@yahyahammad/venastine 0.2.0 → 0.3.8

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (129) hide show
  1. package/.env.example +19 -14
  2. package/CONFIG_ARCHITECTURE.md +2002 -0
  3. package/NOTICE +1 -1
  4. package/README.md +233 -94
  5. package/agents/builtin/build.md +73 -0
  6. package/agents/builtin/general.md +80 -0
  7. package/agents/builtin/plan.md +23 -5
  8. package/agents/builtin/test.md +65 -0
  9. package/agents/builtin/writer.md +67 -0
  10. package/agents/manager.py +261 -195
  11. package/agents/subagent_tool.py +277 -43
  12. package/agents/tui_commands.py +7 -12
  13. package/bin/venastine.mjs +528 -382
  14. package/bin/venastine.py +71 -0
  15. package/cli_commands.py +835 -0
  16. package/config.py +95 -1198
  17. package/config.yaml +1090 -0
  18. package/config_edit.py +1301 -0
  19. package/config_schema.py +966 -0
  20. package/config_update.py +749 -0
  21. package/core/agent_activity.py +158 -20
  22. package/core/ansi.py +452 -0
  23. package/core/approval.py +27 -5
  24. package/core/call_view.py +364 -0
  25. package/core/cancel.py +121 -0
  26. package/core/chatgpt_auth.py +747 -0
  27. package/core/claims_view.py +107 -0
  28. package/core/client.py +1238 -82
  29. package/core/commands.py +454 -0
  30. package/core/compaction.py +830 -66
  31. package/core/config_loader.py +516 -97
  32. package/core/connect.py +686 -0
  33. package/core/events.py +68 -6
  34. package/core/file_mentions.py +280 -0
  35. package/core/interaction.py +87 -4
  36. package/core/line_pattern.py +128 -0
  37. package/core/loop.py +1516 -440
  38. package/core/memory.py +120 -8
  39. package/core/pipeline_models.py +23 -9
  40. package/core/provider_errors.py +234 -0
  41. package/core/reasoning/confidence_scoring.py +70 -54
  42. package/core/reasoning/json_retry.py +19 -1
  43. package/core/reasoning/orchestrator.py +232 -53
  44. package/core/reasoning/output_writer.py +90 -31
  45. package/core/reasoning/payload_validation.py +133 -7
  46. package/core/reasoning/pipeline_storage.py +427 -361
  47. package/core/reasoning/review.py +659 -615
  48. package/core/reasoning/scholar.py +38 -14
  49. package/core/reasoning/source_corpus.py +110 -4
  50. package/core/replay.py +158 -26
  51. package/core/session.py +36 -12
  52. package/core/session_view.py +259 -0
  53. package/core/session_wake.py +259 -0
  54. package/core/shell_sessions.py +1553 -0
  55. package/core/workspace_trust.py +1 -1
  56. package/credentials.py +308 -180
  57. package/database.py +82 -1
  58. package/{PRIVACY.md → docs/PRIVACY.md} +19 -7
  59. package/docs/SECURITY.md +108 -0
  60. package/docs/THIRD_PARTY_NOTICES.md +148 -0
  61. package/env_secrets.py +118 -77
  62. package/json_store.py +236 -16
  63. package/logging_setup.py +121 -11
  64. package/main.py +1671 -199
  65. package/mcp_client/client.py +257 -23
  66. package/mcp_client/config.py +484 -379
  67. package/mcp_client/registration.py +163 -2
  68. package/memories/manager.py +15 -0
  69. package/memories/tui_commands.py +7 -12
  70. package/package.json +234 -4
  71. package/project_init/config_files.py +4 -0
  72. package/project_init/generator.py +201 -37
  73. package/project_init/manifest.py +418 -406
  74. package/project_init/tui_commands.py +7 -4
  75. package/prompts/revalidate.md +1 -1
  76. package/prompts/revise.md +17 -15
  77. package/prompts/system_prompts.py +64 -10
  78. package/providers.json.example +94 -92
  79. package/pyproject.toml +51 -2
  80. package/requirements.txt +252 -29
  81. package/safety/policy_enforcement.py +353 -17
  82. package/security/askpass.py +419 -0
  83. package/security/capability.py +37 -7
  84. package/security/confine.py +484 -0
  85. package/security/hosted.py +182 -0
  86. package/security/permissions.py +54 -7
  87. package/security/posture.py +124 -12
  88. package/security/protected_paths.py +198 -6
  89. package/security/sandbox.py +3262 -139
  90. package/security/secrets.py +855 -0
  91. package/skills/tui_commands.py +3 -4
  92. package/storage.py +368 -35
  93. package/tools/base.py +465 -305
  94. package/tools/builtin/_math_common.py +306 -16
  95. package/tools/builtin/_net_common.py +267 -61
  96. package/tools/builtin/arxiv.py +681 -223
  97. package/tools/builtin/ask_user.py +231 -213
  98. package/tools/builtin/discrete_math.py +35 -4
  99. package/tools/builtin/fetch_url.py +239 -21
  100. package/tools/builtin/file_ops.py +341 -33
  101. package/tools/builtin/geometry.py +184 -148
  102. package/tools/builtin/get_time.py +2 -2
  103. package/tools/builtin/linear_algebra.py +66 -23
  104. package/tools/builtin/load_skill.py +10 -0
  105. package/tools/builtin/logic.py +171 -12
  106. package/tools/builtin/pin.py +46 -12
  107. package/tools/builtin/probability_stats.py +224 -228
  108. package/tools/builtin/project_docs.py +80 -21
  109. package/tools/builtin/remember.py +21 -4
  110. package/tools/builtin/shell.py +586 -67
  111. package/tools/builtin/shell_sessions.py +613 -0
  112. package/tools/builtin/symbolic_math.py +80 -10
  113. package/tools/builtin/todo.py +5 -3
  114. package/tools/builtin/web_search.py +329 -151
  115. package/tools/context.py +23 -0
  116. package/tools/isolation.py +387 -314
  117. package/tools/registry.py +347 -30
  118. package/tui/app.py +4645 -438
  119. package/tui/app.tcss +185 -2
  120. package/tui/commands.py +10 -154
  121. package/tui/diffs.py +286 -265
  122. package/tui/hosted_driver.py +102 -0
  123. package/tui/markdown.py +60 -10
  124. package/tui/meters.py +103 -7
  125. package/tui/screens.py +346 -92
  126. package/tui/themes.py +59 -19
  127. package/tui/widgets.py +1575 -111
  128. package/SECURITY.md +0 -88
  129. package/THIRD_PARTY_NOTICES.md +0 -100
@@ -18,7 +18,11 @@ flag it if you intended something different.
18
18
  from __future__ import annotations
19
19
 
20
20
  from core.reasoning.base import (
21
- CLAIM_TYPES, GROUNDING_STATUSES, Claim, ConfidenceTier, PipelineRun,
21
+ CLAIM_TYPES,
22
+ GROUNDING_STATUSES,
23
+ Claim,
24
+ ConfidenceTier,
25
+ PipelineRun,
22
26
  )
23
27
 
24
28
  # --- Tunable weights -- change these, not the logic below, to retune scoring ---
@@ -28,27 +32,32 @@ CRITIC_WEIGHT_FACTOR = 0.35
28
32
  ASSUMPTION_FLAG_PENALTY = 0.15 # per flag
29
33
  NON_FACTUAL_SCORE_CAP = 0.65 # claims never grounded can't reach HIGH
30
34
 
31
- # ROADMAP §10 (revisit, E8). Cross-candidate disagreement is a PENALTY,
32
- # subtracted like ASSUMPTION_FLAG_PENALTY above -- not a fourth weighted
33
- # term with the others scaled down to make room for it.
35
+ # ROADMAP_v3 §90 (A85, issue #12). Cross-candidate disagreement is ROUTED,
36
+ # NOT SCORED. A factual claim asserted by at most this share of the
37
+ # candidates is `contested`, and D1 sends it to the 6a/6b loop to be
38
+ # checked; nothing is subtracted for it.
34
39
  #
35
- # §10 shipped the other shape: 0.5/0.35 redistributed to 0.4/0.3 with a
36
- # 0.15 consistency term added. Two consequences, neither intended:
40
+ # It was a penalty before (§10's revisit, E8): 0.15 * (1 - consistency),
41
+ # itself the repair of §10's first shape, a fourth weighted term that made
42
+ # grounding count for less whenever ensemble mode was on. Two things were
43
+ # wrong with the penalty, and the second was measured:
37
44
  #
38
- # 1. Turning ensemble mode on made GROUNDING count for less. Grounding is
39
- # the pipeline's strongest evidence, and which generation strategy
40
- # produced a claim is no reason to weigh its verification differently.
41
- # 2. Because the maximum stayed 0.85, consistency could only ever cost a
42
- # well-grounded claim. A claim Pass 3a grounded and Pass 3b found
43
- # nothing wrong with dropped from HIGH to MEDIUM at 2-of-3 agreement.
45
+ # 1. It answered the wrong question. Disagreement is the one signal in
46
+ # the pipeline that points at a specific claim and says "the models do
47
+ # not concur here", which is what a verification pass wants as input,
48
+ # and a deduction spends it on a number instead.
49
+ # 2. It could send a claim into the loop that no revision could get out
50
+ # of. A partially grounded claim scores 0.60 (MEDIUM); at 1 of 3 the
51
+ # penalty took it to 0.50 (LOW), so D1 flagged it, 6a was never told
52
+ # why, and 6b re-scored it under the same penalty -- who asserted a
53
+ # claim does not change when it is reworded -- until D2 published it
54
+ # UNVERIFIED.
44
55
  #
45
- # As a penalty, unanimity changes nothing and dissent subtracts, which is
46
- # also the honest reading of the signal: agreement is the null hypothesis
47
- # (candidates asked the same question usually answer alike), so disagreement
48
- # is the part that carries information. It makes the non-ensemble formula
49
- # literally unchanged rather than coincidentally equal, so §10's
50
- # byte-for-byte regression guard holds by construction.
51
- DISAGREEMENT_PENALTY_FACTOR = 0.15
56
+ # AT MOST HALF, so 1 of 2, 1 of 3 and 2 of 4 are contested and 2 of 3 is
57
+ # not: exactly the shapes where the penalty used to change a tier. The
58
+ # non-ensemble formula is untouched, and the ensemble one is now the same
59
+ # formula, so §10's byte-for-byte regression guard holds for both.
60
+ CONTESTED_SHARE = 0.5
52
61
 
53
62
  # ROADMAP_v2 §45 (SQ3). What a grounding is WORTH.
54
63
  #
@@ -113,7 +122,6 @@ def default_weights() -> dict:
113
122
  "grounding_weight_factor": GROUNDING_WEIGHT_FACTOR,
114
123
  "critic_weight_factor": CRITIC_WEIGHT_FACTOR,
115
124
  "assumption_flag_penalty": ASSUMPTION_FLAG_PENALTY,
116
- "disagreement_penalty_factor": DISAGREEMENT_PENALTY_FACTOR,
117
125
  "non_factual_score_cap": NON_FACTUAL_SCORE_CAP,
118
126
  "source_quality_floor": SOURCE_QUALITY_FLOOR,
119
127
  "authority_full_credit": AUTHORITY_FULL_CREDIT,
@@ -194,9 +202,10 @@ def score_claim(claim: Claim, ensemble_n: int = 0,
194
202
  """Returns (tier, score_breakdown). Pure function -- no side effects,
195
203
  easy to unit test claim-by-claim without running any pass.
196
204
 
197
- When ensemble_n > 0 (ROADMAP §10), cross-candidate DISAGREEMENT -- the
198
- fraction of candidates that did NOT independently assert this claim --
199
- is subtracted from a factual claim's score. ensemble_n is the number of
205
+ When ensemble_n > 0 (ROADMAP §10), the share of candidates that
206
+ asserted this claim is recorded, and a factual claim asserted by at
207
+ most CONTESTED_SHARE of them is marked `contested` for D1 to route
208
+ (ROADMAP_v3 §90). It changes no score. ensemble_n is the number of
200
209
  candidates that actually produced text, which on a run where a
201
210
  candidate's provider failed is fewer than the configured roster.
202
211
 
@@ -251,29 +260,25 @@ def score_claim(claim: Claim, ensemble_n: int = 0,
251
260
  critic_component = 1.0 - claim.critic_severity
252
261
  assumption_penalty = w["assumption_flag_penalty"] * len(claim.assumption_flags)
253
262
 
254
- # None when ensemble mode is off, which is what makes the formula below
255
- # ONE formula rather than a pair that can drift -- and (B7) also when
256
- # ensemble mode is ON and Pass 2 reported nothing for this claim, which
257
- # is a different fact from "no candidate asserted it".
263
+ # None when ensemble mode is off -- and (B7) also when ensemble mode is
264
+ # ON and Pass 2 reported nothing for this claim, which is a different
265
+ # fact from "no candidate asserted it". An unreported claim is never
266
+ # contested.
258
267
  consistency_score = None
259
268
  consistency_reported = None
260
- disagreement_penalty = 0.0
261
269
  if ensemble_n > 0 and claim.asserted_by_candidates:
262
- # B8. Deduplicated and clamped. `[1, 2, 3, 3]` made
263
- # consistency 1.3333, so the PENALTY went NEGATIVE and a MEDIUM
264
- # claim scored HIGH -- the exact property E8's comment says holds
265
- # "by construction". `consistency_reported` keeps the unclamped
266
- # value when it fired, because a numerator above the denominator is
267
- # itself evidence that Pass 2 misread the candidate labelling, and
270
+ # B8. Deduplicated and clamped. `[1, 2, 3, 3]` made consistency
271
+ # 1.3333, which under E8 made the penalty NEGATIVE and scored a
272
+ # MEDIUM claim HIGH; here a duplicate would make one candidate's
273
+ # claim read as two agreeing, and `[1, 1]` of 3 would escape
274
+ # routing. `consistency_reported` keeps the unclamped value when
275
+ # it fired, because a numerator above the denominator is itself
276
+ # evidence that Pass 2 misread the candidate labelling, and
268
277
  # clamping silently would hide the evidence along with the bug.
269
278
  reported = len(set(claim.asserted_by_candidates)) / ensemble_n
270
279
  consistency_score = min(1.0, reported)
271
280
  if reported > 1.0:
272
281
  consistency_reported = round(reported, 4)
273
- # E9: computed only where it is applied, so the breakdown records
274
- # what was subtracted rather than what was available (B10).
275
- if claim_type == "factual":
276
- disagreement_penalty = w["disagreement_penalty_factor"] * (1.0 - consistency_score)
277
282
 
278
283
  if claim_type != "factual":
279
284
  # Cap applies to the critic component BEFORE subtracting the
@@ -283,11 +288,6 @@ def score_claim(claim: Claim, ensemble_n: int = 0,
283
288
  # claim that was already clean under critique. Found this via a
284
289
  # direct test case (two speculative claims, one flagged one not,
285
290
  # both landing at the same tier) before it shipped.
286
- #
287
- # E9: no disagreement penalty here. §10 decided consistency does not
288
- # RESCUE a non-factual claim from the cap; whether dissent should
289
- # penalise a speculative claim is a separate judgment about what
290
- # candidates declining to speculate means, and it was not made.
291
291
  capped_critic = min(critic_component, w["non_factual_score_cap"])
292
292
  raw_score = capped_critic - assumption_penalty
293
293
  formula = "non_factual"
@@ -298,7 +298,6 @@ def score_claim(claim: Claim, ensemble_n: int = 0,
298
298
  (w["grounding_weight_factor"] * grounding_component)
299
299
  + (w["critic_weight_factor"] * critic_component)
300
300
  - assumption_penalty
301
- - disagreement_penalty
302
301
  )
303
302
 
304
303
  # ROUND ONCE, then use that one value for BOTH the tier comparison and
@@ -309,8 +308,9 @@ def score_claim(claim: Claim, ensemble_n: int = 0,
309
308
  # auditing a run artifact read a number that contradicted the published
310
309
  # thresholds, and the tier was the one that mattered.
311
310
  #
312
- # Binary floating point makes this reachable from ordinary weights, not
313
- # from an exotic input: 0.5 + 0.35 - 0.15 * (1/3) is exactly that value.
311
+ # Binary floating point made this reachable from ordinary weights, not
312
+ # from an exotic input: 0.5 + 0.35 - 0.15 * (1/3), E8's 2-of-3 penalty,
313
+ # was exactly that value.
314
314
  # Sweeping the whole non-ensemble input space, exactly ONE shape changes
315
315
  # tier under this fix -- a speculative claim at severity 0.55 with one
316
316
  # assumption flag, which records 0.3 and now tiers LOW instead of
@@ -336,7 +336,7 @@ def score_claim(claim: Claim, ensemble_n: int = 0,
336
336
  # extra key: "anyone auditing a run artifact read a number that
337
337
  # contradicted the published thresholds."
338
338
  #
339
- # factual 0.5*grounding + 0.35*critic - assumption - disagreement
339
+ # factual 0.5*grounding + 0.35*critic - assumption
340
340
  # non_factual capped_critic - assumption
341
341
  breakdown = {
342
342
  "formula": formula,
@@ -371,11 +371,18 @@ def score_claim(claim: Claim, ensemble_n: int = 0,
371
371
  # B7. None means PASS 2 REPORTED NOTHING, which is not 0.0 -- and
372
372
  # the two were indistinguishable, so a claim one candidate raised
373
373
  # uncontested took the maximum 0.15 penalty rather than the 0.05 it
374
- # earned. #12 reads this field; it needs to be able to tell them
375
- # apart too.
374
+ # earned under E8. `contested` needs them apart for the same reason.
376
375
  breakdown["consistency_score"] = (
377
376
  None if consistency_score is None else round(consistency_score, 4))
378
- breakdown["disagreement_penalty"] = round(disagreement_penalty, 4)
377
+ # §90: what D1 routes on, and the record of why a claim that scored
378
+ # HIGH went through 6a/6b anyway. E9 LIVES HERE, and only here:
379
+ # whether candidates declining to speculate says anything about a
380
+ # speculative claim is a judgment §10 never made, so the non-factual
381
+ # branch is never asked and carries no key (B10: a field that did
382
+ # nothing is never present).
383
+ if formula == "factual":
384
+ breakdown["contested"] = (consistency_score is not None
385
+ and consistency_score <= CONTESTED_SHARE)
379
386
  if consistency_reported is not None:
380
387
  breakdown["consistency_reported"] = consistency_reported
381
388
  if unrecognised_status is not None:
@@ -385,6 +392,14 @@ def score_claim(claim: Claim, ensemble_n: int = 0,
385
392
  return tier, breakdown
386
393
 
387
394
 
395
+ def is_contested(claim: Claim) -> bool:
396
+ """Whether Pass 5 marked this claim contested (ROADMAP_v3 §90).
397
+
398
+ Read off the breakdown rather than recomputed, so D1 routes on exactly
399
+ what the stored claim says and the type coercion above happens once."""
400
+ return claim.score_breakdown.get("contested") is True
401
+
402
+
388
403
  def build_coverage_gap_entries(completeness: dict) -> list[dict]:
389
404
  """Gaps from Pass 3c become separate coverage-gap entries, not Claim
390
405
  objects -- they were never asserted, so there's nothing to re-tier."""
@@ -398,8 +413,9 @@ def run_confidence_tiering(
398
413
  weights: dict | None = None,
399
414
  ) -> None:
400
415
  """The actual Pass 5 entry point the orchestrator calls. Mutates
401
- the claims in place; zero LLM calls. Pass ensemble_n > 0 to use
402
- the ensemble formula variant (ROADMAP §10).
416
+ the claims in place; zero LLM calls. Pass ensemble_n > 0 to record
417
+ each claim's consistency and mark the contested ones (ROADMAP §10,
418
+ ROADMAP_v3 §90); the formula is the same either way.
403
419
 
404
420
  `claims` narrows WHICH claims are re-tiered, and defaults to all of
405
421
  them -- which is Pass 5 itself, where every claim is in play. The
@@ -449,8 +465,8 @@ def _report_coercions(run: PipelineRun, claim: Claim, breakdown: dict,
449
465
  if ensemble_n > 0 and breakdown.get("consistency_score") is None:
450
466
  run.log(
451
467
  f"Pass 5: claim {claim.id} carries no asserted_by_candidates on "
452
- f"an ensemble run -- scored WITHOUT a disagreement penalty, "
453
- f"since an unreported numerator is not a numerator of zero."
468
+ f"an ensemble run -- not treated as contested, since an "
469
+ f"unreported numerator is not a numerator of zero."
454
470
  )
455
471
  if "unrecognised_grounding_status" in breakdown:
456
472
  run.log(
@@ -44,7 +44,8 @@ status='failed' record.
44
44
  from __future__ import annotations
45
45
 
46
46
  import json
47
- from typing import Any, Callable, Optional
47
+ from collections.abc import Callable
48
+ from typing import Any, Optional
48
49
 
49
50
  import config
50
51
  from core.reasoning.payload_validation import PayloadShapeError
@@ -85,6 +86,7 @@ def retry_until_json(
85
86
  max_total_tokens: Optional[int] = None,
86
87
  validate: Optional[Callable] = None,
87
88
  effort: Optional[str] = None,
89
+ activity=None,
88
90
  ) -> str:
89
91
  """Returns the first response text that parses as JSON, retrying by
90
92
  continuing `response`'s thread.
@@ -121,6 +123,16 @@ def retry_until_json(
121
123
  PayloadShapeError re-enters this same loop. None means
122
124
  "parse is the only contract", which is what the reviewer
123
125
  wants and what every caller had before §30.
126
+ activity §47's agent-activity sink, forwarded for the same reason
127
+ as everything above it: a retry is the same run
128
+ continuing. It matters MORE here than elsewhere, not
129
+ less -- `continue_conversation` drains its own loop, so
130
+ a subagent spawned while the model is being corrected
131
+ reaches a shell through this channel or through nothing
132
+ (`tui/app.py`'s one-shot path carries it for exactly
133
+ that reason). No span is opened here: the caller's own
134
+ brackets the whole attempt-and-retry sequence, and a row
135
+ per correction would report one run as several.
124
136
  """
125
137
  from core.loop import RunAgentLoop
126
138
 
@@ -165,6 +177,12 @@ def retry_until_json(
165
177
  # continuing, so it carries the pass's effort too --
166
178
  # effort_for re-validates against this thread's model.
167
179
  effort=effort,
180
+ # And the same rule again for §47's sink (batch 75). This
181
+ # hop dropped it, so a spawn inside a corrective attempt
182
+ # was invisible -- and this is the one path where the
183
+ # channel is the ONLY route, because continue_conversation
184
+ # drains its own loop.
185
+ activity=activity,
168
186
  **({} if max_total_tokens is None
169
187
  else {"max_total_tokens": max_total_tokens}),
170
188
  )