Module: PWN::AI::Agent::Mistakes

Defined in:
lib/pwn/ai/agent/mistakes.rb

Overview

PWN::AI::Agent::Mistakes is the negative-feedback half of the pwn-ai learning loop. Where Learning records WHAT WORKED and Metrics records HOW OFTEN a tool worked, Mistakes records SPECIFIC FAILURE PATTERNS with a stable fingerprint so the agent can (a) recognise it is repeating itself, (b) be told exactly what not to do again in every future system prompt, and (c) capture the FIX once one is found so the avoidance lesson becomes an actionable correction.

A "mistake" is keyed by sha12(tool + normalised_error). Normalisation strips volatile bits (paths, hex addresses, line numbers, timestamps, UUIDs, PIDs) so "NoMethodError ... at foo.rb:42" and "... at foo.rb:99" collapse to one signature and its :count climbs — that count IS the repeat detector.

Closed loop (why it does NOT repeat mistakes):

Loop.run --(tool failure)---------> Mistakes.record        (persist + count++)
Loop.run --(same sig fails ≥N)----> guard_repeated_failure (uses PERSISTENT count,
                                                          so triggers on the 1st
                                                          recurrence in a new
                                                          session, not the 3rd)
Loop.run --(failure w/ known fix)-> inline "KNOWN FIX: …"  (self-corrects next iter)
Loop.run --(user says "wrong")----> check_user_correction  (flip last outcome + record)
PromptBuilder <-------------------- Mistakes.to_context    (DO-NOT-REPEAT + KNOWN-FIXES)
model --(tool call)---------------> mistakes_record / mistakes_resolve

Constant Summary collapse

MISTAKES_FILE = 
File.join(Dir.home, '.pwn', 'mistakes.json')
REPEAT_THRESHOLD = 
3
CORRECTION_RX = 
/
  \b(
    no[,.]?\s*(that|this|it)?'?s?\s*(wrong|not\s+right|incorrect)|
    still\s+(broken|failing|wrong|not\s+working|doesn'?t\s+work)|
    (that|it|this)\s+(did(n'?t| not)\s+work|failed|is\s+wrong)|
    not\s+what\s+i\s+(asked|meant|wanted)|
    you\s+(made\s+a|got\s+it)\s+(mistake|wrong)|
    same\s+(mistake|error|problem)\s+again|
    try\s+again|redo\s+that|wrong\s+answer|incorrect
  )\b
/ix

Class Method Summary collapse

Class Method Details

.authors ⇒ Object

Author(s)

0day Inc. support@0dayinc.com



540
541
542
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 540

public_class_method def self.authors
  "AUTHOR(S):\n  0day Inc. <support@0dayinc.com>\n"
end

.check_user_correction(opts = {}) ⇒ Object

Supported Method Parameters

entry = PWN::AI::Agent::Mistakes.check_user_correction( request: 'required - the incoming user message', session_id: 'optional - session to inspect for the previous answer' )

When the user's new message reads like a correction of the previous answer, this (a) flips the most recent Learning outcome for that session to success:false, and (b) records a mistake with source :user_correction whose "error" is the user's own words. This is the strongest available signal that the agent was WRONG.



454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 454

public_class_method def self.check_user_correction(opts = {})
  request    = opts[:request].to_s
  session_id = opts[:session_id]
  return nil unless correction?(request: request)

  prev = previous_assistant(session_id: session_id)
  Learning.flip_last_outcome(session_id: session_id, reason: request[0, 200]) if defined?(Learning)
  # W1 — stash the rejected answer + user prompt so the NEXT final
  # (the correction) completes a (rejected, chosen) preference pair.
  Thread.current[:pwn_pending_pref] = { prompt: previous_user(session_id: session_id).to_s, rejected: prev.to_s } if defined?(Reward)
  record(
    tool: 'assistant_answer',
    error: "user rejected previous answer: #{request.strip[0, 200]}",
    args: prev.to_s[0, 200],
    session_id: session_id,
    source: :user_correction
  )
rescue StandardError => e
  warn "[pwn-ai/mistakes] check_user_correction swallowed: #{e.class}: #{e.message}"
  nil
end

.correction?(opts = {}) ⇒ Boolean

Supported Method Parameters

bool = PWN::AI::Agent::Mistakes.correction?(request: user_text)

Returns:

  • (Boolean) 


435
436
437
438
439
440
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 435

public_class_method def self.correction?(opts = {})
  req = opts[:request].to_s
  return false if req.strip.empty?

  req.match?(CORRECTION_RX) && req.length < 600
end

.correction_hint(opts = {}) ⇒ Object

Supported Method Parameters

str = PWN::AI::Agent::Mistakes.correction_hint( tool: 'required - tool that just failed', error: 'required - raw error it failed with' )

Called by Loop.run immediately after a failed dispatch. Returns a string to append to the tool result telling the model (a) how many times this exact failure has occurred across ALL sessions, and (b) the recorded fix if one exists — so it can self-correct on the very next iteration instead of re-discovering the fix from scratch.



422
423
424
425
426
427
428
429
430
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 422

public_class_method def self.correction_hint(opts = {})
  m = find(tool: opts[:tool], error: opts[:error])
  return '' unless m

  parts = ["seen #{m[:count]}× across #{Array(m[:sessions]).length} session(s), sig=#{m[:signature]}"]
  parts << 'REGRESSED (previous fix did not hold)' if m[:regressed]
  parts << "KNOWN FIX: #{m[:fix]}" if m[:fix].to_s.strip.length.positive?
  "[pwn-ai/mistakes] #{parts.join(' | ')}"
end

.effective_count(opts = {}) ⇒ Object

Age-weighted count for [REPEATING] threshold — a ×8 signature from 6 months ago on a since-rewritten module decays toward zero.



507
508
509
510
511
512
513
514
515
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 507

public_class_method def self.effective_count(opts = {})
  m = opts[:mistake] || find(signature: opts[:signature])
  return 0 unless m

  days = (Time.now.utc - Time.parse(m[:last_seen].to_s)) / 86_400.0
  (m[:count].to_f * (0.5**(days / 30.0))).ceil
rescue StandardError
  m ? m[:count].to_i : 0
end

.find(opts = {}) ⇒ Object

Supported Method Parameters

entry = PWN::AI::Agent::Mistakes.find( signature: 'optional - exact signature to fetch', tool: 'optional - with error:, compute signature and fetch', error: 'optional - raw error text (used with tool:)' )



111
112
113
114
115
116
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 111

public_class_method def self.find(opts = {})
  sig = opts[:signature] || (opts[:tool] && opts[:error] ? signature(tool: opts[:tool], error: opts[:error]) : nil)
  return nil unless sig

  load[sig.to_sym]
end

.for_tool(opts = {}) ⇒ Object

Supported Method Parameters

rows = PWN::AI::Agent::Mistakes.for_tool( tool: 'required - tool name', unresolved_only: 'optional - default false' )



124
125
126
127
128
129
130
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 124

public_class_method def self.for_tool(opts = {})
  tool = opts[:tool].to_s
  only = opts[:unresolved_only] ? true : false
  rows = load.values.select { |m| m[:tool].to_s == tool }
  rows = rows.reject { |m| m[:resolved] } if only
  rows.sort_by { |m| -m[:count].to_i }
end

.help ⇒ Object

Display Usage for this Module



546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 546

public_class_method def self.help
  puts <<~USAGE
    USAGE:
      PWN::AI::Agent::Mistakes.record(tool: 'shell', error: 'nmpa: command not found', args: '{"command":"nmpa -sV"}')
      PWN::AI::Agent::Mistakes.top(limit: 10, unresolved_only: true)
      PWN::AI::Agent::Mistakes.find(signature: 'abc123def456')
      PWN::AI::Agent::Mistakes.for_tool(tool: 'shell')
      PWN::AI::Agent::Mistakes.resolve(signature: 'abc123def456', fix: 'binary is spelled `nmap`, not `nmpa`')
      PWN::AI::Agent::Mistakes.correction_hint(tool: 'shell', error: err)  # inline self-correct
      PWN::AI::Agent::Mistakes.to_context(limit: 6)                        # injected by PromptBuilder
      PWN::AI::Agent::Mistakes.correction?(request: "no that's wrong")
      PWN::AI::Agent::Mistakes.check_user_correction(request: req, session_id: sid)
      PWN::AI::Agent::Mistakes.signature(tool: 'shell', error: err)
      PWN::AI::Agent::Mistakes.reset

      #{self}.authors
  USAGE
end

.load ⇒ Object

Supported Method Parameters

store = PWN::AI::Agent::Mistakes.load



54
55
56
57
58
59
60
61
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 54

public_class_method def self.load
  FileUtils.mkdir_p(File.dirname(MISTAKES_FILE))
  return {} unless File.exist?(MISTAKES_FILE)

  JSON.parse(File.read(MISTAKES_FILE), symbolize_names: true)
rescue StandardError
  {}
end

.park(opts = {}) ⇒ Object

2.5 — park unfixable sigs so nightly practice skips them



308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 308

public_class_method def self.park(opts = {})
  sig = opts[:signature].to_s
  raise 'ERROR: signature is required' if sig.empty?

  store = load
  key = sig.to_sym
  raise "ERROR: unknown mistake signature #{sig}" unless store[key]

  store[key][:parked] = true
  store[key][:needs_code_change] = true
  store[key][:park_reason] = opts[:reason].to_s[0, 300]
  store[key][:parked_at] = Time.now.utc.iso8601
  save(store: store)
  store[key]
end

.record(opts = {}) ⇒ Object

Supported Method Parameters

entry = PWN::AI::Agent::Mistakes.record( tool: 'required - tool/component that produced the failure', error: 'required - error text / message', args: 'optional - args that triggered it (stored truncated as sample)', session_id: 'optional - PWN::Sessions id', source: 'optional - :tool | :user_correction | :loop | :model | :heuristic (default :tool)' )

Returns the FULL persisted entry including its cumulative :count so the caller (Loop.run) can drive cross-session repeat detection.



144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 144

public_class_method def self.record(opts = {})
  tool  = opts[:tool].to_s
  error = opts[:error].to_s
  return nil if tool.empty? || error.strip.empty?

  # 1.1 — reward_signal: never inflate count beyond 1 open fingerprint.
  # Sentinel opens one parked sig; further gaps calibrate, not spam.
  if tool == 'reward_signal'
    existing = load.values.select { |e| e[:tool].to_s == 'reward_signal' && !e[:resolved] && !e[:parked] }
    if existing.any? && !opts[:force]
      e = existing.max_by { |x| x[:count].to_i }
      e[:last_seen] = Time.now.utc.iso8601
      e[:count] = e[:count].to_i # freeze
      e[:meta] = (e[:meta] || {}).merge(opts[:meta] || {})
      store = load
      store[e[:signature].to_sym] = e
      save(store: store)
      return e
    end
  end

  sig   = signature(tool: tool, error: error)
  store = load
  key   = sig.to_sym
  now   = Time.now.utc.iso8601
  norm  = normalize_error(error: error)

  m = store[key] ||= {
    signature: sig, tool: tool, error: norm,
    snippet: error.to_s.strip[0, 300],
    count: 0, drift_count: 0, first_seen: now, sessions: [],
    resolved: false, fix: nil, source: (opts[:source] || :tool).to_s
  }
  was_resolved = m[:resolved]
  # E1 — env-drift-attributed failures are counted separately so
  # they do NOT push the signature toward [REPEATING]. "The world
  # changed under me" is not the same lesson as "I did it wrong".
  cause = (opts[:cause] || :self).to_sym
  if cause == :env_drift
    m[:drift_count] = m[:drift_count].to_i + 1
    m[:cause] = 'env_drift'
  else
    m[:count] += 1
  end
  m[:last_seen]    = now
  m[:snippet]      = error.to_s.strip[0, 300]
  m[:sample_args]  = opts[:args].to_s[0, 200] if opts[:args]
  m[:sessions]     = (Array(m[:sessions]) + [opts[:session_id]]).compact.uniq.last(10)
  # 2.2 — recoverable shape for repair routing
  if opts[:shape]
    m[:shape] = opts[:shape].to_s
  elsif defined?(Reward) && Reward.respond_to?(:recoverable_shape)
    m[:shape] ||= Reward.recoverable_shape(err: error).to_s
  end
  m[:needs_code_change] = true if opts[:needs_code_change]
  m[:meta] = (m[:meta] || {}).merge(opts[:meta] || {}) if opts[:meta]
  # A recurrence of a "resolved" mistake means the fix was wrong /
  # incomplete — reopen it so it re-enters the DO-NOT-REPEAT block.
  # Structured fixes with holdout_tests that still pass stay closed.
  if was_resolved && structured_fix_holds?(mistake: m)
    m[:resolved] = true
    m[:regressed] = false
  else
    m[:resolved]  = false
    m[:regressed] = true if was_resolved
  end
  save(store: store)
  m
end

.reset ⇒ Object

Supported Method Parameters

PWN::AI::Agent::Mistakes.reset



479
480
481
482
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 479

public_class_method def self.reset
  FileUtils.rm_f(MISTAKES_FILE)
  {}
end

.resolve(opts = {}) ⇒ Object

Supported Method Parameters

entry = PWN::AI::Agent::Mistakes.resolve( signature: 'required - mistake signature (from mistakes_list / .top)', fix: 'required - what to do INSTEAD next time' )



220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 220

public_class_method def self.resolve(opts = {})
  sig = opts[:signature].to_s
  fix = opts[:fix].to_s
  raise 'ERROR: signature is required' if sig.empty?
  raise 'ERROR: fix is required' if fix.strip.empty?

  store = load
  key   = sig.to_sym
  raise "ERROR: unknown mistake signature #{sig}" unless store[key]

  store[key][:resolved]    = true
  store[key][:regressed]   = false
  store[key][:fix]         = fix.strip[0, 500]
  store[key][:resolved_at] = Time.now.utc.iso8601
  # 2.3 — structured fix payload (strategy/tool/args_template/holdouts).
  # Prose-only resolve is why shell sigs regressed after auto-curriculum.
  if opts[:structured].is_a?(Hash)
    s = opts[:structured]
    store[key][:structured_fix] = {
      strategy: s[:strategy].to_s[0, 80],
      tool: s[:tool].to_s[0, 60],
      args_template: s[:args_template],
      holdout_tests: Array(s[:holdout_tests]).first(5),
      winning_trace: s[:winning_trace].to_s[0, 2_000]
    }.compact
  end
  store[key][:parked] = false
  store[key][:needs_code_change] = false if opts[:clear_needs_code_change]
  save(store: store)

  if defined?(PWN::Memory)
    PWN::Memory.remember(
      key: :"mistake_fix_#{sig}",
      value: "AVOID: #{store[key][:tool]} → #{store[key][:error]} — FIX: #{fix.strip[0, 300]}",
      category: :lesson,
      source: :resolve,
      confidence: 0.9,
      importance: 0.9
    )
  end
  # W1/P9 — every resolve is a preference pair. Prefer structured
  # winning_trace (+ strategy/tool) over first-line fix prose so DPO
  # learns tool trajectories, not commentary.
  if defined?(Reward)
    sf = store[key][:structured_fix] || {}
    trace = sf[:winning_trace].to_s.strip
    strat = [sf[:strategy], sf[:tool], sf[:args_template]].compact.map(&:to_s).reject(&:empty?).join(' | ')
    # P21/P25 — only write W1 pairs when we have a real winning_trace.
    # Prose-only resolve still updates Memory lesson + structured_fix;
    # it must NOT flood DPO with fix commentary (shape: :fix_prose).
    if trace.length >= 40
      parts = []
      parts << "STRATEGY: #{strat}" unless strat.empty?
      parts << "WINNING_TRACE:\n#{trace[0, 3_500]}"
      parts << "FIX: #{fix.strip[0, 400]}"
      chosen = parts.join("\n")
      rejected = store[key][:snippet].to_s
      rejected = "FAILING: tool=#{store[key][:tool]} err=#{store[key][:error]}" if rejected.strip.empty?
      Reward.record_preference(
        prompt: "#{store[key][:tool]}: #{store[key][:error]}",
        rejected: rejected,
        chosen: chosen,
        source: :mistakes_resolve,
        shape: :winning_trace,
        meta: { signature: sig, strategy: sf[:strategy], tool: sf[:tool] }.compact
      )
    end
  end
  store[key]
end

.save(opts = {}) ⇒ Object

Supported Method Parameters

PWN::AI::Agent::Mistakes.save(store: hash)



66
67
68
69
70
71
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 66

public_class_method def self.save(opts = {})
  store = opts[:store] ||= {}
  FileUtils.mkdir_p(File.dirname(MISTAKES_FILE))
  atomic_write(path: MISTAKES_FILE, body: JSON.pretty_generate(store))
  store
end

.signature(opts = {}) ⇒ Object

Supported Method Parameters

sig = PWN::AI::Agent::Mistakes.signature( tool: 'required - tool/component name that failed', error: 'required - raw error text (will be normalised)' )



98
99
100
101
102
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 98

public_class_method def self.signature(opts = {})
  tool = opts[:tool].to_s
  norm = normalize_error(error: opts[:error])
  Digest::SHA256.hexdigest("#{tool}|#{norm}")[0, 12]
end

.to_context(opts = {}) ⇒ Object

Supported Method Parameters

ctx = PWN::AI::Agent::Mistakes.to_context(limit: 6)

Injected by PromptBuilder into every system prompt. Emits TWO blocks so the model sees both what NOT to do AND what to do INSTEAD:

KNOWN MISTAKES — unresolved, count-sorted, [REPEATING]/[REGRESSED]
KNOWN FIXES    — resolved entries with their fix, so the correction
               survives even after dropping out of the first list.


334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 334

public_class_method def self.to_context(opts = {})
  limit   = opts[:limit] || 6
  request = opts[:request].to_s
  open_rows = top(limit: limit * 3, unresolved_only: true)
  # 2.6 — request-conditioned rank (sim × recency × count), same idea
  # as exemplars_for. Stops injecting loudest scar (reward_signal ×13)
  # on every unrelated turn.
  open = rank_for_request(rows: open_rows, request: request, limit: limit)
  closed = load.values.select { |m| m[:resolved] && m[:fix] }
  closed = rank_for_request(rows: closed, request: request, limit: limit)
  return '' if open.empty? && closed.empty?

  out = +''
  unless open.empty?
    lines = open.map do |m|
      tags = []
      tags << 'REPEATING' if effective_count(mistake: m) >= REPEAT_THRESHOLD
      tags << 'ENV_DRIFT' if m[:cause].to_s == 'env_drift'
      tags << 'REGRESSED' if m[:regressed]
      tags << 'PARKED' if m[:parked] || m[:needs_code_change]
      tag = tags.empty? ? '' : " [#{tags.join(',')}]"
      fix = m[:fix] ? " — last fix (insufficient): #{m[:fix][0, 100]}" : ''
      shape = m[:shape] ? " shape=#{m[:shape]}" : ''
      "  ✗ [#{m[:signature]}] #{m[:tool]} ×#{m[:count]}#{tag}#{shape}: #{m[:error][0, 140]}#{fix}"
    end
    out << "KNOWN MISTAKES (do NOT repeat — call mistakes_resolve once fixed)\n#{lines.join("\n")}\n"
  end
  unless closed.empty?
    lines = closed.map do |m|
      sf = m[:structured_fix]
      extra = sf ? " strategy=#{sf[:strategy]} tool=#{sf[:tool]}" : ''
      "  ✓ [#{m[:signature]}] #{m[:tool]}: #{m[:error][0, 80]} — FIX: #{m[:fix][0, 140]}#{extra}"
    end
    out << "KNOWN FIXES (apply these instead of repeating the mistake)\n#{lines.join("\n")}\n"
  end
  "#{out}\n"
end

.top(opts = {}) ⇒ Object

Supported Method Parameters

rows = PWN::AI::Agent::Mistakes.top( limit: 'optional - max rows (default 10)', unresolved_only: 'optional - default true' )



297
298
299
300
301
302
303
304
305
 
# File 'lib/pwn/ai/agent/mistakes.rb', line 297

public_class_method def self.top(opts = {})
  limit = opts[:limit] || 10
  only  = opts.key?(:unresolved_only) ? opts[:unresolved_only] : true
  rows  = load.values
  rows  = rows.reject { |m| m[:resolved] } if only
  # 2.5 — practice/curriculum skip engineer-only / parked fingerprints
  rows = rows.reject { |m| m[:parked] || m[:needs_code_change] || m[:tool].to_s == 'reward_signal' } if opts[:practiceable_only]
  rows.sort_by { |m| [-m[:count].to_i, m[:last_seen].to_s] }.first(limit)
end