pwn 0.5.779 → 0.5.781

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

@@ -2195,7 +2195,8 @@ module PWN
           pairs.each do |pair|
             tool = pair.find { |m| m[:role].to_s == 'tool' }
             prev = collapsed.last&.find { |m| m[:role].to_s == 'tool' }
-            next if tool && prev && tool[:content].to_s.strip == prev[:content].to_s.strip
+            # Compare raw bytes: existing history may contain malformed tool output.
+            next if tool && prev && tool[:content].to_s.b.strip == prev[:content].to_s.b.strip
 
             collapsed << pair
           end
@@ -2241,8 +2242,10 @@ module PWN
           FileUtils.mkdir_p(dir)
           path = File.join(dir, "#{digest[0, 12]}.txt")
           File.binwrite(path, text) unless File.file?(path)
-          head = text.byteslice(0, 2_048).to_s
-          tail = text.bytesize > 3_072 ? text.byteslice(-1_024, 1_024).to_s : ''
+          # Keep the artifact/digest byte-exact; only the displayed excerpt is text.
+          # Byte slicing can bisect a multibyte character at either boundary.
+          head = text.byteslice(0, 2_048).to_s.dup.force_encoding(Encoding::UTF_8).scrub
+          tail = text.bytesize > 3_072 ? text.byteslice(-1_024, 1_024).to_s.dup.force_encoding(Encoding::UTF_8).scrub : ''
           mid = tail.empty? ? '' : "\n...\n#{tail}"
           "#{head}#{mid}\n[compacted path=#{path} sha256=#{digest} bytes=#{text.bytesize} ref=#{path}]"
         rescue StandardError

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

@@ -87,7 +87,9 @@ module PWN
             File.open(backup, File::WRONLY | File::CREAT | File::EXCL, 0o600) { |file| file.write(baseline) }
           end
           replace(path: live, bytes: candidate)
-          { promoted: true, live_path: live, rollback_path: backup, baseline_sha256: prior_digest,
+          { promoted: true, receipt_version: 1, change_kind: 'routing_policy',
+            acceptance_sha256: Digest::SHA256.hexdigest(JSON.generate(deterministic(value: replayed))),
+            live_path: live, rollback_path: backup, baseline_sha256: prior_digest,
             candidate_sha256: candidate_digest, replayed_seeds: replayed.map { |r| r[:seed] } }
         rescue StandardError => e
           { promoted: false, reason: e.message }
@@ -98,6 +100,8 @@ module PWN
           raise ArgumentError, 'stop all policy writers and set quiescent: true' unless opts[:quiescent] == true
 
           receipt = opts.fetch(:receipt)
+          raise ArgumentError, 'successful versioned policy promotion receipt required' unless receipt[:promoted] == true && receipt[:receipt_version] == 1 && receipt[:change_kind] == 'routing_policy' && receipt[:acceptance_sha256].to_s.match?(/\A[0-9a-f]{64}\z/)
+
           live = File.expand_path(opts.fetch(:live_path))
           prior_digest = receipt.fetch(:baseline_sha256)
           raise ArgumentError, 'invalid rollback digest' unless prior_digest.is_a?(String) && prior_digest.match?(/\A[0-9a-f]{64}\z/)

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

@@ -180,6 +180,9 @@ module PWN
             v[:confidence] = [v[:confidence].to_f, ground[:confidence].to_f].max if ground[:confidence]
           end
 
+          # Preserve the evaluator's original completion claim before verification
+          # can lower its verdict. This is diagnostic, never acceptance evidence.
+          v[:claimed_completion] = v[:verdict].to_s == 'solved'
           v[:judge_score] = v[:score].to_f
           v[:score_quality] = v[:score].to_f
           v[:verification] = if opts[:verification].is_a?(Hash)
@@ -272,6 +275,12 @@ module PWN
                    :pass
                  end
           end
+          v[:verified_completion] = vv.nil? ? nil : vv == :pass
+          v[:false_completion] = if v[:claimed_completion] == true
+                                   vv.nil? ? nil : vv == :fail
+                                 elsif v[:claimed_completion] == false
+                                   false
+                                 end
           v[:verifier_verdict] = vv
           v[:confidence] = 1.0 if vv
           if vv == :fail || v.dig(:grounded, :verdict).to_s == 'refuted'

data/lib/pwn/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module PWN
-  VERSION = '0.5.779'
+  VERSION = '0.5.781'
 end

data/spec/fixtures/harness/completion_regressions.json

@@ -0,0 +1,10 @@
+{
+  "schema": "pwn-completion-regressions-v1",
+  "split": "regression",
+  "provenance": "Source-reproduced HarnessEvaluation.grade false-success accounting defects: vacuous empty checks, nonzero process exit despite correct bytes, and resource cap despite correct bytes. These are public regression stimuli, not held-out learning evidence.",
+  "cases": [
+    {"id": "empty-acceptance-claim", "checks": [], "result": {"stdout": "DONE", "exit_status": 0, "resource_cap": false}, "outcome": "FAIL"},
+    {"id": "failed-process-claim", "checks": [{"path": "result.txt", "bytes": "correct"}], "result": {"stdout": "DONE", "exit_status": 1, "resource_cap": false}, "outcome": "FAIL"},
+    {"id": "capped-process-claim", "checks": [{"path": "result.txt", "bytes": "correct"}], "result": {"stdout": "DONE", "exit_status": 0, "resource_cap": true}, "outcome": "INCOMPLETE"}
+  ]
+}

data/spec/fixtures/harness/completion_regressions.json.sha256

@@ -0,0 +1 @@
+9368d0704929f9ed4ab448f8bfb0958d19de89cb09a934d20d2cb0a36a22ee7b

data/spec/integration/harness_evaluation_spec.rb

@@ -23,6 +23,19 @@ RSpec.describe 'offline harness evaluation' do
     expect { HarnessEvaluation.validate_parity!(pinned, pinned.merge('provider' => 'different')) }.to raise_error(ArgumentError)
   end
 
+  it 'counts claims without checks, nonzero exits, and capped claims as false completion' do
+    Dir.mktmpdir do |root|
+      corpus = HarnessExperiments.locked_json(File.join(HarnessEvaluation::ROOT, 'spec/fixtures/harness/completion_regressions.json'))
+      expect(corpus.fetch('split')).to eq('regression')
+      File.write(File.join(root, 'result.txt'), 'correct')
+      corpus.fetch('cases').each do |failure|
+        task = { 'checks' => failure.fetch('checks'), 'success_marker' => 'DONE' }
+        result = failure.fetch('result').transform_keys(&:to_sym)
+        expect(HarnessEvaluation.grade(task, root, result)).to include(outcome: failure.fetch('outcome'), false_success: true), failure.fetch('id')
+      end
+    end
+  end
+
   it 'grades independent bytes and counts observed repeated identities, not tool attempts' do
     Dir.mktmpdir do |root|
       File.write(File.join(root, 'result.txt'), 'wrong')

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

@@ -5,6 +5,28 @@ require 'tmpdir'
 require 'fileutils'
 
 describe PWN::AI::Agent::Learning do
+  describe 'failure regression tasks' do
+    include_context 'pwn tmp sandbox'
+
+    it 'exports a reviewed replay recipe only for a persisted independently verified failure' do
+      request = 'Write result.txt'
+      checks = [{ requirement: request, kind: :file, path: 'result.txt', expected: 'correct' }]
+      verification = PWN::AI::Agent::Verification.run(request: request, requirements: [request], checks: checks, root: @tmp)
+      row = described_class.note_outcome(task: request, outcome: { score: 0.9, source: :llm_orm, claimed_completion: true, verification: verification })
+      recipe = described_class.regression_task(outcome_id: row[:id], request: request, checks: checks)
+      expect(recipe).to include(schema: 'pwn-failure-regression-v1', split: 'regression', source_outcome_id: row[:id])
+      expect(recipe[:source_verification_sha256]).to eq(Digest::SHA256.hexdigest(JSON.generate(row[:verification])))
+      replay = PWN::AI::Agent::Verification.run(recipe[:contract].merge(root: @tmp))
+      expect(replay[:status]).to eq(:fail)
+      File.write(File.join(@tmp, 'result.txt'), 'correct')
+      expect(PWN::AI::Agent::Verification.run(recipe[:contract].merge(root: @tmp))[:status]).to eq(:pass)
+      expect { described_class.regression_task(outcome_id: row[:id], request: 'other', checks: checks) }.to raise_error(ArgumentError, /request/)
+      expect { described_class.regression_task(outcome_id: row[:id], request: request, checks: []) }.to raise_error(ArgumentError, /coverage/)
+      expect { described_class.regression_task(outcome_id: row[:id], request: request, checks: checks.map { |c| c.merge(path: '../escape') }) }.to raise_error(ArgumentError, /relative/)
+      expect { described_class.regression_task(outcome_id: 'missing', request: request, checks: checks) }.to raise_error(ArgumentError, /verified failure/)
+    end
+  end
+
   describe 'evidence failure safety' do
     before do
       home = Dir.mktmpdir

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

@@ -1304,6 +1304,45 @@ describe PWN::AI::Agent::Loop do # rubocop:disable Metrics/BlockLength
       expect(src).not_to match(/compact_history!\(messages: messages\) if local/)
     end
 
+    it 'keeps UTF-8 previews valid across repeated history compaction' do
+      Dir.mktmpdir('pwn-history-encoding') do |dir|
+        allow(Dir).to receive(:home).and_return(dir)
+        messages = [{ role: 'system', content: 'sys' }, { role: 'user', content: 'goal' }]
+        8.times do |n|
+          messages << { role: 'assistant', content: '', tool_calls: [{ id: "c#{n}" }] }
+          messages << { role: 'tool', tool_call_id: "c#{n}", content: "a#{'é' * 3000}#{n}" }
+        end
+        2.times do
+          expect { described_class.send(:compact_history!, messages: messages) }.not_to output(/compact_history swallowed/).to_stderr
+          expect(messages.all? { |msg| msg[:content].valid_encoding? }).to be(true)
+        end
+      end
+    end
+
+    it 'preserves raw spilled bytes while rendering malformed and binary-tagged text safely' do
+      Dir.mktmpdir('pwn-history-binary') do |dir|
+        allow(Dir).to receive(:home).and_return(dir)
+        ["a#{'é' * 3000}".b, "a\xff#{'é' * 3000}"].each do |text|
+          preview = described_class.send(:spill_tool_body, text: text)
+          expect(preview.encoding).to eq(Encoding::UTF_8)
+          expect(preview.valid_encoding?).to be(true)
+          expect(preview).to include('é', '[compacted path=')
+          path = preview[/\[compacted path=(.*?) sha256=/, 1]
+          expect(File.binread(path)).to eq(text.b)
+        end
+      end
+    end
+
+    it 'compacts existing malformed tool history without swallowing encoding errors' do
+      messages = [{ role: 'system', content: 'sys' }, { role: 'user', content: 'goal' }]
+      8.times do |n|
+        messages << { role: 'assistant', tool_calls: [{ id: "c#{n}" }] }
+        messages << { role: 'tool', tool_call_id: "c#{n}", content: "invalid\xff#{n}" }
+      end
+      expect { described_class.send(:compact_history!, messages: messages) }.not_to output(/compact_history swallowed/).to_stderr
+      expect(messages.count { |msg| msg[:role] == 'tool' }).to eq(6)
+    end
+
     it 'shrinks tool bodies when history is fat' do
       msgs = [
         { role: 'system', content: 'sys' },

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

@@ -100,7 +100,12 @@ describe 'PWN::AI::Agent::PolicyEvaluation' do
     File.unlink(live)
     File.write(live, File.read(baseline))
     result = evaluator.promote(enabled: true, quiescent: true, baseline: baseline, candidate: candidate, reports: reports, live_path: live)
-    expect(result).to include(promoted: true)
+    expect(result).to include(promoted: true, receipt_version: 1, change_kind: 'routing_policy')
+    expect(result[:acceptance_sha256]).to match(/\A[0-9a-f]{64}\z/)
+    [result.merge(promoted: false), result.merge(receipt_version: 2), result.except(:receipt_version), result.merge(change_kind: 'skill'), result.merge(acceptance_sha256: 'invalid')].each do |invalid|
+      expect(evaluator.rollback(enabled: true, quiescent: true, live_path: live, receipt: invalid)).to include(rolled_back: false, reason: /receipt/)
+      expect(File.binread(live)).to eq(File.binread(candidate))
+    end
     expect(File.binread(live)).to eq(File.binread(candidate))
     expect(File.binread(result[:rollback_path])).to eq(File.binread(baseline))
     expect(result[:replayed_seeds]).to eq([0, 1])

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

@@ -4,6 +4,27 @@ require 'spec_helper'
 require 'tmpdir'
 
 describe PWN::AI::Agent::Reward do
+  describe 'completion accounting' do
+    include_context 'pwn tmp sandbox'
+
+    it 'preserves a refuted completion claim through resolution and durable learning' do
+      request = 'Write answer.txt'
+      report = PWN::AI::Agent::Verification.run(request: request, requirements: [request], root: @tmp,
+                                                checks: [{ requirement: request, kind: :file, path: 'answer.txt', expected: 'correct' }])
+      result = described_class.resolve_outcome(outcome: { score: 0.95, source: :llm_orm, claimed_completion: true, verification: report })
+      expect(result).to include(verified_completion: false, false_completion: true)
+      expect(described_class.resolve_outcome(outcome: result)).to eq(result)
+      row = PWN::AI::Agent::Learning.note_outcome(task: request, outcome: result)
+      expect(row).to include(claimed_completion: true, verified_completion: false, false_completion: true)
+      expect(PWN::AI::Agent::Learning.stats).to include(completion_claims: 1, false_completions: 1, verified_completions: 0)
+    end
+
+    it 'keeps unverified claims unknown rather than labeling them false or verified' do
+      result = described_class.resolve_outcome(outcome: { score: 0.95, source: :llm_orm, claimed_completion: true })
+      expect(result).to include(verified_completion: nil, false_completion: nil)
+    end
+  end
+
   describe 'labeled transcript calibration' do
     include_context 'pwn tmp sandbox'
 

data/spec/support/harness/evaluate.rb

@@ -62,7 +62,7 @@ module HarnessEvaluation
                 passed ? 'PASS' : 'FAIL'
               end
     { outcome: outcome, checks: checks, tool_timeouts: timeouts,
-      false_success: claimed && !checks.all? { |row| row[:passed] },
+      false_success: claimed && outcome != 'PASS',
       repeated_effects: effects.tally.values.sum { |count| count - 1 }, observed_effects: effects }
   end
 

data/third_party/pwn_rdoc.jsonl

@@ -357,6 +357,7 @@
 {"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.regression_task Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.regression_task`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.requeue_conflicted Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.requeue_conflicted`: "}]}
 {"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"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.rl_skill_notes Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.rl_skill_notes`: "}]}
@@ -368,7 +369,7 @@
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.should_gc_stores? Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.should_gc_stores?`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.skill_body_without_frontmatter Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.skill_body_without_frontmatter`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Learning.skills_dir Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.skills_dir`: "}]}
-{"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.stats Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.stats`: "}]}
 {"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.update_skill Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Learning.update_skill`: "}]}

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: pwn
 version: !ruby/object:Gem::Version
-  version: 0.5.779
+  version: 0.5.781
 platform: ruby
 authors:
 - 0day Inc.
@@ -505,28 +505,28 @@ dependencies:
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 1.8.0
+        version: 1.8.1
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 1.8.0
+        version: 1.8.1
 - !ruby/object:Gem::Dependency
   name: meshtastic
   requirement: !ruby/object:Gem::Requirement
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 0.0.188
+        version: 0.0.189
   type: :development
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 0.0.188
+        version: 0.0.189
 - !ruby/object:Gem::Dependency
   name: metasm
   requirement: !ruby/object:Gem::Requirement
@@ -953,14 +953,14 @@ dependencies:
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 1.92.0
+        version: 1.92.1
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 1.92.0
+        version: 1.92.1
 - !ruby/object:Gem::Dependency
   name: rubocop-rake
   requirement: !ruby/object:Gem::Requirement
@@ -1447,6 +1447,7 @@ files:
 - documentation/Metasploit.md
 - documentation/Mistakes.md
 - documentation/NmapIt.md
+- documentation/Operability.md
 - documentation/PWN.png
 - documentation/Persistence.md
 - documentation/Plugin-Degradation.md
@@ -6579,6 +6580,8 @@ files:
 - spec/documentation/prompt_version_placeholder_spec.rb
 - spec/fixtures/harness/adapters.json
 - spec/fixtures/harness/adapters.json.sha256
+- spec/fixtures/harness/completion_regressions.json
+- spec/fixtures/harness/completion_regressions.json.sha256
 - spec/fixtures/harness/experiment_gold.json
 - spec/fixtures/harness/experiment_gold.json.sha256
 - spec/fixtures/harness/experiment_tasks.json