@yahyahammad/venastine 0.1.0
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 +14 -0
- package/LICENSE +201 -0
- package/NOTICE +14 -0
- package/PRIVACY.md +66 -0
- package/README.md +878 -0
- package/SECURITY.md +88 -0
- package/THIRD_PARTY_NOTICES.md +100 -0
- package/agents/builtin/compactor.md +75 -0
- package/agents/builtin/explore.md +77 -0
- package/agents/builtin/grill-me.md +27 -0
- package/agents/builtin/initializer.md +94 -0
- package/agents/builtin/pipeline-reviewer.md +100 -0
- package/agents/builtin/plan.md +88 -0
- package/agents/builtin/review.md +74 -0
- package/agents/manager.py +195 -0
- package/agents/subagent_tool.py +215 -0
- package/agents/tui_commands.py +136 -0
- package/bin/venastine.mjs +382 -0
- package/config.py +1198 -0
- package/core/__init__.py +0 -0
- package/core/approval.py +115 -0
- package/core/client.py +1548 -0
- package/core/compaction.py +959 -0
- package/core/config_loader.py +1563 -0
- package/core/events.py +61 -0
- package/core/interaction.py +328 -0
- package/core/loop.py +1353 -0
- package/core/memory.py +370 -0
- package/core/model_windows.py +175 -0
- package/core/pipeline_models.py +176 -0
- package/core/reasoning/__init__.py +0 -0
- package/core/reasoning/authorization.py +130 -0
- package/core/reasoning/base.py +219 -0
- package/core/reasoning/confidence_scoring.py +468 -0
- package/core/reasoning/events.py +149 -0
- package/core/reasoning/json_retry.py +219 -0
- package/core/reasoning/orchestrator.py +1382 -0
- package/core/reasoning/output_writer.py +286 -0
- package/core/reasoning/payload_validation.py +398 -0
- package/core/reasoning/pipeline_storage.py +361 -0
- package/core/reasoning/review.py +615 -0
- package/core/reasoning/scholar.py +434 -0
- package/core/reasoning/source_corpus.py +327 -0
- package/core/reasoning/source_scoring.py +855 -0
- package/core/replay.py +167 -0
- package/core/session.py +153 -0
- package/core/workspace_trust.py +207 -0
- package/credentials.py +180 -0
- package/database.py +252 -0
- package/env_secrets.py +77 -0
- package/json_store.py +123 -0
- package/logging_setup.py +185 -0
- package/main.py +1959 -0
- package/mcp_client/__init__.py +15 -0
- package/mcp_client/client.py +499 -0
- package/mcp_client/config.py +379 -0
- package/mcp_client/registration.py +182 -0
- package/memories/manager.py +132 -0
- package/memories/tui_commands.py +232 -0
- package/package.json +73 -0
- package/project_init/config_files.py +278 -0
- package/project_init/doc_sets.py +330 -0
- package/project_init/generator.py +564 -0
- package/project_init/manifest.py +406 -0
- package/project_init/tui_commands.py +144 -0
- package/prompts/__init__.py +0 -0
- package/prompts/annotate.md +14 -0
- package/prompts/assumption_audit.md +27 -0
- package/prompts/claim_extraction.md +34 -0
- package/prompts/completeness.md +18 -0
- package/prompts/confidence_tiers.md +23 -0
- package/prompts/critic_pass.md +24 -0
- package/prompts/final_synthesis.md +13 -0
- package/prompts/initial_generation.md +14 -0
- package/prompts/preliminary_plan.md +23 -0
- package/prompts/revalidate.md +25 -0
- package/prompts/revise.md +15 -0
- package/prompts/source_grounding.md +66 -0
- package/prompts/system_prompts.py +262 -0
- package/prompts/universal_system_prompt +7 -0
- package/prompts/untrusted_content +1 -0
- package/providers.json.example +92 -0
- package/pyproject.toml +100 -0
- package/requirements.txt +72 -0
- package/safety/__init__.py +0 -0
- package/safety/policy_enforcement.py +682 -0
- package/security/__init__.py +0 -0
- package/security/capability.py +177 -0
- package/security/permissions.py +190 -0
- package/security/posture.py +302 -0
- package/security/protected_paths.py +309 -0
- package/security/sandbox.py +985 -0
- package/skills/builtin/crypto/cryptography-verification.md +66 -0
- package/skills/builtin/math/formal-specification.md +67 -0
- package/skills/builtin/math/numerical-methods.md +66 -0
- package/skills/builtin/math/proof-writing.md +61 -0
- package/skills/builtin/research/experiment-design.md +62 -0
- package/skills/builtin/research/literature-review.md +65 -0
- package/skills/builtin/research/reproducibility.md +55 -0
- package/skills/builtin/research/source-evaluation.md +70 -0
- package/skills/builtin/research/statistical-inference.md +63 -0
- package/skills/builtin/research/technical-writing.md +59 -0
- package/skills/builtin/security/cybersecurity-research.md +63 -0
- package/skills/builtin/software/code-review.md +61 -0
- package/skills/builtin/software/debugging.md +61 -0
- package/skills/builtin/software/test-design.md +67 -0
- package/skills/manager.py +138 -0
- package/skills/tui_commands.py +111 -0
- package/storage.py +1005 -0
- package/tools/base.py +305 -0
- package/tools/builtin/__init__.py +0 -0
- package/tools/builtin/_math_common.py +377 -0
- package/tools/builtin/_net_common.py +61 -0
- package/tools/builtin/arxiv.py +223 -0
- package/tools/builtin/ask_user.py +213 -0
- package/tools/builtin/discrete_math.py +134 -0
- package/tools/builtin/fetch_url.py +153 -0
- package/tools/builtin/file_ops.py +412 -0
- package/tools/builtin/geometry.py +148 -0
- package/tools/builtin/get_time.py +20 -0
- package/tools/builtin/linear_algebra.py +155 -0
- package/tools/builtin/load_skill.py +66 -0
- package/tools/builtin/logic.py +96 -0
- package/tools/builtin/pin.py +193 -0
- package/tools/builtin/probability_stats.py +228 -0
- package/tools/builtin/project_docs.py +405 -0
- package/tools/builtin/remember.py +158 -0
- package/tools/builtin/shell.py +461 -0
- package/tools/builtin/symbolic_math.py +119 -0
- package/tools/builtin/todo.py +147 -0
- package/tools/builtin/web_search.py +151 -0
- package/tools/context.py +136 -0
- package/tools/isolation.py +314 -0
- package/tools/registry.py +897 -0
- package/tui/app.py +3423 -0
- package/tui/app.tcss +458 -0
- package/tui/commands.py +61 -0
- package/tui/diffs.py +265 -0
- package/tui/preferences.py +182 -0
- package/tui/ravens.py +146 -0
- package/tui/screens.py +943 -0
- package/tui/themes.py +366 -0
- package/tui/widgets.py +1239 -0
package/core/events.py
ADDED
|
@@ -0,0 +1,61 @@
|
|
|
1
|
+
"""
|
|
2
|
+
core/events.py
|
|
3
|
+
|
|
4
|
+
LoopEvent is the single event type yielded by RunAgentLoop._run() as it
|
|
5
|
+
progresses through the call-dispatch-repeat cycle. Exactly one field is
|
|
6
|
+
meaningfully populated per event, matching what happened at that point
|
|
7
|
+
in the loop.
|
|
8
|
+
|
|
9
|
+
Consumers (CLI, TUI, research pipeline) iterate the generator and react
|
|
10
|
+
to whichever field is set. run_to_completion() in core/loop.py drains
|
|
11
|
+
the generator for callers that only need the final ModelResponse.
|
|
12
|
+
|
|
13
|
+
Error handling: there is deliberately no error variant. Exceptions
|
|
14
|
+
raised mid-loop propagate naturally out of the generator, unchanged
|
|
15
|
+
from the pre-streaming behavior. orchestrator.py's failure-path
|
|
16
|
+
persistence (except Exception: update_pipeline_run(..., status="failed");
|
|
17
|
+
raise) depends on a real exception propagating -- converting failures
|
|
18
|
+
into data-shaped events would require every consumer to know to unpack
|
|
19
|
+
and re-raise. Consumers that want graceful error display (the TUI) wrap
|
|
20
|
+
their own consumption loop in try/except.
|
|
21
|
+
"""
|
|
22
|
+
|
|
23
|
+
from dataclasses import dataclass
|
|
24
|
+
from typing import Optional, Any
|
|
25
|
+
|
|
26
|
+
|
|
27
|
+
@dataclass
|
|
28
|
+
class LoopEvent:
|
|
29
|
+
"""One event yielded by RunAgentLoop._run()."""
|
|
30
|
+
token_delta: Optional[str] = None
|
|
31
|
+
# ROADMAP_v2 §38 (O4) -- the EIGHTH field, and P1 re-made rather than
|
|
32
|
+
# widened by accretion. P1 rejected putting §22's and §23's kinds here
|
|
33
|
+
# because those describe a ten-pass RUN and live for the run; a
|
|
34
|
+
# thinking delta describes one model call's progress and lives for a
|
|
35
|
+
# turn, which is token_delta's family exactly. So this is the case P1's
|
|
36
|
+
# rule was never about, and the ninth field still has to argue for
|
|
37
|
+
# itself in tests/test_pipeline_events.py.
|
|
38
|
+
#
|
|
39
|
+
# Reasoning text, streamed as it arrives, for the shells that want to
|
|
40
|
+
# show it. Display-only: it is not persisted, not part of
|
|
41
|
+
# ModelResponse.text, and never sent back on the wire. Empty on
|
|
42
|
+
# providers that return no reasoning (OPENAI, GOOGLE) -- see
|
|
43
|
+
# StreamToken's note.
|
|
44
|
+
thinking_delta: Optional[str] = None
|
|
45
|
+
tool_call_start: Optional[dict] = None # {"id", "name", "input"}
|
|
46
|
+
tool_result: Optional[dict] = None # {"id", "result"}
|
|
47
|
+
permission_request: Optional[dict] = None # {"tool_name", "params"}
|
|
48
|
+
# ROADMAP_v2 §21: {"kind", "text"} -- a compaction that happened, an
|
|
49
|
+
# early warning that one is coming, or a compaction that could not
|
|
50
|
+
# help. §21's "no silent compaction, ever": both automatic and manual
|
|
51
|
+
# compaction show a marker inline.
|
|
52
|
+
#
|
|
53
|
+
# ALSO carried on the ModelResponse, because run_to_completion()
|
|
54
|
+
# discards every non-final event -- so a notice delivered only here
|
|
55
|
+
# would be invisible to the CLI and the pipeline, which drain the
|
|
56
|
+
# generator rather than watching it. That is the same defect §20 and
|
|
57
|
+
# §25 each hit once; the event is for live display, the response field
|
|
58
|
+
# is for everyone else.
|
|
59
|
+
notice: Optional[dict] = None
|
|
60
|
+
final_response: Optional[Any] = None # ModelResponse on terminal event
|
|
61
|
+
stop_reason: Optional[str] = None # "complete" | "max_steps_reached" | "token_budget_exceeded"
|
|
@@ -0,0 +1,328 @@
|
|
|
1
|
+
"""
|
|
2
|
+
core/interaction.py
|
|
3
|
+
|
|
4
|
+
ROADMAP_v2 §23 (AC1): the ONE way anything in this harness asks a human a
|
|
5
|
+
question and blocks for the answer.
|
|
6
|
+
|
|
7
|
+
Before this there were five, and each re-derived the same four rules --
|
|
8
|
+
what a dismissal carries, what a timeout means, how shutdown releases a
|
|
9
|
+
parked worker, and how a malformed answer decodes:
|
|
10
|
+
|
|
11
|
+
* `permission_channel`, a queue.Queue of booleans (§13)
|
|
12
|
+
* `ApprovalProvider.ask(tool_name, params, notice) -> bool` (§25), which
|
|
13
|
+
exists only because run_to_completion() DISCARDS the LoopEvent carrying
|
|
14
|
+
the question, so the CLI and the pipeline would block on a queue with
|
|
15
|
+
nothing displayed
|
|
16
|
+
* `ReviewConsent.decide(finding, round) -> (decision, notes)` (§20)
|
|
17
|
+
* two more in §24's /init, added by the section immediately before this one
|
|
18
|
+
|
|
19
|
+
§23's spec put the generalisation first for a reason it has now been proved
|
|
20
|
+
right about twice: "doing the tools first and the generalisation afterwards
|
|
21
|
+
produces two bespoke channels and a third when §18 wants one."
|
|
22
|
+
|
|
23
|
+
WHAT IS ACTUALLY SHARED IS `decode`, NOT THE DATACLASS. A request/response
|
|
24
|
+
dataclass pair is nearly free to write badly -- five call sites can each
|
|
25
|
+
build one and still each invent their own failure handling. The consolidation
|
|
26
|
+
that pays is that every route into an answer (a queue, a callable, a timeout,
|
|
27
|
+
a shutdown release, a callback that raised) funnels through ONE function that
|
|
28
|
+
owns the safe default per kind. "The inability to ask is not permission to
|
|
29
|
+
proceed" (§25's V6) stops being a habit five places remember and becomes a
|
|
30
|
+
property of this module.
|
|
31
|
+
|
|
32
|
+
THE DEFAULT IS ALWAYS THE DECLINING ANSWER, never the permissive one. That
|
|
33
|
+
asymmetry is the whole design: an approval that cannot be obtained is a
|
|
34
|
+
denial, a sign-off nobody answered grants nothing, a review correction nobody
|
|
35
|
+
approved is not applied.
|
|
36
|
+
|
|
37
|
+
Deliberately a leaf module with no project imports, matching core/approval.py
|
|
38
|
+
-- core/loop.py, main.py, tui/app.py and core/reasoning/ all depend on it and
|
|
39
|
+
it depends on none of them.
|
|
40
|
+
"""
|
|
41
|
+
|
|
42
|
+
from dataclasses import dataclass, field
|
|
43
|
+
from typing import Any, Callable, Optional
|
|
44
|
+
|
|
45
|
+
import logging
|
|
46
|
+
|
|
47
|
+
logger = logging.getLogger(__name__)
|
|
48
|
+
|
|
49
|
+
|
|
50
|
+
# ---------------------------------------------------------------------------
|
|
51
|
+
# ---- Kinds ----------------------------------------------------------------
|
|
52
|
+
# ---------------------------------------------------------------------------
|
|
53
|
+
|
|
54
|
+
APPROVAL = "approval"
|
|
55
|
+
SUBAGENT_SIGNOFF = "subagent_signoff"
|
|
56
|
+
REVIEW = "review"
|
|
57
|
+
# §24's two, migrated in §23 slice 1. CONFIRM and APPROVAL both answer with
|
|
58
|
+
# a bool and are still separate kinds, because the KIND is what tells a
|
|
59
|
+
# shell how to render: APPROVAL is a gated tool call with params, CONFIRM
|
|
60
|
+
# is a titled yes/no over a block of text the caller composed.
|
|
61
|
+
CONFIRM = "confirm"
|
|
62
|
+
CHOICE = "choice"
|
|
63
|
+
# §23 slice 2. Separate from CHOICE, which answers with exactly one of the
|
|
64
|
+
# options offered and nothing else. A QUESTION can come back as several
|
|
65
|
+
# options, as free text, as both, or as "I would rather talk about this" --
|
|
66
|
+
# four affordances the spec asks for by name, and a shape CHOICE's decoder
|
|
67
|
+
# cannot express without becoming two decoders wearing one name.
|
|
68
|
+
QUESTION = "question"
|
|
69
|
+
|
|
70
|
+
# Review decisions, moved here from core/reasoning/review.py so the decoder
|
|
71
|
+
# and the pipeline cannot disagree about what a valid answer is.
|
|
72
|
+
ACCEPT = "accept"
|
|
73
|
+
REJECT = "reject"
|
|
74
|
+
REFINE = "refine"
|
|
75
|
+
REJECT_ALL = "reject_all"
|
|
76
|
+
REVIEW_DECISIONS = (ACCEPT, REJECT, REFINE, REJECT_ALL)
|
|
77
|
+
|
|
78
|
+
# What each kind means when nobody answered, answered incoherently, or could
|
|
79
|
+
# not be asked at all. Every entry declines.
|
|
80
|
+
#
|
|
81
|
+
# INVARIANT, pinned by a test: `decode(Request(kind), None) ==
|
|
82
|
+
# SAFE_DEFAULTS[kind]` for
|
|
83
|
+
# every kind. That is what lets `ask()` route the no-channel and the
|
|
84
|
+
# callback-raised paths through `decode` rather than reading this table -- so
|
|
85
|
+
# there is genuinely one function deciding what a non-answer means, and a new
|
|
86
|
+
# kind cannot be added with a default the decoder disagrees about.
|
|
87
|
+
SAFE_DEFAULTS = {
|
|
88
|
+
APPROVAL: False,
|
|
89
|
+
# None, not set(): the two are DIFFERENT answers. None denies the spawn
|
|
90
|
+
# outright; an empty set spawns the subagent with nothing pre-granted,
|
|
91
|
+
# which is a legitimate choice a user can make. Defaulting to set() would
|
|
92
|
+
# turn "nobody answered" into "go ahead, just ask me later" -- and in a
|
|
93
|
+
# headless run there is nobody to ask later either.
|
|
94
|
+
SUBAGENT_SIGNOFF: None,
|
|
95
|
+
REVIEW: (REJECT, ""),
|
|
96
|
+
CONFIRM: False,
|
|
97
|
+
# None, not the first option: "nobody answered" is not a choice, and
|
|
98
|
+
# picking one on the user's behalf is exactly what a declining default
|
|
99
|
+
# exists to avoid.
|
|
100
|
+
CHOICE: None,
|
|
101
|
+
# None, for CHOICE's reason. Note what this makes indistinguishable, on
|
|
102
|
+
# purpose: an unanswered question and a dismissed one both arrive here,
|
|
103
|
+
# and the TOOL turns that into an error result the model can work
|
|
104
|
+
# around. "I would rather discuss this" is NOT this -- it is a real
|
|
105
|
+
# answer carrying defer=True, because a user who says that has told the
|
|
106
|
+
# model something, and a user who closed the modal has not.
|
|
107
|
+
QUESTION: None,
|
|
108
|
+
}
|
|
109
|
+
|
|
110
|
+
|
|
111
|
+
@dataclass
|
|
112
|
+
class Request:
|
|
113
|
+
"""One question, addressed to whoever can answer it.
|
|
114
|
+
|
|
115
|
+
`payload` is kind-specific and is not validated here: this module owns
|
|
116
|
+
the ANSWER contract, not the question's. A shell renders the payload it
|
|
117
|
+
understands; an unrecognised kind is a programming error, and `decode`
|
|
118
|
+
raises on one rather than inventing a default for a question nobody
|
|
119
|
+
designed.
|
|
120
|
+
"""
|
|
121
|
+
|
|
122
|
+
kind: str
|
|
123
|
+
payload: dict = field(default_factory=dict)
|
|
124
|
+
# Extra text explaining what answering actually authorises. Supplied by
|
|
125
|
+
# ToolSpec.approval_notice for tool calls -- the tool's knowledge, not
|
|
126
|
+
# the shell's (spawn_subagent uses it to name the tools a child could
|
|
127
|
+
# then run unprompted).
|
|
128
|
+
notice: Optional[str] = None
|
|
129
|
+
|
|
130
|
+
|
|
131
|
+
@dataclass
|
|
132
|
+
class ResponseChannel:
|
|
133
|
+
"""A caller-supplied way to put one Request to a human and block.
|
|
134
|
+
|
|
135
|
+
ask(request) -> raw answer. Implementations live in the shells, because
|
|
136
|
+
only they know how to reach the person: main.py reads stdin, tui/app.py
|
|
137
|
+
posts to the UI thread and waits on a queue.
|
|
138
|
+
|
|
139
|
+
The raw answer is never trusted. Callers go through this module's
|
|
140
|
+
`ask()`, which decodes it and substitutes the declining default for
|
|
141
|
+
anything unrecognised.
|
|
142
|
+
|
|
143
|
+
honour_run_scope mirrors ToolSpec.grant_scope, carried over unchanged
|
|
144
|
+
from §25's ApprovalProvider. A channel that sets it False makes every
|
|
145
|
+
call ask, even for a tool whose approval would normally cover the rest
|
|
146
|
+
of the run. Attended mode sets it False: a mode whose entire purpose is
|
|
147
|
+
per-call supervision must not let one yes silently cover later calls,
|
|
148
|
+
however sensible that shortcut is inside a chat turn.
|
|
149
|
+
"""
|
|
150
|
+
|
|
151
|
+
ask: Callable[[Request], Any]
|
|
152
|
+
honour_run_scope: bool = True
|
|
153
|
+
|
|
154
|
+
|
|
155
|
+
# ---------------------------------------------------------------------------
|
|
156
|
+
# ---- Decoding --------------------------------------------------------------
|
|
157
|
+
# ---------------------------------------------------------------------------
|
|
158
|
+
|
|
159
|
+
def decode(request: "Request", raw: Any) -> Any:
|
|
160
|
+
"""Normalise whatever a shell returned into this kind's answer type.
|
|
161
|
+
|
|
162
|
+
Anything unrecognised becomes SAFE_DEFAULTS[kind]. That covers a
|
|
163
|
+
dismissal carrying None, a modal torn down by shutdown (which puts a
|
|
164
|
+
bare False), a timeout, and a shell that returned something of the
|
|
165
|
+
wrong shape entirely.
|
|
166
|
+
|
|
167
|
+
TAKES THE REQUEST, NOT JUST THE KIND, because two kinds cannot be
|
|
168
|
+
validated without knowing what was ASKED. A choice is only valid
|
|
169
|
+
against the options it was offered, and a sign-off subset is only
|
|
170
|
+
valid against the candidates that were listed -- a shell returning a
|
|
171
|
+
tool nobody offered must not thereby grant it. Passing the kind alone
|
|
172
|
+
made those two unverifiable, which is the sort of gap that shows up
|
|
173
|
+
later as "the answer was honoured but nothing had proposed it".
|
|
174
|
+
|
|
175
|
+
Raises on an unknown KIND, deliberately, rather than returning a
|
|
176
|
+
default. An unanswerable question is a runtime condition to fail safe
|
|
177
|
+
on; a question nobody defined is a bug, and inventing "no" for it would
|
|
178
|
+
hide the typo in whichever call site built the Request.
|
|
179
|
+
"""
|
|
180
|
+
kind = request.kind
|
|
181
|
+
if kind not in SAFE_DEFAULTS:
|
|
182
|
+
raise ValueError(
|
|
183
|
+
f"Unknown request kind {kind!r}; expected one of "
|
|
184
|
+
f"{', '.join(sorted(SAFE_DEFAULTS))}.")
|
|
185
|
+
|
|
186
|
+
if kind in (APPROVAL, CONFIRM):
|
|
187
|
+
# bool(), not `is True`: a shell answering with a truthy value it
|
|
188
|
+
# considers a yes should not be silently denied. The permissive
|
|
189
|
+
# direction is safe here ONLY because the alternative is also a
|
|
190
|
+
# decision the user made -- unlike the shapes below, there is no
|
|
191
|
+
# third state to confuse it with.
|
|
192
|
+
return bool(raw)
|
|
193
|
+
|
|
194
|
+
if kind == CHOICE:
|
|
195
|
+
# Valid ONLY against what was offered. A shell returning something
|
|
196
|
+
# nobody proposed is answering a different question, and the
|
|
197
|
+
# options live on the request precisely so this can be checked --
|
|
198
|
+
# `decode(kind, raw)` could not, which is why it takes the request.
|
|
199
|
+
options = request.payload.get("options") or ()
|
|
200
|
+
return raw if raw in options else None
|
|
201
|
+
|
|
202
|
+
if kind == SUBAGENT_SIGNOFF:
|
|
203
|
+
# Three answers, and the middle one is the easy one to lose: None
|
|
204
|
+
# denies the spawn, an EMPTY collection spawns with nothing
|
|
205
|
+
# granted, a non-empty one grants those names. The emptiness test
|
|
206
|
+
# is therefore `isinstance`, never truthiness -- `if raw:` would
|
|
207
|
+
# fold the middle answer into the first.
|
|
208
|
+
#
|
|
209
|
+
# Everything else denies by falling through, booleans included: a
|
|
210
|
+
# shell that answered this like an approval must not hand the
|
|
211
|
+
# child its whole gated set. That needs no explicit `raw is True`
|
|
212
|
+
# branch, since bool subclasses int rather than any collection --
|
|
213
|
+
# one was written here and removed, because a guard that cannot
|
|
214
|
+
# fire still reads as the thing protecting you.
|
|
215
|
+
if not isinstance(raw, (set, frozenset, list, tuple)):
|
|
216
|
+
return None
|
|
217
|
+
chosen = {str(name) for name in raw}
|
|
218
|
+
# INTERSECTED with what was offered. A shell returning a name that
|
|
219
|
+
# was never a candidate must not grant it -- the sign-off's whole
|
|
220
|
+
# substance is that the list shown and the list granted are the
|
|
221
|
+
# same, which agents/manager.py's candidate_approvals() says in its
|
|
222
|
+
# own docstring. An empty `candidates` means nothing was offered,
|
|
223
|
+
# so nothing can be granted.
|
|
224
|
+
offered = {str(name) for name in
|
|
225
|
+
(request.payload.get("candidates") or ())}
|
|
226
|
+
unoffered = chosen - offered
|
|
227
|
+
if unoffered:
|
|
228
|
+
logger.warning(
|
|
229
|
+
"Sign-off answer named %s, which %s not offered; ignoring "
|
|
230
|
+
"them.", sorted(unoffered),
|
|
231
|
+
"was" if len(unoffered) == 1 else "were")
|
|
232
|
+
return chosen & offered
|
|
233
|
+
|
|
234
|
+
if kind == QUESTION:
|
|
235
|
+
# A dict, or nothing. Three answers again, and again the middle one
|
|
236
|
+
# is the one to lose:
|
|
237
|
+
#
|
|
238
|
+
# None nobody answered -- dismissed, timed
|
|
239
|
+
# out, no channel, or a shape this
|
|
240
|
+
# cannot read. The TOOL turns it into an
|
|
241
|
+
# error result.
|
|
242
|
+
# {..., "defer": True} "let's talk about it instead". A REAL
|
|
243
|
+
# answer: the user engaged and declined
|
|
244
|
+
# to pick, which is different from not
|
|
245
|
+
# being there. This is the spec's "chat
|
|
246
|
+
# about this" escape.
|
|
247
|
+
# {"options": [...], ...} a choice, or free text, or both.
|
|
248
|
+
#
|
|
249
|
+
# Deliberately NOT truthiness on the dict: an answer of
|
|
250
|
+
# {"options": [], "text": ""} is someone submitting a blank form,
|
|
251
|
+
# which is still not "nobody answered", and the tool can say so.
|
|
252
|
+
if not isinstance(raw, dict):
|
|
253
|
+
return None
|
|
254
|
+
if raw.get("defer"):
|
|
255
|
+
# Nothing else is read. A deferral is not a partial answer, and
|
|
256
|
+
# carrying half-filled options alongside it would invite a
|
|
257
|
+
# caller to act on both.
|
|
258
|
+
return {"options": [], "text": "", "defer": True}
|
|
259
|
+
|
|
260
|
+
# INTERSECTED with what was offered, in OFFERED order -- the same
|
|
261
|
+
# rule as SUBAGENT_SIGNOFF above, and for J3's reason: a shell
|
|
262
|
+
# returning something nobody proposed is answering a different
|
|
263
|
+
# question. Offered order rather than answer order so the model
|
|
264
|
+
# reads options back in the sequence it wrote them, which is the
|
|
265
|
+
# sequence its own prompt explained.
|
|
266
|
+
offered = [str(o) for o in (request.payload.get("options") or ())]
|
|
267
|
+
answered = raw.get("options")
|
|
268
|
+
if not isinstance(answered, (set, frozenset, list, tuple)):
|
|
269
|
+
answered = ()
|
|
270
|
+
chosen = {str(o) for o in answered}
|
|
271
|
+
unoffered = chosen - set(offered)
|
|
272
|
+
if unoffered:
|
|
273
|
+
logger.warning(
|
|
274
|
+
"Question answer named %s, which %s not offered; ignoring "
|
|
275
|
+
"them.", sorted(unoffered),
|
|
276
|
+
"was" if len(unoffered) == 1 else "were")
|
|
277
|
+
|
|
278
|
+
text = raw.get("text")
|
|
279
|
+
return {
|
|
280
|
+
"options": [o for o in offered if o in chosen],
|
|
281
|
+
"text": str(text) if text else "",
|
|
282
|
+
"defer": False,
|
|
283
|
+
}
|
|
284
|
+
|
|
285
|
+
# REVIEW. Strict: a well-formed (decision, notes) pair or nothing.
|
|
286
|
+
#
|
|
287
|
+
# The two decoders this replaces DISAGREED -- review.py's accepted a
|
|
288
|
+
# bare "accept" string and a 1-tuple, tui/app.py's required exactly two
|
|
289
|
+
# elements. Neither behaviour was tested, and the strict one wins
|
|
290
|
+
# because the unsafe direction here is ACCEPTING: this is the only kind
|
|
291
|
+
# whose permissive failure silently applies an edit to a finished
|
|
292
|
+
# research run rather than merely declining something.
|
|
293
|
+
if not isinstance(raw, (tuple, list)) or len(raw) != 2:
|
|
294
|
+
return SAFE_DEFAULTS[REVIEW]
|
|
295
|
+
decision, notes = raw
|
|
296
|
+
if decision not in REVIEW_DECISIONS:
|
|
297
|
+
return SAFE_DEFAULTS[REVIEW]
|
|
298
|
+
return (decision, notes or "")
|
|
299
|
+
|
|
300
|
+
|
|
301
|
+
def ask(channel: Optional[ResponseChannel], request: Request) -> Any:
|
|
302
|
+
"""Put `request` to `channel` and return a decoded answer.
|
|
303
|
+
|
|
304
|
+
THE ONE ENTRY POINT. Callers never invoke `channel.ask` directly, so
|
|
305
|
+
there is exactly one place that knows a missing channel declines, a
|
|
306
|
+
raising callback declines, and a malformed answer declines.
|
|
307
|
+
|
|
308
|
+
A channel of None is not an error: it is the ordinary state of every
|
|
309
|
+
headless run, and §13 pairs it with hiding approval-gated tools rather
|
|
310
|
+
than calling them and failing. Anything that reaches here anyway --
|
|
311
|
+
a path-dependent approval_check that only resolved at call time -- gets
|
|
312
|
+
the declining default.
|
|
313
|
+
"""
|
|
314
|
+
if channel is None:
|
|
315
|
+
return decode(request, None)
|
|
316
|
+
try:
|
|
317
|
+
raw = channel.ask(request)
|
|
318
|
+
except Exception: # noqa: BLE001 -- a shell's failure is not the run's
|
|
319
|
+
# Logged with the traceback, because a shell raising here is a real
|
|
320
|
+
# bug worth finding, and at ERROR because the user's answer was
|
|
321
|
+
# silently discarded. The run continues on the declining default:
|
|
322
|
+
# a broken prompt must not take down a ten-pass pipeline, and must
|
|
323
|
+
# not be read as a yes either.
|
|
324
|
+
logger.exception(
|
|
325
|
+
"Response channel raised answering a %r request; "
|
|
326
|
+
"treating as declined.", request.kind)
|
|
327
|
+
return decode(request, None)
|
|
328
|
+
return decode(request, raw)
|