pwn 0.5.663 → 0.5.664

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: d14d73b9f076c18a1248f868c7febb2666758cfcf522493b850d3bc3e0a09ed2
-  data.tar.gz: ae83a33e35025aa612b82d48d0ccab99bea2e5a4b20d1d2ffff81189233739b1
+  metadata.gz: 7e88bcfaae1b1859df50e16528f8460a31cab0d7a8168206b9f48c3541c00b88
+  data.tar.gz: c143b22d04b09c1c30a7f6b2f7bfe17cf3dfb8812484db4e42daf9565ba12668
 SHA512:
-  metadata.gz: 9d39cbfff72ee966e15990f6d2329edc0fc3671bdc7b5c7f11ebb29587d4e93f64a447f9e5bb6fde2dac84160e1f4764a9143001866c85b730681fd903541340
-  data.tar.gz: d8f6e3f3e50605efde9a8db3fa96daf4a385d599b98a5fe787af62a1d7ed9c0232da987f7320cb9a07001208d0a21f96767530a943d819b14c43841fe0d867d8
+  metadata.gz: 06d533eb382b1b6509684e15db4dcb8e9bdd24d9ac628e1f07c090bfd7f099e7a608ef61e2c87969e54f245736fb6a10f170042cd650f1dbd023816c788e662c
+  data.tar.gz: c3243b7fd533a14ad0825546b60b88b465ae8bdc5b476f001dd00cad98788d73173b578c981467a6ec664f1e60815d4bd2f3efa6a6b659e17b85e6ab6d4ad09d

data/lib/pwn/ai/agent/curriculum.rb

@@ -292,13 +292,21 @@ module PWN
               end
             end
             if commit && defined?(Learning)
+              # P29 — keep verdict tag score-aligned (same as auto_introspect).
+              sc = v[:score].to_f
+              verd = if defined?(Learning) && Learning.respond_to?(:verdict_for_score, true)
+                       Learning.send(:verdict_for_score, score: sc).to_s
+                     elsif sc >= 0.6 then 'solved'
+                     elsif sc >= 0.3 then 'partial'
+                     else 'wrong'
+                     end
               Learning.note_outcome(
                 task: req[0, 120],
-                success: v[:score].to_f >= 0.6,
-                score: v[:score],
-                details: "offline_judge #{v[:verdict]}(#{v[:score]}) #{v[:rationale]}",
+                success: sc >= 0.6,
+                score: sc,
+                details: "offline_judge #{verd}(#{sc.round(2)}) #{v[:rationale]}",
                 session_id: sid,
-                tags: %w[offline_judge auto]
+                tags: ['offline_judge', 'auto', verd]
               )
             end
             scored << { session_id: sid, score: v[:score], verdict: v[:verdict] }

data/lib/pwn/ai/agent/learning.rb

@@ -160,13 +160,33 @@ module PWN
 
         public_class_method def self.to_context(opts = {})
           limit = opts[:limit] || 5
-          rows  = outcomes(limit: limit)
-          fails = outcomes(limit: 200, success: false).first(limit)
+          # Fetch a wider window so prefer_primary_tasks can drop critic/red_team
+          # envelope rows (REQUEST:/GOAL: prefixes) without starving the block.
+          rows  = prefer_primary_tasks(rows: outcomes(limit: limit * 4)).first(limit)
+          fails = prefer_primary_tasks(rows: outcomes(limit: 200, success: false))
+          # Do not mirror the same ids under both headings — that doubled the
+          # failure signal and made RECENT OUTCOMES == RECENT FAILURES when the
+          # last N attempts all failed (the injected block looked "stuck").
+          row_ids = rows.map { |r| r[:id] }.compact
+          fails = fails.reject { |r| row_ids.include?(r[:id]) }.first(limit)
           return '' if rows.empty? && fails.empty?
 
           fmt = lambda do |r|
-            flag = r[:success] ? '✓' : '✗'
-            "  #{flag} #{r[:task].to_s[0, 100]} (#{r[:timestamp]})"
+            flag = case r[:success]
+                   when true then '✓'
+                   when 'soft', :soft then '∼'
+                   else '✗'
+                   end
+            score = r.key?(:score) ? format('%.2f', r[:score].to_f) : '-'
+            task  = display_task(task: r[:task])
+            line  = "  #{flag} [#{score}] #{task} (#{r[:timestamp]})"
+            # Surface a one-line cause crumb so the agent can actually learn
+            # from failures instead of only seeing that they failed.
+            if r[:success] != true
+              crumb = cause_crumb(details: r[:details])
+              line += "\n      cause: #{crumb}" unless crumb.empty?
+            end
+            line
           end
           s   = stats
           jm  = s[:judge_mean]
@@ -468,7 +488,12 @@ module PWN
           v = Reward.judge(request: opts[:request], final: opts[:final], session_id: session_id, proxy_ok: proxy_ok) if defined?(Reward)
           v ||= { score: proxy_ok ? 1.0 : 0.0, success: proxy_ok, verdict: proxy_ok ? :solved : :wrong }
           v[:score] = [v[:score], 0.3].min if crit[:verdict] == :flaw
-          ok = v[:score] >= 0.6
+          # P29 — critic floor used to leave stale verdict=:solved at score=0.3,
+          # producing learning.jsonl rows tagged "solved" with success=false
+          # (116+ rows). Always resync verdict/success from the final score.
+          v[:verdict] = verdict_for_score(score: v[:score])
+          v[:success] = v[:score].to_f >= 0.6
+          ok = v[:success]
 
           # W1 pending user_correction pair
           pend = Thread.current[:pwn_pending_pref]
@@ -515,11 +540,14 @@ module PWN
           outcome_tags = ['auto', 'loop', v[:verdict].to_s]
           outcome_tags << plan_cov[:tag] if plan_cov && plan_cov[:tag]
           outcome_tags << "plan_cover=#{plan_cov[:score]}" if plan_cov && plan_cov[:total].to_i.positive?
+          # P29 — persist the bare user ask (strip REQUEST:/GOAL: envelopes at write time)
+          task_txt = display_task(task: opts[:request].to_s)
+          task_txt = opts[:request].to_s[0, 100] if task_txt.empty?
           note_outcome(
-            task: opts[:request].to_s[0, 120],
+            task: task_txt,
             success: ok,
             score: v[:score],
-            details: "#{v[:verdict]}(#{v[:score].round(2)}) #{v[:rationale]} | #{opts[:final].to_s[0, 200]}",
+            details: "#{v[:verdict]}(#{v[:score].to_f.round(2)}) #{v[:rationale]} | #{opts[:final].to_s[0, 200]}",
             session_id: session_id,
             tags: outcome_tags
           )
@@ -794,6 +822,103 @@ module PWN
           nil
         end
 
+        # P29 — map score → verdict with the same thresholds as Reward.judge.
+        private_class_method def self.verdict_for_score(opts = {})
+          s = opts[:score].to_f
+          return :solved if s >= 0.6
+          return :partial if s >= 0.3
+
+          :wrong
+        end
+
+        # Strip critic/red_team envelope prefixes so the injected block shows
+        # the human ask, not "REQUEST:\n…\nANSWER:" / "GOAL:\n…\nPLAN:".
+        private_class_method def self.display_task(opts = {})
+          t = opts[:task].to_s.gsub(/\s+/, ' ').strip
+          if t.match?(/\AREQUEST:\s*/i)
+            body = t.sub(/\AREQUEST:\s*/i, '')
+            body = body.split(/\bANSWER:\s*/i, 2).first.to_s
+            t = body.strip
+          elsif t.match?(/\AGOAL:\s*/i)
+            body = t.sub(/\AGOAL:\s*/i, '')
+            body = body.split(/\bPLAN:\s*/i, 2).first.to_s
+            t = body.strip
+          end
+          t[0, 100]
+        end
+
+        # Prefer bare user goals over REQUEST:/GOAL: swarm envelopes when both
+        # describe the same underlying attempt (offline_judge + critic sessions).
+        private_class_method def self.prefer_primary_tasks(opts = {})
+          rows = Array(opts[:rows])
+          return rows if rows.empty?
+
+          scored = rows.map do |r|
+            t = r[:task].to_s
+            envelope = t.match?(/\A\s*(REQUEST:|GOAL:)/i) ? 1 : 0
+            # Higher is better: bare task first, then newer (rows already newest-first)
+            [r, -envelope]
+          end
+          # stable: keep relative order within same envelope rank
+          scored.sort_by.with_index { |(_, rank), i| [rank, i] }.map(&:first)
+        end
+
+        private_class_method def self.cause_crumb(opts = {})
+          d = opts[:details].to_s.gsub(/\s+/, ' ').strip
+          return '' if d.empty?
+
+          # Prefer explicit FLAW / CORRECTED crumbs; else verdict(score) head.
+          if (m = d.match(/\bFLAW:\s*(.+)\z/i)) || (m = d.match(/\bFLAW:\s*([^|]+)/i))
+            return m[1].to_s.strip[0, 120]
+          end
+          if (m = d.match(/\bCORRECTED:\s*(.+)\z/i))
+            return "corrected: #{m[1].to_s.strip[0, 100]}"
+          end
+
+          d[0, 120]
+        end
+
+        # One-shot / on-load repair: rewrite tags+details where verdict label
+        # disagrees with score (solved @ 0.3 etc.). Safe to call repeatedly.
+        public_class_method def self.reconcile_verdict_tags!(opts = {})
+          return { repaired: 0 } unless File.exist?(LEARNING_FILE)
+
+          dry = opts[:dry_run] ? true : false
+          repaired = 0
+          lines = File.readlines(LEARNING_FILE)
+          out = lines.map do |l|
+            r = JSON.parse(l, symbolize_names: true)
+            score = r.key?(:score) ? r[:score].to_f : nil
+            next l if score.nil?
+
+            want = verdict_for_score(score: score).to_s
+            tags = Array(r[:tags]).map(&:to_s)
+            stale = tags & %w[solved partial wrong unknown]
+            next l if stale.empty? || stale.include?(want)
+
+            repaired += 1
+            next l if dry
+
+            cleaned = tags - %w[solved partial wrong unknown]
+            cleaned << want
+            r[:tags] = cleaned
+            # Fix leading "solved(0.3)" style details head when present
+            det = r[:details].to_s
+            r[:details] = det.sub(
+              /\A(solved|partial|wrong|unknown)\(\d+(?:\.\d+)?\)/i,
+              "#{want}(#{format('%.2f', score)})"
+            )
+            r[:success] = (score >= 0.6) if [true, false].include?(r[:success])
+            "#{JSON.generate(r)}\n"
+          rescue StandardError
+            l
+          end
+          File.write(LEARNING_FILE, out.join) if !dry && repaired.positive?
+          { repaired: repaired, dry_run: dry }
+        rescue StandardError => e
+          { repaired: 0, error: "#{e.class}: #{e.message}" }
+        end
+
         private_class_method def self.auto_introspect_enabled?
           return false unless defined?(PWN::Env) && PWN::Env.is_a?(Hash)
 

data/lib/pwn/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module PWN
-  VERSION = '0.5.663'
+  VERSION = '0.5.664'
 end

data/spec/lib/pwn/ai/agent/learning_spec.rb

@@ -86,4 +86,65 @@ describe PWN::AI::Agent::Learning do
   ensure
     FileUtils.rm_rf(tmp) if defined?(tmp) && tmp
   end
+
+  it 'P29 resyncs verdict after critic floor and strips request envelopes in to_context' do
+    tmp = Dir.mktmpdir
+    stub_const('PWN::AI::Agent::Learning::LEARNING_FILE', File.join(tmp, 'learning.jsonl'))
+    stub_const('PWN::Memory::MEMORY_FILE', File.join(tmp, 'memory.json'))
+
+    PWN::AI::Agent::Learning.reset
+
+    # Critic-capped solved-at-0.3 pattern (pre-P29 poison)
+    PWN::AI::Agent::Learning.note_outcome(
+      task: "REQUEST:\nwhats the bottom line?\n\nANSWER:\nfake bottom line",
+      success: false,
+      score: 0.3,
+      details: 'solved(0.3) heuristic overlap=0.75 ratio=1.0 | fake bottom line',
+      tags: %w[auto loop solved plan_cover_high]
+    )
+    PWN::AI::Agent::Learning.note_outcome(
+      task: 'real bare goal that failed',
+      success: false,
+      score: 0.22,
+      details: 'wrong(0.22) heuristic overlap=0.06 | blew up',
+      tags: %w[auto loop wrong]
+    )
+    PWN::AI::Agent::Learning.note_outcome(
+      task: 'successful scan',
+      success: true,
+      score: 0.9,
+      details: 'solved(0.9) ok',
+      tags: %w[auto loop solved]
+    )
+
+    # reconcile should flip solved→partial on the 0.3 row
+    rep = PWN::AI::Agent::Learning.reconcile_verdict_tags!
+    expect(rep[:repaired]).to be >= 1
+
+    rows = PWN::AI::Agent::Learning.outcomes(limit: 20)
+    fixed = rows.find { |r| (r[:score].to_f - 0.3).abs < 0.001 }
+    expect(Array(fixed[:tags])).to include('partial')
+    expect(Array(fixed[:tags])).not_to include('solved')
+    expect(fixed[:details]).to match(/\Apartial\(0\.30?\)/)
+
+    ctx = PWN::AI::Agent::Learning.to_context(limit: 5)
+    # envelope stripped
+    expect(ctx).to include('whats the bottom line?')
+    expect(ctx).not_to match(/REQUEST:\nwhats/)
+    # score visible
+    expect(ctx).to match(/\[0\.30\]/)
+    # RECENT OUTCOMES and RECENT FAILURES must not be identical dumps
+    # (dedupe by id) — with 1 success + 2 fails, outcomes shows mix
+    expect(ctx).to include('successful scan')
+    expect(ctx).to include('cause:')
+  ensure
+    FileUtils.rm_rf(tmp) if defined?(tmp) && tmp
+  end
+
+  it 'P29 verdict_for_score thresholds match Reward.judge' do
+    expect(PWN::AI::Agent::Learning.send(:verdict_for_score, score: 0.6)).to eq(:solved)
+    expect(PWN::AI::Agent::Learning.send(:verdict_for_score, score: 0.59)).to eq(:partial)
+    expect(PWN::AI::Agent::Learning.send(:verdict_for_score, score: 0.3)).to eq(:partial)
+    expect(PWN::AI::Agent::Learning.send(:verdict_for_score, score: 0.29)).to eq(:wrong)
+  end
 end

data/third_party/pwn_rdoc.jsonl

@@ -260,10 +260,12 @@
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.auto_introspect Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.auto_introspect`: Supported Method Parameters\n\nPWN::AI::Agent::Learning.auto_introspect(\n\nsession_id: 'required - id of the just-completed session',\nrequest: 'optional - original user request (for outcome logging)',\nfinal: 'optional - final assistant answer (for outcome logging)'\n\n)\n\nCalled by Loop.run when PWN::Env[:agent] is truthy. Never raises — learning must not break the primary loop.\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.auto_introspect_enabled? Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.auto_introspect_enabled?`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.build_skill_from_session Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.build_skill_from_session`: "}]}
+{"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.cause_crumb Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.cause_crumb`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.checkable_claim? Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.checkable_claim?`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.compress_exemplar Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.compress_exemplar`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.compress_finetune_trace Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.compress_finetune_trace`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.consolidate Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.consolidate`: Supported Method Parameters\n\nremoved = PWN::AI::Agent::Learning.consolidate(\n\nmax_entries: 'optional - hard cap on PWN::Memory size (default MAX_MEMORY_ENTRIES)'\n\n)\n\nDeduplicates near-identical lesson values and prunes the oldest entries once the cap is exceeded so the injected MEMORY block stays high-signal.\n"}]}
+{"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.display_task Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.display_task`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.distill_skill Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.distill_skill`: Supported Method Parameters\n\nskill = PWN::AI::Agent::Learning.distill_skill(\n\nname: 'required - snake_case name for the new skill',\nsession_id: 'optional - PWN::Sessions id to mine (uses its transcript)',\ncontent: 'optional - explicit markdown body; overrides transcript mining',\nreferences: 'optional - Array of reference URLs / CWE / CVE / ATT&CK ids'\n\n)\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.exemplars_for Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.exemplars_for`: Supported Method Parameters\n\nmsgs = PWN::AI::Agent::Learning.exemplars_for(\n\nrequest: 'required - current user request',\nlimit: 'optional - max exemplar traces to return (default 1)',\nmax_msgs: 'optional - cap on messages per exemplar (default 6)'\n\n)\n\nRetrieval-augmented BEHAVIOUR: keyword-matches request against prior successful outcomes in learning.jsonl, loads the matching session, and compresses its (user, tool, assistant) trace into a short few-shot exemplar Loop.run splices between system and user. Local models are dramatically better with 1 concrete example than with 25 abstract lessons.\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.export_finetune Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.export_finetune`: "}]}
@@ -277,9 +279,11 @@
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.merge_cluster Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.merge_cluster`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.note_outcome Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.note_outcome`: Supported Method Parameters\n\nentry = PWN::AI::Agent::Learning.note_outcome(\n\ntask: 'required - short description of what was attempted',\nsuccess: 'required - Boolean, did the attempt achieve its goal',\ndetails: 'optional - free-form notes / error / evidence',\nsession_id: 'optional - PWN::Sessions id this outcome belongs to',\ntags: 'optional - Array of String labels for later retrieval'\n\n)\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.outcomes Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.outcomes`: Supported Method Parameters\n\nrows = PWN::AI::Agent::Learning.outcomes(\n\nlimit: 'optional - max entries returned newest-first (default 50)',\nsuccess: 'optional - filter by Boolean outcome',\ntag: 'optional - filter by tag substring'\n\n)\n"}]}
+{"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.prefer_primary_tasks Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.prefer_primary_tasks`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.process_sop_text? Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.process_sop_text?`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.promote_process_lesson Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.promote_process_lesson`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.purge_noise Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.purge_noise`: Supported Method Parameters\n\nPWN::AI::Agent::Learning.purge_noise\n\nOne-shot GC of the pre-R1 garbage: drops every PWN::Memory entry matching the old ‘SUCCESS: <req> — <final>` / `Avoid repeating failure pattern from <tool>: {“success”:true` shapes. Run once after upgrading; subsequent writes never produce these.\n"}]}
+{"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.reconcile_verdict_tags! Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.reconcile_verdict_tags!`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.recover_predicted_from_session Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.recover_predicted_from_session`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.reflect Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.reflect`: Supported Method Parameters\n\nreport = PWN::AI::Agent::Learning.reflect(\n\nsession_id: 'required - PWN::Sessions id to analyse',\ndry_run: 'optional - when true, do not write to Memory/Skills (default false)'\n\n)\n\nUses PWN::AI::Agent::Reflect (when available) to LLM-summarise the session into structured lessons. Falls back to a heuristic extractor when module_reflection is disabled so learning never stops.\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.reset Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.reset`: Supported Method Parameters\n\nPWN::AI::Agent::Learning.reset\n"}]}
@@ -289,6 +293,7 @@
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.stats Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.stats`: Supported Method Parameters\n\nstats = PWN::AI::Agent::Learning.stats\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.to_context Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.to_context`: Supported Method Parameters\n\nctx = PWN::AI::Agent::Learning.to_context(\n\nlimit: 'optional - number of recent outcomes to surface (default 5)'\n\n)\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.transcript_text Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.transcript_text`: "}]}
+{"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.verdict_for_score Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.verdict_for_score`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.active_engine Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.active_engine`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.agent_flag Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.agent_flag`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.append_session Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.append_session`: "}]}

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: pwn
 version: !ruby/object:Gem::Version
-  version: 0.5.663
+  version: 0.5.664
 platform: ruby
 authors:
 - 0day Inc.