@yahyahammad/venastine 0.2.0 → 0.3.9
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/.env.example +9 -4
- package/CONFIG_ARCHITECTURE.md +2030 -0
- package/LICENSE +201 -201
- package/NOTICE +1 -1
- package/README.md +1034 -895
- package/agents/builtin/build.md +73 -0
- package/agents/builtin/compactor.md +75 -75
- package/agents/builtin/explore.md +77 -77
- package/agents/builtin/general.md +80 -0
- package/agents/builtin/pipeline-reviewer.md +100 -100
- package/agents/builtin/plan.md +106 -88
- package/agents/builtin/review.md +74 -74
- package/agents/builtin/test.md +65 -0
- package/agents/builtin/writer.md +67 -0
- package/agents/manager.py +68 -2
- package/agents/subagent_tool.py +472 -238
- package/agents/tui_commands.py +238 -151
- package/bin/venastine.mjs +150 -4
- package/bin/venastine.py +71 -0
- package/cli_commands.py +838 -0
- package/config.py +95 -1198
- package/config.yaml +1103 -0
- package/config_edit.py +1301 -0
- package/config_schema.py +971 -0
- package/config_update.py +749 -0
- package/core/agent_activity.py +256 -118
- package/core/ansi.py +452 -0
- package/core/approval.py +137 -115
- package/core/call_view.py +364 -0
- package/core/cancel.py +121 -0
- package/core/chatgpt_auth.py +747 -0
- package/core/claims_view.py +107 -0
- package/core/client.py +2715 -1559
- package/core/commands.py +460 -0
- package/core/compaction.py +1735 -971
- package/core/config_loader.py +1982 -1563
- package/core/connect.py +686 -0
- package/core/events.py +123 -61
- package/core/file_mentions.py +280 -0
- package/core/interaction.py +411 -328
- package/core/line_pattern.py +128 -0
- package/core/loop.py +2566 -1490
- package/core/memory.py +482 -370
- package/core/pipeline_models.py +190 -176
- package/core/provider_errors.py +234 -0
- package/core/reasoning/authorization.py +130 -130
- package/core/reasoning/base.py +219 -219
- package/core/reasoning/confidence_scoring.py +484 -468
- package/core/reasoning/events.py +149 -149
- package/core/reasoning/json_retry.py +19 -1
- package/core/reasoning/orchestrator.py +1561 -1382
- package/core/reasoning/output_writer.py +345 -286
- package/core/reasoning/payload_validation.py +524 -398
- package/core/reasoning/pipeline_storage.py +88 -22
- package/core/reasoning/review.py +659 -615
- package/core/reasoning/scholar.py +458 -434
- package/core/reasoning/source_corpus.py +433 -327
- package/core/reasoning/source_scoring.py +855 -855
- package/core/replay.py +315 -183
- package/core/session.py +36 -12
- package/core/session_view.py +259 -0
- package/core/session_wake.py +259 -0
- package/core/shell_sessions.py +1553 -0
- package/core/workspace_trust.py +1 -1
- package/credentials.py +148 -20
- package/database.py +333 -252
- package/{PRIVACY.md → docs/PRIVACY.md} +78 -66
- package/docs/SECURITY.md +108 -0
- package/docs/THIRD_PARTY_NOTICES.md +148 -0
- package/env_secrets.py +50 -9
- package/json_store.py +343 -123
- package/logging_setup.py +295 -185
- package/main.py +3433 -1961
- package/mcp_client/client.py +733 -499
- package/mcp_client/config.py +126 -21
- package/mcp_client/registration.py +343 -182
- package/memories/manager.py +147 -132
- package/memories/tui_commands.py +319 -244
- package/package.json +303 -73
- package/project_init/config_files.py +282 -278
- package/project_init/generator.py +728 -564
- package/project_init/manifest.py +18 -6
- package/project_init/tui_commands.py +147 -144
- package/prompts/critic_pass.md +24 -24
- package/prompts/revalidate.md +25 -25
- package/prompts/revise.md +2 -0
- package/prompts/source_grounding.md +66 -66
- package/prompts/system_prompts.py +316 -262
- package/providers.json.example +4 -2
- package/pyproject.toml +149 -100
- package/requirements.txt +295 -72
- package/safety/policy_enforcement.py +1082 -746
- package/security/askpass.py +419 -0
- package/security/capability.py +207 -177
- package/security/confine.py +484 -0
- package/security/hosted.py +182 -0
- package/security/permissions.py +237 -190
- package/security/posture.py +124 -12
- package/security/protected_paths.py +501 -309
- package/security/sandbox.py +4108 -985
- package/security/secrets.py +855 -0
- package/skills/builtin/math/formal-specification.md +67 -67
- package/skills/builtin/math/numerical-methods.md +66 -66
- package/skills/builtin/research/experiment-design.md +62 -62
- package/skills/builtin/research/reproducibility.md +55 -55
- package/skills/builtin/research/source-evaluation.md +70 -70
- package/skills/builtin/research/statistical-inference.md +63 -63
- package/skills/builtin/research/technical-writing.md +59 -59
- package/skills/builtin/software/code-review.md +61 -61
- package/skills/builtin/software/debugging.md +61 -61
- package/skills/builtin/software/test-design.md +67 -67
- package/skills/tui_commands.py +110 -111
- package/storage.py +1338 -1005
- package/tools/base.py +161 -1
- package/tools/builtin/_math_common.py +667 -377
- package/tools/builtin/_net_common.py +225 -19
- package/tools/builtin/arxiv.py +497 -39
- package/tools/builtin/ask_user.py +231 -213
- package/tools/builtin/discrete_math.py +35 -4
- package/tools/builtin/fetch_url.py +371 -153
- package/tools/builtin/file_ops.py +720 -412
- package/tools/builtin/geometry.py +48 -12
- package/tools/builtin/get_time.py +2 -2
- package/tools/builtin/linear_algebra.py +66 -23
- package/tools/builtin/load_skill.py +76 -66
- package/tools/builtin/logic.py +171 -12
- package/tools/builtin/pin.py +227 -193
- package/tools/builtin/probability_stats.py +12 -16
- package/tools/builtin/project_docs.py +464 -405
- package/tools/builtin/remember.py +175 -158
- package/tools/builtin/shell.py +980 -461
- package/tools/builtin/shell_sessions.py +613 -0
- package/tools/builtin/symbolic_math.py +80 -10
- package/tools/builtin/todo.py +149 -147
- package/tools/builtin/web_search.py +191 -13
- package/tools/context.py +159 -136
- package/tools/isolation.py +107 -34
- package/tools/registry.py +1242 -925
- package/tui/app.py +8318 -4100
- package/tui/app.tcss +747 -556
- package/tui/commands.py +26 -170
- package/tui/diffs.py +34 -13
- package/tui/history.py +128 -128
- package/tui/hosted_driver.py +102 -0
- package/tui/markdown.py +1329 -1279
- package/tui/meters.py +364 -268
- package/tui/ravens.py +146 -146
- package/tui/screens.py +1255 -943
- package/tui/themes.py +599 -559
- package/tui/widgets.py +3813 -2332
- package/SECURITY.md +0 -88
- package/THIRD_PARTY_NOTICES.md +0 -100
package/core/pipeline_models.py
CHANGED
|
@@ -1,176 +1,190 @@
|
|
|
1
|
-
"""
|
|
2
|
-
core/pipeline_models.py
|
|
3
|
-
|
|
4
|
-
The
|
|
5
|
-
|
|
6
|
-
ROADMAP_v2 §45 (SQ7). A research run reaches for three models: the
|
|
7
|
-
generator (the session pair, `/model`), the CRITIC that grounds and
|
|
8
|
-
critiques its claims (§11), and — since §45 — the EMBEDDER that scores how
|
|
9
|
-
close a source is to the claim it was cited for. Only the first has ever
|
|
10
|
-
been reachable without editing installed source. `CRITIC_MODEL` has sat in
|
|
11
|
-
`config.py` since §11, which means the routing this project's own
|
|
12
|
-
methodology calls essential ("a model checking its own output for errors
|
|
13
|
-
shares that model's blind spots") was available only to whoever could edit
|
|
14
|
-
the checkout.
|
|
15
|
-
|
|
16
|
-
IN `core/`, NOT `tui/preferences.py`, AND THAT IS A LAYERING FACT. The
|
|
17
|
-
pipeline reads these and D12 makes the TUI a shell that core must never
|
|
18
|
-
import — WS6's argument for `core/model_windows.py`, unchanged, and §25's
|
|
19
|
-
lesson about a capability the pipeline could not reach. It is equally not
|
|
20
|
-
`core/session.py`, whose contract is that nothing there reaches disk.
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
RM4's reason: a writer serialising only its own role would
|
|
24
|
-
the
|
|
25
|
-
|
|
26
|
-
which part in a run — and `model_windows.py` is next door as the example
|
|
27
|
-
of when a second file is right instead (it holds a different kind of fact,
|
|
28
|
-
keyed on a pair rather than naming one).
|
|
29
|
-
|
|
30
|
-
PRECEDENCE: THE STORE OUTRANKS `config.py`. The opposite of
|
|
31
|
-
`model_windows`, and for the opposite reason — there, a configured value
|
|
32
|
-
could be confidently wrong about a deployment, so the person won. Here
|
|
33
|
-
`config.py` is the harness's own default and the store is a choice
|
|
34
|
-
someone made at a command prompt for this account, which is `/model`
|
|
35
|
-
outranking `config.MODEL_NAME` exactly.
|
|
36
|
-
|
|
37
|
-
WHY NOT `settings.json`. Choosing a provider is a grant (E2): a project's
|
|
38
|
-
settings.json outranks the user's, so a cloned repo could otherwise point
|
|
39
|
-
this harness's embedding calls — claim text and fetched page text — at a
|
|
40
|
-
provider of its choosing. `embedder_model` and `critic_model` are
|
|
41
|
-
therefore rejected there BY NAME, like `ensemble_models`, and this
|
|
42
|
-
user-tier store is the route instead.
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
from
|
|
51
|
-
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
#
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
102
|
-
|
|
103
|
-
|
|
104
|
-
"""
|
|
105
|
-
|
|
106
|
-
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
|
|
142
|
-
|
|
143
|
-
|
|
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
|
-
|
|
1
|
+
"""
|
|
2
|
+
core/pipeline_models.py
|
|
3
|
+
|
|
4
|
+
The model roles that are not the session model, remembered.
|
|
5
|
+
|
|
6
|
+
ROADMAP_v2 §45 (SQ7). A research run reaches for three models: the
|
|
7
|
+
generator (the session pair, `/model`), the CRITIC that grounds and
|
|
8
|
+
critiques its claims (§11), and — since §45 — the EMBEDDER that scores how
|
|
9
|
+
close a source is to the claim it was cited for. Only the first has ever
|
|
10
|
+
been reachable without editing installed source. `CRITIC_MODEL` has sat in
|
|
11
|
+
`config.py` since §11, which means the routing this project's own
|
|
12
|
+
methodology calls essential ("a model checking its own output for errors
|
|
13
|
+
shares that model's blind spots") was available only to whoever could edit
|
|
14
|
+
the checkout.
|
|
15
|
+
|
|
16
|
+
IN `core/`, NOT `tui/preferences.py`, AND THAT IS A LAYERING FACT. The
|
|
17
|
+
pipeline reads these and D12 makes the TUI a shell that core must never
|
|
18
|
+
import — WS6's argument for `core/model_windows.py`, unchanged, and §25's
|
|
19
|
+
lesson about a capability the pipeline could not reach. It is equally not
|
|
20
|
+
`core/session.py`, whose contract is that nothing there reaches disk.
|
|
21
|
+
|
|
22
|
+
EVERY ROLE IN ONE FILE, so every writer reads, modifies and writes. RM4's
|
|
23
|
+
shape and RM4's reason: a writer serialising only its own role would
|
|
24
|
+
silently drop the others', each correct in its own test. They are one file
|
|
25
|
+
rather than a module each because they are the same KIND of fact — which model plays
|
|
26
|
+
which part in a run — and `model_windows.py` is next door as the example
|
|
27
|
+
of when a second file is right instead (it holds a different kind of fact,
|
|
28
|
+
keyed on a pair rather than naming one).
|
|
29
|
+
|
|
30
|
+
PRECEDENCE: THE STORE OUTRANKS `config.py`. The opposite of
|
|
31
|
+
`model_windows`, and for the opposite reason — there, a configured value
|
|
32
|
+
could be confidently wrong about a deployment, so the person won. Here
|
|
33
|
+
`config.py` is the harness's own default and the store is a choice
|
|
34
|
+
someone made at a command prompt for this account, which is `/model`
|
|
35
|
+
outranking `config.MODEL_NAME` exactly.
|
|
36
|
+
|
|
37
|
+
WHY NOT `settings.json`. Choosing a provider is a grant (E2): a project's
|
|
38
|
+
settings.json outranks the user's, so a cloned repo could otherwise point
|
|
39
|
+
this harness's embedding calls — claim text and fetched page text — at a
|
|
40
|
+
provider of its choosing. `embedder_model` and `critic_model` are
|
|
41
|
+
therefore rejected there BY NAME, like `ensemble_models`, and this
|
|
42
|
+
user-tier store is the route instead.
|
|
43
|
+
|
|
44
|
+
A THIRD ROLE SINCE ROADMAP_v3 §57: `subagent`, set by `/submodel` in
|
|
45
|
+
either shell. It is the model a SPAWNED child runs on when its own file
|
|
46
|
+
pins neither half of a pair (SM6), at every depth and whatever its parent
|
|
47
|
+
runs on (SM7) -- so a cheap child no longer has to run on its parent's
|
|
48
|
+
expensive model. It is the same kind of fact as the other two, which part
|
|
49
|
+
a model plays, and so it shares their file and their read-modify-write.
|
|
50
|
+
It has NO `config.yaml` tier (SM5): `resolve` answers from this store or
|
|
51
|
+
not at all, which is why its `_FALLBACK_ATTR` entry is None rather than a
|
|
52
|
+
constant nobody sets. Nothing else reads it: the compactor, the pipeline
|
|
53
|
+
reviewer and the initializer keep the session model (SM8).
|
|
54
|
+
"""
|
|
55
|
+
|
|
56
|
+
import logging
|
|
57
|
+
import os
|
|
58
|
+
from typing import Optional
|
|
59
|
+
|
|
60
|
+
import config
|
|
61
|
+
from json_store import read_versioned, write_json_atomic
|
|
62
|
+
|
|
63
|
+
logger = logging.getLogger(__name__)
|
|
64
|
+
|
|
65
|
+
# Bumped when the SHAPE changes. An unknown version reads as an empty
|
|
66
|
+
# store, which costs the remembered roles and falls back to config.py --
|
|
67
|
+
# there is nothing here that was consented to, so the MCP store's careful
|
|
68
|
+
# legacy handling has no counterpart.
|
|
69
|
+
STORE_VERSION = 1
|
|
70
|
+
|
|
71
|
+
# The roles this store knows. A role outside it is refused rather than
|
|
72
|
+
# stored: a typo'd `/embeder` writing a key nothing reads would look like
|
|
73
|
+
# it worked and change nothing, which is the failure mode this whole
|
|
74
|
+
# module exists to remove for CRITIC_MODEL.
|
|
75
|
+
ROLES = ("critic", "embedder", "subagent")
|
|
76
|
+
|
|
77
|
+
# Which config.py constant answers when the store is silent. None means
|
|
78
|
+
# nothing does: `subagent` is remembered or unset (§57, SM5).
|
|
79
|
+
_FALLBACK_ATTR = {"critic": "CRITIC_MODEL", "embedder": "EMBEDDER_MODEL",
|
|
80
|
+
"subagent": None}
|
|
81
|
+
|
|
82
|
+
|
|
83
|
+
def store_path() -> str:
|
|
84
|
+
# Call time, not import time: tests redirect it, and a moved config
|
|
85
|
+
# directory is picked up without a restart. Every sibling store
|
|
86
|
+
# resolves theirs the same way.
|
|
87
|
+
return os.path.expanduser("~/.config/venastine/pipeline_models.json")
|
|
88
|
+
|
|
89
|
+
|
|
90
|
+
def _read_raw() -> dict:
|
|
91
|
+
data = read_versioned(store_path(), STORE_VERSION,
|
|
92
|
+
"falling back to config.py for pipeline models.")
|
|
93
|
+
roles = data.get("roles")
|
|
94
|
+
return roles if isinstance(roles, dict) else {}
|
|
95
|
+
|
|
96
|
+
|
|
97
|
+
def _write(roles: dict, action: str) -> bool:
|
|
98
|
+
"""Persist the whole map. True if it reached disk.
|
|
99
|
+
|
|
100
|
+
A failure WARNS and returns False rather than raising, matching
|
|
101
|
+
model_windows: the TUI routes WARNING+ into the transcript, so someone
|
|
102
|
+
whose choice could not be saved is told at the moment they make it
|
|
103
|
+
rather than discovering it at the next launch.
|
|
104
|
+
"""
|
|
105
|
+
path = store_path()
|
|
106
|
+
try:
|
|
107
|
+
write_json_atomic(path, {"version": STORE_VERSION, "roles": roles},
|
|
108
|
+
sort_keys=True)
|
|
109
|
+
except OSError as exc:
|
|
110
|
+
logger.warning("Could not %s the pipeline model in %s (%s).",
|
|
111
|
+
action, path, exc)
|
|
112
|
+
return False
|
|
113
|
+
return True
|
|
114
|
+
|
|
115
|
+
|
|
116
|
+
def _valid(entry) -> bool:
|
|
117
|
+
"""A whole pair or nothing.
|
|
118
|
+
|
|
119
|
+
tui/preferences.py's rule, for its reason: a model name means nothing
|
|
120
|
+
against the wrong provider, so half a record is no record. A
|
|
121
|
+
hand-edited file with only one of the two is ignored rather than
|
|
122
|
+
half-applied.
|
|
123
|
+
"""
|
|
124
|
+
return (isinstance(entry, dict)
|
|
125
|
+
and isinstance(entry.get("provider_name"), str)
|
|
126
|
+
and isinstance(entry.get("model"), str)
|
|
127
|
+
and bool(entry["provider_name"]) and bool(entry["model"]))
|
|
128
|
+
|
|
129
|
+
|
|
130
|
+
def remembered(role: str) -> Optional[dict]:
|
|
131
|
+
"""What the user chose for `role`, or None. Ignores config.py."""
|
|
132
|
+
entry = _read_raw().get(role)
|
|
133
|
+
if not _valid(entry):
|
|
134
|
+
return None
|
|
135
|
+
return {"provider_name": entry["provider_name"], "model": entry["model"]}
|
|
136
|
+
|
|
137
|
+
|
|
138
|
+
def resolve(role: str) -> Optional[dict]:
|
|
139
|
+
"""The pair actually in force for `role`: the store, else config.py,
|
|
140
|
+
else None.
|
|
141
|
+
|
|
142
|
+
THE ONE READER FOR THE PIPELINE. `orchestrator` and every shell ask
|
|
143
|
+
this rather than reading `config.CRITIC_MODEL` directly, so a
|
|
144
|
+
remembered choice reaches the CLI, the TUI and a spawned run alike --
|
|
145
|
+
and so there is one precedence rule rather than one per caller.
|
|
146
|
+
"""
|
|
147
|
+
if role not in ROLES:
|
|
148
|
+
raise ValueError(f"unknown pipeline model role {role!r}; "
|
|
149
|
+
f"expected one of {', '.join(ROLES)}")
|
|
150
|
+
chosen = remembered(role)
|
|
151
|
+
if chosen is not None:
|
|
152
|
+
return chosen
|
|
153
|
+
attr = _FALLBACK_ATTR[role]
|
|
154
|
+
fallback = getattr(config, attr, None) if attr else None
|
|
155
|
+
return fallback if _valid(fallback) else None
|
|
156
|
+
|
|
157
|
+
|
|
158
|
+
def remember(role: str, provider_name: str, model: str) -> bool:
|
|
159
|
+
"""Record a pair for `role`. True if it reached disk."""
|
|
160
|
+
if role not in ROLES:
|
|
161
|
+
raise ValueError(f"unknown pipeline model role {role!r}; "
|
|
162
|
+
f"expected one of {', '.join(ROLES)}")
|
|
163
|
+
roles = _read_raw()
|
|
164
|
+
roles[role] = {"provider_name": provider_name, "model": model}
|
|
165
|
+
return _write(roles, "save")
|
|
166
|
+
|
|
167
|
+
|
|
168
|
+
def forget(role: str) -> str:
|
|
169
|
+
"""Drop `role`'s entry. "removed", "absent" or "failed".
|
|
170
|
+
|
|
171
|
+
Three outcomes rather than a bool, following model_windows: the shell
|
|
172
|
+
says a different sentence for each, and a failed write reported as
|
|
173
|
+
"nothing was set" would be a lie about a file the user can go and
|
|
174
|
+
read. Note that "removed" does not mean the role is unset -- config.py
|
|
175
|
+
may still answer, and the shell says so.
|
|
176
|
+
"""
|
|
177
|
+
if role not in ROLES:
|
|
178
|
+
raise ValueError(f"unknown pipeline model role {role!r}; "
|
|
179
|
+
f"expected one of {', '.join(ROLES)}")
|
|
180
|
+
roles = _read_raw()
|
|
181
|
+
if roles.pop(role, None) is None:
|
|
182
|
+
return "absent"
|
|
183
|
+
return "removed" if _write(roles, "clear") else "failed"
|
|
184
|
+
|
|
185
|
+
|
|
186
|
+
def all_roles() -> dict:
|
|
187
|
+
"""{role: {"provider_name", "model"}} for every role in force, from
|
|
188
|
+
whichever tier answered. For a report."""
|
|
189
|
+
return {role: pair for role in ROLES
|
|
190
|
+
if (pair := resolve(role)) is not None}
|
|
@@ -0,0 +1,234 @@
|
|
|
1
|
+
"""
|
|
2
|
+
core/provider_errors.py
|
|
3
|
+
|
|
4
|
+
Which provider failures are worth another attempt, how long to wait before
|
|
5
|
+
it, and the one line that describes a failure (batch 91).
|
|
6
|
+
|
|
7
|
+
WHY THIS IS NOT IN core/client.py. AGENTS.md's ownership table gives
|
|
8
|
+
client.py one call and provider format translation, and says in so many
|
|
9
|
+
words that it does not own retry. The loop owns the attempt; this module
|
|
10
|
+
answers the two questions the attempt asks -- "is this one worth another
|
|
11
|
+
try?" and "how long do I wait?" -- and holds no state of its own.
|
|
12
|
+
|
|
13
|
+
WHY NO SDK IS IMPORTED, HERE OR AT CALL TIME. core/client.py imports each SDK
|
|
14
|
+
lazily inside api_initialization, so a harness configured for one provider
|
|
15
|
+
never loads the other two -- and the suite installs FAKE openai, anthropic,
|
|
16
|
+
google.genai and httpx modules before any test imports anything (the root
|
|
17
|
+
conftest). An isinstance check against a real SDK class would load what
|
|
18
|
+
client.py deliberately does not, and under test would be comparing against
|
|
19
|
+
a fake. So an exception is read by the class names in its MRO and by the
|
|
20
|
+
attributes the SDKs themselves define: `status_code` (openai / anthropic
|
|
21
|
+
APIStatusError), an int `code` (google.genai errors.APIError), `body` (the
|
|
22
|
+
error object an already-open stream carried) and `response.headers`.
|
|
23
|
+
|
|
24
|
+
WHAT COUNTS AS TRANSIENT (owner decision, batch 91):
|
|
25
|
+
|
|
26
|
+
* a connection failure or a timeout -- openai / anthropic
|
|
27
|
+
APIConnectionError and APITimeoutError, an httpx transport error raised
|
|
28
|
+
raw while a stream is being read (and by google-genai, whose transport
|
|
29
|
+
httpx is since 2.x), a requests ConnectionError / Timeout (google-auth's
|
|
30
|
+
token refresh; google-genai 1.0's transport), and the builtin ConnectionError /
|
|
31
|
+
TimeoutError;
|
|
32
|
+
* HTTP 408, 409, 429 and any 5xx -- the statuses the openai and anthropic
|
|
33
|
+
SDKs already retry for themselves before a stream opens;
|
|
34
|
+
* an error event inside a stream that had ALREADY opened with 200. The
|
|
35
|
+
openai SDK raises its base APIError for it, with no status at all;
|
|
36
|
+
anthropic raises APIStatusError carrying the 200 the stream opened with.
|
|
37
|
+
Neither SDK retries it, and that is the gap this module exists for:
|
|
38
|
+
OpenRouter delivered "The service is temporarily unavailable" and
|
|
39
|
+
"Upstream idle timeout exceeded" exactly this way. Retried unless the
|
|
40
|
+
error object itself names a 4xx code or a permanent error type.
|
|
41
|
+
|
|
42
|
+
Everything else is NOT: a bad request, an auth failure or an unknown model
|
|
43
|
+
fails identically on every attempt, and a harness error -- D21's
|
|
44
|
+
RuntimeError, a ValueError from message translation -- is a bug, not
|
|
45
|
+
weather, and retrying it would only delay the traceback.
|
|
46
|
+
"""
|
|
47
|
+
|
|
48
|
+
from random import random
|
|
49
|
+
from time import sleep
|
|
50
|
+
from typing import Optional
|
|
51
|
+
|
|
52
|
+
#: Beside any 5xx. The same set the openai and anthropic SDKs retry.
|
|
53
|
+
RETRYABLE_STATUSES = frozenset({408, 409, 429})
|
|
54
|
+
|
|
55
|
+
#: A server's retry-after is honoured up to this, as the SDKs do.
|
|
56
|
+
RETRY_AFTER_CAP_S = 60.0
|
|
57
|
+
|
|
58
|
+
#: Up to a quarter of the delay either way, so parallel siblings that failed
|
|
59
|
+
#: together -- three children of one spawn batch, which is the case that
|
|
60
|
+
#: prompted this -- do not all come back in the same instant.
|
|
61
|
+
JITTER = 0.25
|
|
62
|
+
|
|
63
|
+
#: A transcript line, not a traceback. Long enough for a provider's own
|
|
64
|
+
#: sentence and the error object google-genai folds into its message.
|
|
65
|
+
DESCRIBE_LIMIT = 500
|
|
66
|
+
|
|
67
|
+
_CONNECTION_NAMES = frozenset({
|
|
68
|
+
"APIConnectionError", "APITimeoutError", # openai, anthropic
|
|
69
|
+
"TransportError", "TimeoutException", # httpx: raw mid-stream; google-genai
|
|
70
|
+
"ConnectionError", "Timeout", # requests (google-auth); builtin
|
|
71
|
+
"ChunkedEncodingError", # requests, a body cut off mid-read
|
|
72
|
+
"TimeoutError", # builtin
|
|
73
|
+
})
|
|
74
|
+
|
|
75
|
+
#: Error types an open stream's error object can name. Listed in both
|
|
76
|
+
#: directions rather than inferred from a suffix: openai's "server_error" and
|
|
77
|
+
#: anthropic's "invalid_request_error" share one.
|
|
78
|
+
_TRANSIENT_TYPES = frozenset({
|
|
79
|
+
"overloaded_error", "api_error", "rate_limit_error", "server_error",
|
|
80
|
+
})
|
|
81
|
+
_PERMANENT_TYPES = frozenset({
|
|
82
|
+
"invalid_request_error", "authentication_error", "permission_error",
|
|
83
|
+
"not_found_error", "request_too_large",
|
|
84
|
+
})
|
|
85
|
+
|
|
86
|
+
#: A 200 that failed schema validation is a mismatch between the SDK and the
|
|
87
|
+
#: response, and it will be the same response next time.
|
|
88
|
+
_NEVER_TRANSIENT_NAMES = frozenset({"APIResponseValidationError"})
|
|
89
|
+
|
|
90
|
+
|
|
91
|
+
def _names(exc: BaseException) -> set:
|
|
92
|
+
return {cls.__name__ for cls in type(exc).__mro__}
|
|
93
|
+
|
|
94
|
+
|
|
95
|
+
def _as_status(value) -> Optional[int]:
|
|
96
|
+
if isinstance(value, str) and value.isdigit():
|
|
97
|
+
value = int(value)
|
|
98
|
+
if isinstance(value, int) and not isinstance(value, bool) \
|
|
99
|
+
and 100 <= value < 600:
|
|
100
|
+
return value
|
|
101
|
+
return None
|
|
102
|
+
|
|
103
|
+
|
|
104
|
+
def _retryable_status(status: int) -> bool:
|
|
105
|
+
return status in RETRYABLE_STATUSES or status >= 500
|
|
106
|
+
|
|
107
|
+
|
|
108
|
+
def _status(exc: BaseException, names: set) -> Optional[int]:
|
|
109
|
+
"""The HTTP status an exception carries, or None.
|
|
110
|
+
|
|
111
|
+
`code` is read only on something named APIError, because google-genai
|
|
112
|
+
is the one SDK that spells its status that way and the attribute name is
|
|
113
|
+
far too common to trust anywhere else.
|
|
114
|
+
"""
|
|
115
|
+
status = _as_status(getattr(exc, "status_code", None))
|
|
116
|
+
if status is None and "APIError" in names:
|
|
117
|
+
status = _as_status(getattr(exc, "code", None))
|
|
118
|
+
return status
|
|
119
|
+
|
|
120
|
+
|
|
121
|
+
def _body_verdict(body) -> Optional[bool]:
|
|
122
|
+
"""What an open stream's error object says, or None when it is silent.
|
|
123
|
+
|
|
124
|
+
openai hands over the error object itself (`data["error"]`); anthropic
|
|
125
|
+
hands over the whole event (`{"type": "error", "error": {...}}`), so the
|
|
126
|
+
nested object is preferred when there is one.
|
|
127
|
+
"""
|
|
128
|
+
if not isinstance(body, dict):
|
|
129
|
+
return None
|
|
130
|
+
error = body["error"] if isinstance(body.get("error"), dict) else body
|
|
131
|
+
status = _as_status(error.get("code"))
|
|
132
|
+
if status is not None:
|
|
133
|
+
return _retryable_status(status)
|
|
134
|
+
kind = error.get("type")
|
|
135
|
+
if kind in _TRANSIENT_TYPES:
|
|
136
|
+
return True
|
|
137
|
+
if kind in _PERMANENT_TYPES:
|
|
138
|
+
return False
|
|
139
|
+
return None
|
|
140
|
+
|
|
141
|
+
|
|
142
|
+
def is_transient(exc: BaseException) -> bool:
|
|
143
|
+
"""Would the same call plausibly succeed if simply made again?"""
|
|
144
|
+
names = _names(exc)
|
|
145
|
+
if names & _NEVER_TRANSIENT_NAMES:
|
|
146
|
+
return False
|
|
147
|
+
if names & _CONNECTION_NAMES:
|
|
148
|
+
return True
|
|
149
|
+
status = _status(exc, names)
|
|
150
|
+
if status is not None and status != 200:
|
|
151
|
+
return _retryable_status(status)
|
|
152
|
+
if "APIError" in names:
|
|
153
|
+
# An error event inside a stream that opened: no status (openai) or
|
|
154
|
+
# the stream's own 200 (anthropic). The provider has said something
|
|
155
|
+
# went wrong mid-generation; unless it also said the request itself
|
|
156
|
+
# is at fault, that is the weather this retry is for.
|
|
157
|
+
return open_stream_error_is_transient(getattr(exc, "body", None))
|
|
158
|
+
return False
|
|
159
|
+
|
|
160
|
+
|
|
161
|
+
def open_stream_error_is_transient(body) -> bool:
|
|
162
|
+
"""The open-stream rule above, for a caller holding only the error object.
|
|
163
|
+
|
|
164
|
+
ROADMAP_v3 §83: the ChatGPT plan's `error` event and `response.failed`
|
|
165
|
+
arrive as EVENTS -- the openai SDK raises only on a top-level `error`
|
|
166
|
+
object, which neither has -- so `core/client.py` reads them and asks
|
|
167
|
+
here, rather than keeping a second copy of the rule.
|
|
168
|
+
"""
|
|
169
|
+
verdict = _body_verdict(body)
|
|
170
|
+
return True if verdict is None else verdict
|
|
171
|
+
|
|
172
|
+
|
|
173
|
+
def retry_after_s(exc: BaseException) -> Optional[float]:
|
|
174
|
+
"""A server's `retry-after`, in seconds, when the failure carried one.
|
|
175
|
+
|
|
176
|
+
Only the delta-seconds form. The HTTP-date form is legal and rare enough
|
|
177
|
+
on these APIs that a date parser here would be code with no caller; an
|
|
178
|
+
unreadable value is simply ignored and the backoff decides.
|
|
179
|
+
"""
|
|
180
|
+
headers = getattr(getattr(exc, "response", None), "headers", None)
|
|
181
|
+
if headers is None:
|
|
182
|
+
return None
|
|
183
|
+
try:
|
|
184
|
+
seconds = float(headers.get("retry-after"))
|
|
185
|
+
except (AttributeError, TypeError, ValueError):
|
|
186
|
+
return None
|
|
187
|
+
if seconds < 0:
|
|
188
|
+
return None
|
|
189
|
+
return min(seconds, RETRY_AFTER_CAP_S)
|
|
190
|
+
|
|
191
|
+
|
|
192
|
+
def backoff_delay(retry: int, base_s: float,
|
|
193
|
+
retry_after: Optional[float] = None) -> float:
|
|
194
|
+
"""Seconds to wait before retry number `retry` (1-based).
|
|
195
|
+
|
|
196
|
+
`base_s`, doubled for each retry after the first, with JITTER either way
|
|
197
|
+
(owner decision: longer than the SDKs' own half-second start, so a
|
|
198
|
+
temporary rate limit is not met with a burst of attempts).
|
|
199
|
+
|
|
200
|
+
A server's retry-after can make the wait LONGER, never shorter: a
|
|
201
|
+
`retry-after: 0` answering a rate limit is exactly the hammering the
|
|
202
|
+
longer base exists to avoid.
|
|
203
|
+
"""
|
|
204
|
+
delay = base_s * (2 ** (retry - 1))
|
|
205
|
+
delay *= 1 + JITTER * (2 * random() - 1)
|
|
206
|
+
if retry_after is not None:
|
|
207
|
+
delay = max(delay, retry_after)
|
|
208
|
+
return delay
|
|
209
|
+
|
|
210
|
+
|
|
211
|
+
def wait(seconds: float) -> None:
|
|
212
|
+
"""Sleep before a retry.
|
|
213
|
+
|
|
214
|
+
Through `sleep` bound HERE, so a test patches `core.provider_errors.sleep`
|
|
215
|
+
and reaches nothing else -- AGENTS.md's `from time import monotonic`
|
|
216
|
+
rule: patching the global `time` module froze textual's event loop and
|
|
217
|
+
conftest's `settle` along with it, and the suite hung rather than failed.
|
|
218
|
+
"""
|
|
219
|
+
sleep(seconds)
|
|
220
|
+
|
|
221
|
+
|
|
222
|
+
def describe(exc: BaseException) -> str:
|
|
223
|
+
"""`Type: message`, on one line -- the ONE spelling of a failure.
|
|
224
|
+
|
|
225
|
+
Shared by the transcript's log handler, the failure stored on a thread
|
|
226
|
+
and the retry warning, so the three never describe the same exception
|
|
227
|
+
three ways. Not redacted: every caller that shows or
|
|
228
|
+
stores it redacts the whole line it is part of.
|
|
229
|
+
"""
|
|
230
|
+
text = " ".join(str(exc).split())
|
|
231
|
+
if len(text) > DESCRIBE_LIMIT:
|
|
232
|
+
text = text[:DESCRIBE_LIMIT - 1].rstrip() + "…"
|
|
233
|
+
name = type(exc).__name__
|
|
234
|
+
return f"{name}: {text}" if text else name
|