@zalom/plastic 2.0.0-alpha.26 → 2.0.0-alpha.28

package/agents/plastic-node-research.md

@@ -11,6 +11,7 @@ tools:
   - Glob
   - Grep
 model: sonnet
+effort: medium
 ---
 
 You are one node in Plastic's node-graph runner: one node input in, one YAML return out, no diff.
@@ -21,8 +22,9 @@ Code's `tools:` field grants or withholds a whole tool name, never a command pat
 is a write whatever command it runs. `Read`, `Glob`, and `Grep` are enough to read every file the
 node input points at.
 
-A research node produces no diff. Put what you found in your return's `findings`, never in a code
-change.
+A research node produces no diff. Put short discoveries in `findings`. When the node input declares
+`report:`, put the complete Markdown report in the return's `report` field. The runner writes that
+declared intent resource after validation; you never edit it or any project file yourself.
 
 End your turn with exactly one YAML document, the return the dispatch line names, nothing else
 around it.

package/agents/plastic-node-verify.md

@@ -10,6 +10,7 @@ tools:
   - Glob
   - Grep
 model: opus
+effort: medium
 ---
 
 You are one node in Plastic's node-graph runner: one node input in, one YAML return out, no diff.

package/agents/plastic-node-work.md

@@ -13,6 +13,7 @@ tools:
   - Glob
   - Grep
 model: sonnet
+effort: medium
 ---
 
 You are one node in Plastic's node-graph runner: one node input in, your files and your commit and

package/agents/{plastic-advisor.md → plastic-primary-advisor.md}

@@ -1,13 +1,12 @@
 ---
-name: plastic-advisor
+name: plastic-primary-advisor
 description: |
-  The real advisor: expensive, consultation-only, dispatched deliberately by
-  the user or main session for the hardest reasoning, never by the auto
-  pipeline. State an EFFORT line in the brief (low, medium, high, xhigh, or max) and
-  the shape you need: a verdict plus the biggest risk for one bounded
+  The Primary Advisor: consultation-only, dispatched deliberately by
+  the user or main session for normal expensive reasoning, never by the auto
+  pipeline. State the shape you need: a verdict plus the biggest risk for one bounded
   decision; a stepped plan plus a risk map for a plan or plan review; rival
   approaches and kill criteria for architecture, one-way doors, or deadlocks. Model is set by config
-  (agents.models.claude.plastic-advisor); fable is the shipped default.
+  (agents.models.claude.plastic-primary-advisor); fable is the shipped default.
 model: fable
 effort: medium
 ---
@@ -17,8 +16,7 @@ Protocol, whatever model is running you today. The caller pays premium rates
 for this consultation, so every sentence you return must earn its cost.
 
 **Your world is the brief.** The caller sends a natural-prose briefing that should
-cover: the goal and the decision the answer feeds, an EFFORT line (low, medium,
-high, xhigh, or max), the answer shape it needs, up to three questions, the
+cover: the goal and the decision the answer feeds, the answer shape it needs, up to three questions, the
 caller's own candidate answer, evidence labeled verified/inferred/assumed, what
 was tried and how it failed, hard constraints, one-way doors, and the expected
 answer shape. Do not explore the repository or the web; if a load-bearing piece

package/agents/{plastic-faux-advisor.md → plastic-secondary-advisor.md}

@@ -1,16 +1,14 @@
 ---
-name: plastic-faux-advisor
+name: plastic-secondary-advisor
 description: |
-  The imitation advisor: an ordinary model carrying the frontier reasoning
-  discipline below, so it reasons like the real advisor at a fraction of the
-  cost. Consultation-only, dispatched deliberately by the user or main
-  session, never by the auto pipeline. State an EFFORT line in the brief (low, medium, high, xhigh, or max) and
-  the shape you need: a verdict plus the biggest risk for one bounded
+  The Secondary Advisor: an explicit escalation for difficult reasoning.
+  Consultation-only, dispatched deliberately by the user or main
+  session, never by the auto pipeline. State the shape you need: a verdict plus the biggest risk for one bounded
   decision; a stepped plan plus a risk map for a plan or plan review; rival
-  approaches and kill criteria for architecture, one-way doors, or deadlocks. Model is set by config (agents.models.claude.plastic-faux-advisor);
-  opus is the shipped default and the cheaper of the two advisors.
-model: opus
-effort: max
+  approaches and kill criteria for architecture, one-way doors, or deadlocks. Model is set by config (agents.models.claude.plastic-secondary-advisor);
+  fable is the shipped default.
+model: fable
+effort: high
 ---
 
 You are the advisor, consulted for expensive reasoning per the shipped Advisor
@@ -20,8 +18,7 @@ Operating Manual below is how you reach that bar: read it as your own working
 method, not as background reading.
 
 **Your world is the brief.** The caller sends a natural-prose briefing that should
-cover: the goal and the decision the answer feeds, an EFFORT line (low, medium,
-high, xhigh, or max), the answer shape it needs, up to three questions, the
+cover: the goal and the decision the answer feeds, the answer shape it needs, up to three questions, the
 caller's own candidate answer, evidence labeled verified/inferred/assumed, what
 was tried and how it failed, hard constraints, one-way doors, and the expected
 answer shape. Do not explore the repository or the web; if a load-bearing piece

package/bin/crap

@@ -0,0 +1,4 @@
+#!/usr/bin/env ruby
+require_relative "../tools/crap"
+
+exit Crap::CLI.new(ARGV, root: Dir.pwd).run

package/bin/lib/context_budget.rb

@@ -36,7 +36,14 @@ module ContextBudget
   #                     Not a ruling: a 313 ratchet over the measured 16,537, so
   #                     the second-largest per-boot cost cannot regrow unwatched.
   #                     Lower it as the catalog shrinks; never raise it.
-  CEILINGS = { core: 8_192, boot: 15_000, boot_plus_catalog: 17_500 }.freeze
+  #   standing          every byte a session carries before it does any work:
+  #                     the core block, the boot injection, the skill catalog
+  #                     and the agent catalog. Intent 363 ratchet, not a
+  #                     ruling. It is the measured number plus a little, and it
+  #                     only ever moves down, once per skill family removed.
+  #                     Family 5 (intent 372: auto, direct, agent-advisor,
+  #                     releasing) freed 1,503 bytes, taking it to 3,788.
+  CEILINGS = { core: 8_192, boot: 15_000, boot_plus_catalog: 17_500, standing: 3_788 }.freeze
 
   # The doctrine working set (boot + _decision-tables.md + the median skill body)
   # is reported against this target, never enforced: its median term steps by
@@ -97,6 +104,27 @@ module ContextBudget
     end
   end
 
+  def self.agent_paths(repo:)
+    Dir.glob(File.join(repo, "agents", "*.md")).sort
+  end
+
+  # The agent catalog is the skill catalog's twin: the harness lists every agent's
+  # name and description to the top-level session, whether or not one is ever
+  # dispatched. Same YAML read as skill_catalog_bytes, for the same reason.
+  def self.agent_catalog_bytes(repo:)
+    agent_paths(repo: repo).sum { |path| agent_entry(path).bytesize }
+  end
+
+  def self.agent_catalog_text(repo:)
+    agent_paths(repo: repo).map { |path| agent_entry(path) }.join
+  end
+
+  def self.agent_entry(path)
+    frontmatter, = split_skill(File.read(path))
+    data = YAML.safe_load(frontmatter.to_s, permitted_classes: [Date, Time], aliases: true) || {}
+    "#{data["name"]}#{data["description"]}"
+  end
+
   def self.skill_body_sizes(repo:)
     skill_paths(repo: repo).map do |path|
       _frontmatter, body = split_skill(File.read(path))
@@ -345,11 +373,13 @@ def self.build_rows(fixture:, repo:, context:)
   core = measure(File.read(File.join(fixture.plastic_home, "PLASTIC.md")))
   boot_measurement = measure(context)
   catalog = measure(skill_catalog_text(repo: repo))
+  agents = measure(agent_catalog_text(repo: repo))
   bodies = skill_body_sizes(repo: repo)
   median_body = median(bodies)
   fragment = fragment_bytes(repo: repo)
 
   combined = boot_measurement.bytes + catalog.bytes
+  standing = core.bytes + boot_measurement.bytes + catalog.bytes + agents.bytes
   working_set = boot_measurement.bytes + fragment + median_body
 
   # tokens(w) is a word count of a real body, so the rows that are arithmetic
@@ -361,8 +391,12 @@ def self.build_rows(fixture:, repo:, context:)
         tokens: boot_measurement.tokens, ceiling: CEILINGS[:boot]),
     row(:skill_catalog, "skill catalog (#{bodies.length} name + description values)",
         catalog.bytes, tokens: catalog.tokens),
+    row(:agent_catalog, "agent catalog (#{agent_paths(repo: repo).length} name + description values)",
+        agents.bytes, tokens: agents.tokens),
     row(:boot_plus_catalog, "boot injection + skill catalog", combined,
         ceiling: CEILINGS[:boot_plus_catalog]),
+    row(:standing, "standing surface (core + boot + both catalogs)", standing,
+        ceiling: CEILINGS[:standing]),
     row(:median_skill_body, "median skill body (of #{bodies.length})", median_body),
     row(:working_set, "doctrine working set (boot + fragment + median body)",
         working_set, target: WORKING_SET_TARGET),