#         # Conditional routing on the consumed skill's Haiku score (1-5). A step
#         # with when: runs only if the condition holds; a missing score does NOT
#         # fire, so always pair branches with an unconditional fall-through.
#         # Operators: == != < > <= >= (ordering needs an integer on both sides).
#         - { skill: polish,  consume: [skill-c], when: "score > 3" }
#         - { skill: rewrite, consume: [skill-c], when: "score <= 3" }
chains:
  # Provider-neutral, operator-dispatched coding loop: the feature and review
  # skills use whichever harnesses the operator configured for those skills.
  # A human still decides whether to merge. An explicit external target is
  # required so feature can never fall back to every watched repository.
  dev-loop:
    schedule: "workflow_dispatch"
    on_error: fail-fast
    max_dispatches: 5
    steps:
      - { skill: feature, var: "$chain_target" }
      - { skill: pr-review, var: "$feature_pr", consume: [feature] }

  # morning-digest:
  #   schedule: "0 7 * * *"
  #   on_error: fail-fast
  #   steps:
  #     - { parallel: [token-movers, github-trending] }
  #     - { skill: digest, consume: [token-movers, github-trending] }

# Default model for all skills. Override per-run via workflow_dispatch.
# Options: claude-sonnet-5-5, claude-opus-5-5, claude-haiku-5-5
# (older claude-sonnet-5 / claude-opus-4-8 still accepted)
# (Grok harness model: grok-4.7, the flagship reasoning model and grok's current
#  default, run with --no-subagents in CI; grok-4.6 is also offered, and an existing
#  grok-4.5 pin is still accepted on an X-account login. grok-4.7 is the gateway
#  default too. The older coding ids (grok-build-0.1, grok-composer-2.5-fast,
#  grok-4.3) are gateway-only, reachable via XAI_API_KEY + the GROK_MODEL var.)
# The non-claude harnesses use their OWN model ids (codex openai/*, kimi moonshotai/*,
# vibe mistralai/*, pi deepseek/*) — surfaced in the dashboard picker per harness; see
# docs/harnesses.md. The value below only applies to the claude/grok harnesses.
model: claude-sonnet-5-5

# Agent harness — which coding-agent CLI runs your skills.
#   claude (default)  — Claude Code (`claude -p`); uses the AI Gateway below.
#   grok              — Grok Build (`grok`); logs in with your X account
#                       (GROK_CREDENTIALS, captured in the dashboard) or an
#                       XAI_API_KEY. Bypasses the gateway (grok has its own auth).
#   codex/pi/vibe/kimi - OpenAI Codex, Pi, Mistral Vibe, Moonshot Kimi. A single
#                       OPENROUTER_API_KEY drives all four, or each can use its own
#                       native auth (ChatGPT/Moonshot login, provider key).
#   fx                - Vercel fx (AI_GATEWAY_API_KEY; no OpenRouter fallback).
#   cursor            - Cursor CLI (CURSOR_API_KEY).
#   hermes            - Nous Research Hermes Agent (HERMES_AUTH or OpenRouter).
# All nine run through harness-adapter/run-harness (one Claude-Code-shaped
# contract). GLM is a gateway provider, not a harness. See docs/harnesses.md for
# models, auth, and verification.
# Override per-skill with `harness: "<name>"` in the skill entry, or per-run via
# the workflow_dispatch `harness` input. Default stays claude — everything
# already configured keeps working unchanged.
harness: claude

# AI Gateway — route Claude Code requests through an LLM provider.
#
# provider: auto (default) resolves at RUN TIME from whichever provider secret
# is set, in priority order — no need to re-pin when you add/remove a key:
#   claude (CLAUDE_CODE_OAUTH_TOKEN) → anthropic (ANTHROPIC_API_KEY) →
#   openrouter → bankr → usepod → venice → surplus → grok → glm → direct (fallback)
# Override the order with the repo variable GATEWAY_ORDER (space-separated).
#
# Or PIN one explicitly: direct | bankr | openrouter | usepod | venice | surplus | grok | glm |
# hivemindos. hivemindos routes Claude Code at HivemindOS Models, an
# OpenAI-compatible endpoint billed to a credit balance rather than to a provider
# account of your own: set the secret HIVEMINDOS_CREDIT_TOKEN, and optionally the
# repo variables HIVEMINDOS_MODEL (a catalog id such as anthropic/claude-sonnet-5.5;
# default inclusionai/ling-3.0-flash) and HIVEMINDOS_BASE_URL. On a model that
# honours it, HIVEMINDOS_REASONING=keep drops the reasoning tokens an unattended
# run does not need.
# grok routes Claude Code at xAI's Anthropic-compatible api.x.ai (XAI_API_KEY) —
# this is the gateway path; to run the grok CLI itself, set `harness: grok` above.
# glm routes Claude Code at Z.AI's Anthropic endpoint (GLM_API_KEY / ZAI_API_KEY).
# For Anthropic-compatible endpoints (direct), set repo variable ANTHROPIC_BASE_URL
# + secret ANTHROPIC_API_KEY. Example: https://api.deepseek.com/anthropic
# Bankr docs: https://docs.bankr.bot/llm-gateway/overview
gateway:
  provider: auto

# Output channels
channels:
  jsonrender:
    enabled: true # Set to true (or JSONRENDER_ENABLED=true as repo var) to write json-render specs

# Messaging integration
# Polls for messages every 5 minutes on Telegram/Discord/Slack.
telegram:
  enabled: true

# Observability (optional) — no config lives here; it's credential-driven.
# Set the repo secrets LANGFUSE_PUBLIC_KEY + LANGFUSE_SECRET_KEY and every
# Claude Code run streams to your Langfuse project as a trace (LLM calls,
# tokens, cost, prompts/responses). Region/toggles are repo variables:
# LANGFUSE_HOST (default EU cloud), LANGFUSE_TRACING (0 to disable),
# LANGFUSE_LOG_CONTENT (0 = metadata only). See docs/langfuse.md.