config

agentic-loop:config — feature flags

All optional loop features live behind explicit, default-off flags in ONE file: .agentic/config.json in the current project. This skill is the only writer of that file. Invocations:

/agentic-loop:config                          → status (all features)
/agentic-loop:config <feature> on|off         → toggle
/agentic-loop:config <feature> status         → one feature
/agentic-loop:config grill deep on|off        → sub-toggle: deep interviews
                                                 for large/domain-heavy ideas
/agentic-loop:config render [--tty]           → observability report (delegates
                                                 to scripts/observe_render.sh)

The features

Feature	What it enables	Third-party dependency
observability	Unified JSONL event log (.agentic/observability/) capturing every subagent, shim call, headless iteration and factory transition; renderable as an HTML/tty tree	none
minimize	The code-minimization decision ladder is injected into build-stage worker briefs (smallest sufficient diff)	ponytail (rules content; plugin optional)
grill	A relentless pre-planning interview runs before loop-planner decomposes ambiguous or high-stakes requests	none — native behavior in the spec skill (interview protocol inspired by grill-with-docs in mattpocock/skills; nothing to install)
grill deep	large/new-domain ideas escalate to a deep interview: grill-with-docs (/grilling + /domain-modeling, emits glossary + ADRs into factory/specs/<id>/) when Matt Pocock's collection is installed; otherwise native grilling with the question cap lifted	mattpocock-skills (optional — enhances; native fallback always works)
guards	Clean-code + test quality-gate criteria are added to the reviewer's blind-review checklist	guard-skills (criteria content; plugin optional)
summarize	The report renderer fills summary-less nodes via local Ollama (free)	ollama running locally

Config file shape

{
  "observability": { "enabled": true, "all_agents": false },
  "minimize":      { "enabled": false, "agent_judgment": false },
  "grill":         { "enabled": false, "deep": false, "agent_judgment": false },
  "guards":        { "enabled": false },
  "summarize":     { "enabled": false },
  "_meta":         { "updated": "<iso date>" }
}

• agent_judgment: true permits the ORCHESTRATOR to enable that feature per-task on its own judgment. Every judgment toggle MUST be recorded:

  ./scripts/observe.sh emit feature_toggle \
    '{"detail":{"feature":"minimize","scope":"task 003","reason":"mechanical bulk edit","decided_by":"agent"}}'
  

  Never judgment-enable anything metered in unattended (factory/headless) stages — same rule as metered escalation.

• install_declined: true on a feature records that the user declined installing its third-party dependency: never re-offer, run without it.

Steps

1. Read .agentic/config.json if it exists (create .agentic/ and the file with all features enabled: false on first write; never create it for a pure status read — report "no config, all features off" instead).
2. status: print a short table — feature, enabled, dependency state (installed / missing / declined). Detect dependencies cheaply: observability and base grill have none (native — nothing to detect); grill with deep: true → grep claude plugin list for mattpocock-skills (missing = native-uncapped fallback, not an error); summarize → curl -sS --max-time 2 http://localhost:11434/api/tags; minimize/guards → grep claude plugin list for the plugin name (ponytail) or, for skills-dir installs, the @skills-dir entry (guard-skills).
3. <feature> on (and grill deep on, which sets .grill.deep instead of .enabled and also sets .grill.enabled = true — deep implies on): a. Base grill has no dependency — it's native behavior in the spec skill; just enable it (skip straight to step b). For a feature (or grill deep) that DOES have a third-party dependency that is MISSING and install_declined is not set: show the install command and ask the user install now / enable anyway (applies once installed) / cancel. If they decline installing but still want the feature, set install_declined: true alongside the flag (for grill deep, the spec skill then runs its native grilling with the question cap lifted — the feature still works). The install command depends on how the dependency ships — verify the repo layout, don't assume:
   • marketplace plugins (have .claude-plugin/marketplace.json) — minimize→ponytail, grill deep→Matt Pocock's collection:

     claude plugin marketplace add DietrichGebert/ponytail
     claude plugin install ponytail@ponytail
     
     claude plugin marketplace add mattpocock/skills
     claude plugin install mattpocock-skills@mattpocock
     

   • bare skills (a SKILL.md or a skills/ dir, NO marketplace.json — plugin marketplace add FAILS on these) — guards→amElnagdy/guard-skills: clone and symlink into the skills directory instead:

     git clone https://github.com/amElnagdy/guard-skills /path/to/guard-skills
     ln -s /path/to/guard-skills ~/.claude/skills/guard-skills
     

   Note: minimize/guards fold the dependency's content (ladder rules / guard criteria) into our own briefs and reviewer — the third-party install is optional, the feature works from baked-in content either way. b. Update the file with jq (preserve unknown keys): jq '.<feature>.enabled = true | ._meta.updated = "<today>"' … c. If observability was just enabled, mention: events start with the NEXT session/shim call; AGENTIC_OBSERVE=1 forces it for one-off runs; .agentic/ is gitignored so nothing lands in git.
4. <feature> off: set enabled: false. Do not delete the entry (history of install_declined etc. must survive). grill deep off sets only .grill.deep = false — base grill stays as it was.
5. render: run ./scripts/observe_render.sh (add --tty if asked for a terminal view; --summarize only if the summarize flag is enabled). If the scripts are not scaffolded into this project yet, run them from the plugin root instead (this file lives at <plugin-root>/skills/config/SKILL.md).

Keep output terse: a status table or a one-line confirmation. Never enable a feature the user didn't name. If asked about a feature not in the table, say it doesn't exist rather than inventing one.