@yahyahammad/venastine 0.3.13 → 0.3.15

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.
package/main.py CHANGED
@@ -57,6 +57,8 @@ from core import ansi, config_loader, shell_sessions, workspace_trust
57
57
  from core.cancel import CancelToken
58
58
  from core.loop import (
59
59
  DEFAULT_PROVIDER,
60
+ HARNESS_REASONING_ONLY,
61
+ REASONING_ONLY_END,
60
62
  RunAgentLoop,
61
63
  advertisement_facts,
62
64
  agent_turn,
@@ -245,6 +247,10 @@ class _LiveTurn:
245
247
  # is popped out of its result and referenced by nothing else, so
246
248
  # its id could be reused by a later one.
247
249
  self._noticed: list = []
250
+ # ROADMAP_v3 §95 (RO5). The turn's last step already printed, ahead
251
+ # of the notice saying it ended the turn; `finish` then has only
252
+ # the notices left to print.
253
+ self._final_printed = False
248
254
 
249
255
  def __call__(self, event) -> None:
250
256
  try:
@@ -282,7 +288,25 @@ class _LiveTurn:
282
288
  return
283
289
  self._noticed.append(notice)
284
290
  text = notice.get("text") if isinstance(notice, dict) else None
285
- if text:
291
+ kind = notice.get("kind") if isinstance(notice, dict) else None
292
+ if kind in (HARNESS_REASONING_ONLY, REASONING_ONLY_END):
293
+ # ROADMAP_v3 §95 (RO5). Both describe the step that just
294
+ # FINISHED with no tool call, so nothing else will print it:
295
+ # print it first, in the order a replay draws it. Only what
296
+ # streamed -- a turn nobody watched has nothing held, and its
297
+ # last step still prints at `finish`.
298
+ reasoning = "".join(self._reasoning).strip()
299
+ held_text = "".join(self._text)
300
+ if reasoning or held_text:
301
+ self._step(reasoning, held_text, [])
302
+ self._final_printed = kind == REASONING_ONLY_END
303
+ if not text:
304
+ return
305
+ if kind == HARNESS_REASONING_ONLY:
306
+ # The first line of the row the harness wrote, which a replay
307
+ # prints in this role.
308
+ _print_entry("wake", text)
309
+ else:
286
310
  print(f"— {ansi.for_display(text)} —")
287
311
 
288
312
  def finish(self, response) -> None:
@@ -295,10 +319,13 @@ class _LiveTurn:
295
319
  self._notice(notice)
296
320
  calls = [{"id": c.id, "name": c.name, "input": c.input}
297
321
  for c in getattr(response, "tool_calls", None) or ()]
298
- if any(c["id"] in self._announced for c in calls):
322
+ if self._final_printed or any(c["id"] in self._announced
323
+ for c in calls):
299
324
  # This step's reasoning and text printed with its first tool
300
325
  # line: the turn ran out of steps or was stopped between its
301
- # tool calls, after the step itself had finished.
326
+ # tool calls, after the step itself had finished. Or, since
327
+ # §95, ahead of the notice that its reasoning-only reply ended
328
+ # the turn.
302
329
  reasoning, text = "", ""
303
330
  else:
304
331
  reasoning = reasoning_text(getattr(response, "thinking", None))
@@ -368,7 +395,7 @@ def _announce_posture() -> None:
368
395
  app.log and, on the CLI, to stderr through the root handler -- it is
369
396
  the record a bug report is triaged against. The stderr banner is for
370
397
  the human at the terminal, who may have set one of these months ago
371
- and forgotten, or may have inherited a config.py from somewhere.
398
+ and forgotten, or may have inherited a config.yaml from somewhere.
372
399
 
373
400
  Silent when the posture is shipped, so this costs nothing in the case
374
401
  that needs nothing said. That is also why it reads `unsafe_reasons()`
@@ -381,7 +408,7 @@ def _announce_posture() -> None:
381
408
  if not reasons:
382
409
  return
383
410
  print("", file=sys.stderr)
384
- print(" !! REDUCED SECURITY POSTURE -- set in config.py", file=sys.stderr)
411
+ print(" !! REDUCED SECURITY POSTURE -- set in config.yaml", file=sys.stderr)
385
412
  for _label, detail in reasons:
386
413
  logger.warning("Reduced security posture: %s", detail)
387
414
  print(f" - {detail}", file=sys.stderr)
@@ -2206,7 +2233,7 @@ def run_research(
2206
2233
  """Run the full deep-research pipeline, write artifacts to disk,
2207
2234
  and print the results. ensemble_mode/ensemble_n come from
2208
2235
  settings.json (§14); None means the pipeline falls back to the
2209
- config.py defaults. authorization (§25) comes from
2236
+ config.yaml defaults. authorization (§25) comes from
2210
2237
  build_research_authorization(); None means no gated tool is reachable,
2211
2238
  which is the behaviour every research run had before §25.
2212
2239
 
@@ -2367,7 +2394,7 @@ def build_parser() -> argparse.ArgumentParser:
2367
2394
  )
2368
2395
  # Defaults are None so that after parsing we can tell an explicit
2369
2396
  # CLI choice from argparse fill-in; resolution happens in main via
2370
- # resolve_runtime_defaults (CLI > settings.json > config.py).
2397
+ # resolve_runtime_defaults (CLI > settings.json > config.yaml).
2371
2398
  parser.add_argument(
2372
2399
  "--provider",
2373
2400
  default=None,
@@ -2393,7 +2420,7 @@ def build_parser() -> argparse.ArgumentParser:
2393
2420
  metavar="LEVEL",
2394
2421
  help="Reasoning effort for chat AND research (typically "
2395
2422
  "low|medium|high; 'auto' sends none). Default: settings.json "
2396
- "'effort' key, else config.DEFAULT_EFFORT.",
2423
+ "'effort' key, else default_effort in config.yaml.",
2397
2424
  )
2398
2425
  # ROADMAP_v2 §25. TWO flags onto ONE dest, in a mutually exclusive
2399
2426
  # group, rather than a single nargs="?" flag with an optional value.
@@ -2430,7 +2457,7 @@ def build_parser() -> argparse.ArgumentParser:
2430
2457
  "Never persisted.",
2431
2458
  )
2432
2459
  # R9/R12. Default None so settings.json can supply it and an explicit
2433
- # flag still wins (§14's CLI > settings.json > config.py precedence,
2460
+ # flag still wins (§14's CLI > settings.json > config.yaml precedence,
2434
2461
  # which only works because argparse defaults stay None).
2435
2462
  #
2436
2463
  # The MODE is persistable where the grant list is not, and the
@@ -2469,7 +2496,7 @@ def build_parser() -> argparse.ArgumentParser:
2469
2496
  action="store_const",
2470
2497
  const=False,
2471
2498
  help="Research mode: skip the post-pipeline review even if "
2472
- "settings.json or config.py turns it on.",
2499
+ "settings.json or config.yaml turns it on.",
2473
2500
  )
2474
2501
  parser.add_argument(
2475
2502
  "--tui",
@@ -2519,7 +2546,7 @@ def build_parser() -> argparse.ArgumentParser:
2519
2546
  metavar="THREAD",
2520
2547
  action="append",
2521
2548
  help="Attach another thread's summary to this session as context. "
2522
- "Repeatable, up to config.MAX_INJECTED_REFS.",
2549
+ "Repeatable, up to max_injected_refs in config.yaml.",
2523
2550
  )
2524
2551
  # ROADMAP_v2 §24. A flag for M21's reason, and the two kind flags exist
2525
2552
  # so a piped run can skip the question rather than being refused.
@@ -2557,7 +2584,7 @@ def build_parser() -> argparse.ArgumentParser:
2557
2584
  def resolve_runtime_defaults(args, settings: dict) -> tuple[str, str]:
2558
2585
  """Provider/model precedence: explicit CLI flag > settings.json
2559
2586
  (trusted project tier already beat user tier inside the loader) >
2560
- config.py constants."""
2587
+ config.yaml's values."""
2561
2588
  provider = args.provider or settings.get("default_provider") or DEFAULT_PROVIDER
2562
2589
  model = args.model or settings.get("default_model") or config.MODEL_NAME
2563
2590
  return provider, model
@@ -70,7 +70,7 @@ SHUTDOWN_TIMEOUT_S = 15.0
70
70
  # F6 (#64): the WHOLE teardown -- polite close, force-cancel, loop-thread
71
71
  # join -- shares this one wall-clock budget, instead of three sequential
72
72
  # SHUTDOWN_TIMEOUT_S waits whose worst case hung a quitting harness ~45s.
73
- # Master value lives in config.py beside the other harness bounds; the
73
+ # Master value lives in config.yaml beside the other harness bounds; the
74
74
  # re-export here keeps it beside its sibling timeouts and gives tests a
75
75
  # single name to shrink.
76
76
  TEARDOWN_BUDGET_S = _harness_config.TEARDOWN_BUDGET_S
@@ -131,7 +131,7 @@ class MemoryManager:
131
131
  if total > cap:
132
132
  logger.warning(
133
133
  "Injecting %s of %s memories; the oldest %s are not in this "
134
- "prompt. Raise config.MAX_INJECTED_MEMORIES or prune with "
134
+ "prompt. Raise max_injected_memories in config.yaml or prune with "
135
135
  "/forget.", cap, total, total - cap)
136
136
  header = (f"{header}\n\nShowing the {cap} most recent of {total}. "
137
137
  f"Older ones exist and are not listed here.")
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@yahyahammad/venastine",
3
- "version": "0.3.13",
3
+ "version": "0.3.15",
4
4
  "description": "A from-scratch agentic harness: tool-using chat plus a ten-pass self-critiquing research pipeline.",
5
5
  "keywords": [
6
6
  "agent",
@@ -106,9 +106,9 @@ OMITTED = {
106
106
  "disagreement routes a claim to be checked and subtracts nothing, "
107
107
  "and the loader warns when this is set",
108
108
  "ensemble_mode":
109
- "resolved as `config.ENSEMBLE_MODE if value is None`, so writing "
110
- "`false` is not a restatement of the default but a decision to "
111
- "outrank whatever config.py says later",
109
+ "an unset value defers to config.yaml's own ensemble_mode, so "
110
+ "writing `false` is not a restatement of the default but a "
111
+ "decision to outrank whatever config.yaml says later",
112
112
  "compaction.trigger_tokens":
113
113
  "_trigger_is_configured() is a PRESENCE test: the window-derived "
114
114
  "trigger only applies while no human has named a number. Writing "
@@ -117,7 +117,7 @@ OMITTED = {
117
117
  "much earlier fold with nothing on screen to explain it",
118
118
  "research.subagent_review":
119
119
  "same `is not None` shape as ensemble_mode: an explicit false "
120
- "beats config.SUBAGENT_REVIEW rather than deferring to it",
120
+ "beats config.yaml's subagent_review rather than deferring to it",
121
121
  "tui.theme":
122
122
  "no shipped default to write -- absent, _startup_theme() resolves "
123
123
  "one -- and an invented value would outrank a theme the user "
package/pyproject.toml CHANGED
@@ -18,7 +18,7 @@ build-backend = "setuptools.build_meta"
18
18
 
19
19
  [project]
20
20
  name = "venastine-research-harness"
21
- version = "0.3.13"
21
+ version = "0.3.15"
22
22
  description = "A from-scratch agentic harness: tool-using chat plus a ten-pass self-critiquing research pipeline."
23
23
  readme = "README.md"
24
24
  # The same words as package.json's `keywords`, in the same order: what a
@@ -19,9 +19,10 @@ the precedent: `tools/builtin/_math_common.py` exists because "an
19
19
  expression string is code", and its docstring lists the escapes that were
20
20
  found before the AST allowlist went in.
21
21
 
22
- So the posture is bound once and frozen (UN1). The `config.py` constants
23
- are still where a HUMAN writes the value; they simply stop being what is
24
- read at the moment a decision is made.
22
+ So the posture is bound once and frozen (UN1). `config.yaml` is still
23
+ where a HUMAN writes the value, and the `config.py` constants that
24
+ publish it simply stop being what is read at the moment a decision is
25
+ made.
25
26
 
26
27
  WHAT THE GUARANTEE IS, AND IS NOT
27
28
  ---------------------------------
@@ -254,7 +255,7 @@ class Posture:
254
255
  if self.shell_approval_mode == "never":
255
256
  reasons.append((
256
257
  "shell: never ask",
257
- "SHELL_APPROVAL_MODE is 'never' -- no shell command is "
258
+ "shell_approval_mode is 'never' -- no shell command is "
258
259
  "ever asked about, including one that reads outside the "
259
260
  "workspace on the host"))
260
261
  elif self.shell_approval_mode == "contained":
@@ -264,7 +265,7 @@ class Posture:
264
265
  # holds a `.venastine/` that code can read it.
265
266
  reasons.append((
266
267
  "shell: contained",
267
- "SHELL_APPROVAL_MODE is 'contained' -- a command that runs "
268
+ "shell_approval_mode is 'contained' -- a command that runs "
268
269
  "in the container without network runs without asking, "
269
270
  "including code the agent wrote itself; host reads and "
270
271
  "network commands still ask"))
@@ -279,7 +280,7 @@ class Posture:
279
280
  "asks, as under tiered")
280
281
  reasons.append((
281
282
  "shell: classified",
282
- "SHELL_APPROVAL_MODE is 'classified' -- a decision model "
283
+ "shell_approval_mode is 'classified' -- a decision model "
283
284
  f"({who}) decides whether a single command that runs in the "
284
285
  "container, and that tiered would ask about, runs without "
285
286
  "asking; compound commands, host reads, the host fallback, "
@@ -293,14 +294,14 @@ class Posture:
293
294
  and self.shell_approval_mode != "always"):
294
295
  reasons.append((
295
296
  "host shell, no ask",
296
- "ALLOW_INSECURE_SANDBOX_FALLBACK and "
297
- "AUTO_APPROVE_SANDBOX_FALLBACK are both on -- unprompted "
297
+ "allow_insecure_sandbox_fallback and "
298
+ "auto_approve_sandbox_fallback are both on -- unprompted "
298
299
  "arbitrary code on the host, with your own file access, "
299
300
  "including writes into this harness's own source"))
300
301
  elif self.allow_insecure_fallback:
301
302
  reasons.append((
302
303
  "host shell fallback",
303
- "ALLOW_INSECURE_SANDBOX_FALLBACK is on -- if no container "
304
+ "allow_insecure_sandbox_fallback is on -- if no container "
304
305
  "runtime (Docker or Podman) can run the sandbox, commands "
305
306
  "run on the host with no filesystem or network isolation"))
306
307
  # ROADMAP_v3 §49 slice 3 (SS41). Reported when it is ENABLED rather
@@ -321,7 +322,7 @@ class Posture:
321
322
  # this is §48 (CE6)'s correction applied to a second pair.
322
323
  if (self.auto_approve_fallback
323
324
  and self.shell_approval_mode not in ("always", "never")):
324
- asking = ("AUTO_APPROVE_SANDBOX_FALLBACK is on as well, so "
325
+ asking = ("auto_approve_sandbox_fallback is on as well, so "
325
326
  "NOTHING there asks first -- that opt-in answers "
326
327
  "for the whole backend, not only for the host "
327
328
  "fallback it is named after. Running as root is "
@@ -332,7 +333,7 @@ class Posture:
332
333
  "runs as root always asks")
333
334
  reasons.append((
334
335
  "wsl backend",
335
- "ALLOW_WSL_BACKEND is on -- the agent can ask for a command "
336
+ "allow_wsl_backend is on -- the agent can ask for a command "
336
337
  "to run in WSL, which is not a sandbox: it reaches this "
337
338
  "harness's own config and authority files, has no memory or "
338
339
  "process limits, and can start Windows programs with your "
@@ -345,7 +346,7 @@ class Posture:
345
346
  if self.allow_ssh_backend:
346
347
  reasons.append((
347
348
  "ssh backend",
348
- "ALLOW_SSH_BACKEND is on -- the agent can ask for a command "
349
+ "allow_ssh_backend is on -- the agent can ask for a command "
349
350
  "to run on the configured remote host, which this harness "
350
351
  "cannot inspect: no isolation, no memory or process limits, "
351
352
  "and credentials it did not issue. Nothing runs there "
@@ -355,7 +356,7 @@ class Posture:
355
356
  if not self.redact_tool_outputs:
356
357
  reasons.append((
357
358
  "no redaction",
358
- "REDACT_TOOL_OUTPUTS is off -- credential-shaped strings "
359
+ "redact_tool_outputs is off -- credential-shaped strings "
359
360
  "in tool results are passed through unredacted"))
360
361
  elif self.redact_off_env:
361
362
  reasons.append((
@@ -3365,7 +3365,7 @@ def _unavailable_message(backend: str = BACKEND_CONTAINER) -> str:
3365
3365
  return (
3366
3366
  "No container runtime can run the sandbox"
3367
3367
  + (f" ({reason})" if reason else "")
3368
- + " and ALLOW_INSECURE_SANDBOX_FALLBACK is False. Install Docker or "
3368
+ + " and allow_insecure_sandbox_fallback is false. Install Docker or "
3369
3369
  "Podman for strong isolation, or set "
3370
3370
  "allow_insecure_sandbox_fallback: true in config.yaml to use the "
3371
3371
  "weaker subprocess fallback (no filesystem isolation, no network "
package/storage.py CHANGED
@@ -660,6 +660,36 @@ def _decode_harness(raw):
660
660
  return value if isinstance(value, dict) and value.get("kind") else None
661
661
 
662
662
 
663
+ def starts_turn(row: dict) -> bool:
664
+ """Whether this row OPENS a turn: a user row, unless the harness wrote it
665
+ inside one (ROADMAP_v3 §95, RO3).
666
+
667
+ A turn starts at a user message -- the wrapper adds one, then the loop
668
+ appends assistant and tool rows until it stops -- and six readers were
669
+ written on that rule: the turn list behind `completed_turns`, the fold
670
+ boundaries and the pin cap in core/compaction.py, the request a
671
+ mid-turn watermark re-adds, and the row a failure is recorded on, here
672
+ and in core/memory.py. Since §95 the loop can write a user row
673
+ INSIDE a turn, to ask the model to follow a reasoning-only reply, and
674
+ such a row carries `mid_turn: true` in its harness record. Read as a
675
+ turn start, it would move the current-turn floor past the turn's own
676
+ earlier steps while the turn is still running -- the floor the mid-turn
677
+ compaction valve calls safe by construction -- and a mid-turn
678
+ watermark would re-add it in place of the person's request.
679
+
680
+ Takes either shape: an `_ordered_rows` dict, whose `harness` is the
681
+ stored JSON string, or a neutral message, whose `harness` is decoded.
682
+ `is True`, A29's rule: a record carrying a quoted "true" still opens a
683
+ turn, as every harness row did before §95.
684
+ """
685
+ if row.get("role") != "user":
686
+ return False
687
+ harness = row.get("harness")
688
+ if isinstance(harness, str):
689
+ harness = _decode_harness(harness)
690
+ return not (isinstance(harness, dict) and harness.get("mid_turn") is True)
691
+
692
+
663
693
  def _decode_images(raw):
664
694
  """A tool row's stored images (ROADMAP_v3 §81, IM2), or None.
665
695
 
@@ -874,10 +904,10 @@ def _turn_opener_index(rows: list[dict], span: list[dict]) -> list[int]:
874
904
  """
875
905
  if not span or len(span) >= len(rows):
876
906
  return []
877
- if rows[len(span)]["role"] == "user":
907
+ if starts_turn(rows[len(span)]):
878
908
  return []
879
909
  for index in range(len(span) - 1, -1, -1):
880
- if span[index]["role"] == "user":
910
+ if starts_turn(span[index]):
881
911
  return [index]
882
912
  return []
883
913
 
@@ -995,13 +1025,14 @@ def turn_start_ids(thread_id: UUID) -> list[UUID]:
995
1025
  """Ids of the rows that BEGIN each turn, oldest first.
996
1026
 
997
1027
  A turn starts at a user message: the wrapper adds one, then the loop
998
- appends assistant and tool rows until it stops. §21's M4 needs this
1028
+ appends assistant and tool rows until it stops. `starts_turn` says
1029
+ which user rows, since §95 wrote one inside a turn. §21's M4 needs this
999
1030
  because a fold boundary that lands anywhere else can separate an
1000
1031
  assistant message from its tool results, and a tool_result with no
1001
1032
  preceding tool_use is an HTTP 400 on Anthropic -- a hard wire error,
1002
1033
  not a degraded answer. M5's "keep the last N turns" needs it too.
1003
1034
  """
1004
- return [m["id"] for m in _ordered_rows(thread_id) if m["role"] == "user"]
1035
+ return [m["id"] for m in _ordered_rows(thread_id) if starts_turn(m)]
1005
1036
 
1006
1037
 
1007
1038
  def message_ids_from(thread_id: UUID, message_id: UUID) -> list[UUID]:
@@ -1047,6 +1078,10 @@ def mark_turn_failed(thread_id: UUID, error: str) -> bool:
1047
1078
 
1048
1079
  Returns False for a thread with no user row. Does not stamp
1049
1080
  last_activity_at the way save_message does: a failure is not a message.
1081
+
1082
+ The newest row that STARTS a turn (ROADMAP_v3 §95, RO3), which skips a
1083
+ harness row the loop wrote inside the run -- the run did not start
1084
+ there, and replay draws the failure under the request that did.
1050
1085
  """
1051
1086
  with Session(engine) as session:
1052
1087
  statement = (
@@ -1055,7 +1090,9 @@ def mark_turn_failed(thread_id: UUID, error: str) -> bool:
1055
1090
  .where(MessageLog.role == "user")
1056
1091
  .order_by(MessageLog.created_at.desc(), MessageLog.id.desc())
1057
1092
  )
1058
- row = session.exec(statement).first()
1093
+ row = next((candidate for candidate in session.exec(statement)
1094
+ if starts_turn({"role": candidate.role,
1095
+ "harness": candidate.harness})), None)
1059
1096
  if row is None:
1060
1097
  return False
1061
1098
  row.error = error
@@ -281,7 +281,7 @@ def _read_rich(resolved_path: str) -> str:
281
281
  consumer, it sits behind `read`, and `read` is globally denied and
282
282
  cannot be re-enabled at runtime -- so a default install carries neither
283
283
  it nor its eleven transitive distributions. Anyone who reaches here has
284
- enabled `read` in config.py, and the bare ModuleNotFoundError would name
284
+ enabled `read` in config.yaml, and the bare ModuleNotFoundError would name
285
285
  a package they never asked for and cannot place.
286
286
  """
287
287
  try:
@@ -144,7 +144,7 @@ logger = logging.getLogger(__name__)
144
144
  # read the posture at import time**, pinned by
145
145
  # test_posture.py::test_no_module_reads_the_posture_at_import.
146
146
  capability.validate_mode(config.SHELL_APPROVAL_MODE,
147
- "config.SHELL_APPROVAL_MODE")
147
+ "shell_approval_mode in config.yaml")
148
148
 
149
149
  # Compute the shell name once at import time to avoid blocking on
150
150
  # subprocess probes every time the schema is accessed. On Windows this
@@ -625,7 +625,7 @@ def _shell_approval_check(tool_name: str, params: dict) -> bool:
625
625
  """
626
626
  active = posture.current()
627
627
  mode = capability.validate_mode(active.shell_approval_mode,
628
- "config.SHELL_APPROVAL_MODE")
628
+ "shell_approval_mode in config.yaml")
629
629
  if mode == capability.ALWAYS:
630
630
  return True
631
631
  if mode == capability.NEVER:
@@ -878,7 +878,7 @@ def _interactive_approval_check(tool_name: str, params: dict) -> bool:
878
878
  """
879
879
  active = posture.current()
880
880
  mode = capability.validate_mode(active.shell_approval_mode,
881
- "config.SHELL_APPROVAL_MODE")
881
+ "shell_approval_mode in config.yaml")
882
882
  if mode == capability.ALWAYS:
883
883
  return True
884
884
  if mode == capability.NEVER:
@@ -1055,8 +1055,8 @@ def fallback_changed_refusal(command: str, docker_up: bool) -> str | None:
1055
1055
  if active.allow_insecure_fallback and not active.auto_approve_fallback:
1056
1056
  return ("Docker became unavailable after the approval check. "
1057
1057
  "The insecure subprocess fallback requires explicit "
1058
- "per-run approval (AUTO_APPROVE_SANDBOX_FALLBACK is "
1059
- "False). Please retry — the approval prompt will "
1058
+ "per-run approval (auto_approve_sandbox_fallback is "
1059
+ "false in config.yaml). Please retry — the approval prompt will "
1060
1060
  "appear this time.")
1061
1061
  return None
1062
1062
 
@@ -455,7 +455,7 @@ def input_approval_check(tool_name: str, params: dict) -> bool:
455
455
  return False
456
456
  active = posture.current()
457
457
  mode = capability.validate_mode(active.shell_approval_mode,
458
- "config.SHELL_APPROVAL_MODE")
458
+ "shell_approval_mode in config.yaml")
459
459
  if mode == capability.ALWAYS:
460
460
  return True
461
461
  if mode == capability.NEVER:
package/tui/app.py CHANGED
@@ -71,6 +71,7 @@ from core.file_mentions import FileIndex, active_mention
71
71
  from core.loop import (
72
72
  DEFAULT_PROVIDER,
73
73
  DEFAULT_SYSTEM_PROMPT,
74
+ HARNESS_REASONING_ONLY,
74
75
  RunAgentLoop,
75
76
  advertisement_facts,
76
77
  agent_turn,
@@ -943,7 +944,7 @@ class VenastineApp(App):
943
944
  self._animations = tui_settings.get("animations", True)
944
945
  # §38 (O6). Render a turn's reasoning inline, or collapse it to the
945
946
  # animated one-line indicator. Defaulted HERE rather than in
946
- # config.py, following `animations` immediately above: config.py
947
+ # config.yaml, following `animations` immediately above: config.yaml
947
948
  # holds no TUI values and is plain-values-only. /thinking flips it
948
949
  # for the session and persists nothing, matching /effort.
949
950
  self._show_thinking = tui_settings.get("show_thinking", True)
@@ -1543,7 +1544,7 @@ class VenastineApp(App):
1543
1544
  from the trace.
1544
1545
 
1545
1546
  "(remembered)" marks a pair restored from the preference store
1546
- rather than resolved from settings.json or config.py (§43, RM3).
1547
+ rather than resolved from settings.json or config.yaml (§43, RM3).
1547
1548
  #138's rule: a launch should not have to be guessed at, and a
1548
1549
  session silently coming up on a model no config file names is
1549
1550
  exactly the confusion this feature would otherwise trade for the
@@ -3522,6 +3523,13 @@ class VenastineApp(App):
3522
3523
  elif notice["kind"] in ("compaction_failed",
3523
3524
  "compaction_blocked"):
3524
3525
  transcript.write_error(f"— {notice['text']} —")
3526
+ elif notice["kind"] == HARNESS_REASONING_ONLY:
3527
+ # ROADMAP_v3 §95 (RO5). The harness wrote a row asking the
3528
+ # model to follow its reasoning-only reply, and this notice's
3529
+ # text IS that row's first line -- so it is drawn the way
3530
+ # replay draws the row, and a resumed thread shows the line
3531
+ # the live turn showed, in the same role.
3532
+ transcript.write_role("wake", notice["text"])
3525
3533
  else:
3526
3534
  transcript.write_system(f"— {notice['text']} —")
3527
3535
 
@@ -6120,7 +6128,7 @@ def _cmd_window(app: VenastineApp, args: str) -> None:
6120
6128
  # the best reason to have this command -- a gateway or a self-hosted
6121
6129
  # endpoint can report its own default rather than the deployment's
6122
6130
  # real window -- so refusing would block precisely the case it exists
6123
- # for. The risk is named in the same terms config.py uses for the qwen
6131
+ # for. The risk is named in the same terms config.yaml uses for the qwen
6124
6132
  # entry: erring high means a hard context-limit error mid-run.
6125
6133
  reported = _reported_window(app)
6126
6134
  if reported and tokens > reported:
@@ -6289,9 +6297,10 @@ def _report_override(app: VenastineApp, key: str) -> None:
6289
6297
  elif _reported_window(app) is not None:
6290
6298
  source = f"reported by {app.provider_name}"
6291
6299
  elif compaction._normalized(app.model) in config.MODEL_CONTEXT_WINDOWS:
6292
- source = "config.MODEL_CONTEXT_WINDOWS"
6300
+ source = "model_context_windows in config.yaml"
6293
6301
  else:
6294
- source = "config.DEFAULT_CONTEXT_WINDOW (no entry for this model)"
6302
+ source = ("default_context_window in config.yaml "
6303
+ "(no entry for this model)")
6295
6304
  app._transcript.write_system(
6296
6305
  f"Context window: {effective:,} — {source}.")
6297
6306
  app._transcript.write_system(
@@ -6646,7 +6655,7 @@ def _clear_pipeline_role(app: VenastineApp, role: str, label: str) -> None:
6646
6655
  remaining = pipeline_models.resolve(role)
6647
6656
  if remaining:
6648
6657
  app._transcript.write_system(
6649
- f"config.py still sets {label}: {remaining['provider_name']} | "
6658
+ f"config.yaml still sets {label}: {remaining['provider_name']} | "
6650
6659
  f"{remaining['model']}.")
6651
6660
 
6652
6661