Give each role its own view of the transcript (M1.2)
CI / checks (push) Skipped
CI / checks (pull_request) Successful in 1m27s
CI / checks (push) Skipped
CI / checks (pull_request) Successful in 1m27s
A model no longer reads another role's turns as its own (#78). The log stays one record per work item; what a role is shown is projected from it by Transcript.view() (views.py), which the planner, executor and helper now build their context from: - own events unchanged; - operator messages labelled with their channel, e.g. [operator, Matrix] or [operator, comment on PR #165]; - orchestrator facts (a new `fact` event) as [orchestrator] statements; - the executor sees a helper's handback as [helper report], not its turns; - a helper sees the approved plan and the executor's milestone work as one-line narration; - the planner sees no other role's work at all. The orchestrator now writes facts at the points the models must know about: work item opened, plan approved or rejected, milestone started, finished or stopped, published (with PR number and URL) or not, review comment answered, next milestone, task completed or abandoned. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PPQ14L3LikyefCMFBqEbKE
This commit is contained in:
1 parent
51e144e26b
commit
3fba958958
15 files changed
+517
-13
No files matched your search
@@ -73,6 +73,15 @@ class DelegationCompleted:
|
||||
exit_status: str
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Fact:
|
||||
"""Something the orchestrator knows that the models must too (L6, L7):
|
||||
a plan approved, a milestone finished, a PR opened. Written by the
|
||||
orchestrator; every role sees it, as a statement rather than a turn."""
|
||||
|
||||
text: str
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Compaction:
|
||||
summary: str
|
||||
@@ -89,6 +98,7 @@ CanonicalEvent = (
|
||||
| GateOpened
|
||||
| DelegationStarted
|
||||
| DelegationCompleted
|
||||
| Fact
|
||||
| Compaction
|
||||
)
|
||||
|
||||
@@ -101,6 +111,7 @@ EVENT_TYPES: dict[str, type] = {
|
||||
"gate_opened": GateOpened,
|
||||
"delegation_started": DelegationStarted,
|
||||
"delegation_completed": DelegationCompleted,
|
||||
"fact": Fact,
|
||||
"compaction": Compaction,
|
||||
}
|
||||
|
||||
|
||||
@@ -22,7 +22,7 @@ from dataclasses import dataclass
|
||||
from typing import Any, Protocol
|
||||
|
||||
from nisse_dev.cancel import Cancel
|
||||
from nisse_dev.canonical import GateOpened, PlanSubmitted, UserMessage
|
||||
from nisse_dev.canonical import Fact, GateOpened, PlanSubmitted, UserMessage
|
||||
from nisse_dev.gates import (
|
||||
Approved,
|
||||
Gate,
|
||||
@@ -457,6 +457,11 @@ class Orchestrator:
|
||||
item = self._store.active_item()
|
||||
return self._store.transcript(item.id if item else None, actor)
|
||||
|
||||
def _fact(self, text: str) -> None:
|
||||
"""Record what the models must know and cannot see for themselves:
|
||||
every role is shown it, labelled as the orchestrator's (L6)."""
|
||||
self._log("orchestrator").append(Fact(text=text))
|
||||
|
||||
def _close_item(self, outcome: str, cause: str) -> None:
|
||||
item = self._store.active_item()
|
||||
if item is not None and item.status == "open":
|
||||
@@ -601,6 +606,7 @@ class Orchestrator:
|
||||
self._store.set_active(task_id, cause=f"operator command: use {project}")
|
||||
# What the operator said before `use` is part of this task (D1).
|
||||
self._store.carry_idle_messages(task_id)
|
||||
self._fact(f"work item opened: project {project}, task {task_id}")
|
||||
if self._board is not None:
|
||||
try:
|
||||
self._data["board"] = {
|
||||
@@ -738,6 +744,10 @@ class Orchestrator:
|
||||
channel,
|
||||
f"{name} rev-{rev} rejected by {result.approver}: {result.feedback}",
|
||||
)
|
||||
self._fact(
|
||||
f"{name} plan rev-{rev} rejected by {result.approver}; "
|
||||
"their feedback is the operator's message above"
|
||||
)
|
||||
self._data["gate"] = None
|
||||
back = PLANNING_ARCH if name == "arch" else MILESTONE_PLANNING
|
||||
self._transition(back, cause=f"plan rejected ({name} rev-{rev})")
|
||||
@@ -754,6 +764,7 @@ class Orchestrator:
|
||||
rev=f"rev-{rev}",
|
||||
channel=result.channel,
|
||||
)
|
||||
self._fact(f"{name} plan rev-{rev} approved by {result.approver}")
|
||||
if name == "arch":
|
||||
self._data["gate"] = None
|
||||
self._transition(
|
||||
@@ -774,6 +785,7 @@ class Orchestrator:
|
||||
)
|
||||
|
||||
def _execute_milestone(self, milestone: int) -> None:
|
||||
self._fact(f"milestone {milestone} started")
|
||||
if self._executor is None:
|
||||
# The stub completes at once, so it still leaves EXECUTING.
|
||||
self._transport.send_message(f"[stub] executing milestone {milestone}")
|
||||
@@ -783,6 +795,7 @@ class Orchestrator:
|
||||
with self._long_work():
|
||||
outcome = self._executor.run(self._log("executor"), milestone)
|
||||
if outcome.status == "stopped":
|
||||
self._fact(f"milestone {milestone} stopped by the operator")
|
||||
self._halt_for_stop(
|
||||
what=f"milestone {milestone}",
|
||||
note=outcome.note,
|
||||
@@ -794,9 +807,11 @@ class Orchestrator:
|
||||
)
|
||||
return
|
||||
if outcome.status == "completed":
|
||||
self._fact(f"milestone {milestone} finished: {outcome.note}")
|
||||
self._transport.send_message(f"milestone {milestone} done: {outcome.note}")
|
||||
self._publish_milestone(milestone)
|
||||
return
|
||||
self._fact(f"milestone {milestone} {outcome.status}: {outcome.note}")
|
||||
self._await_input(
|
||||
cause=f"milestone {milestone} {outcome.status}",
|
||||
prompt=f"milestone {milestone} stopped ({outcome.status}): {outcome.note}",
|
||||
@@ -860,8 +875,11 @@ class Orchestrator:
|
||||
):
|
||||
return # not taken: a watch attached later still sees it
|
||||
self._take_comment(board_state, comment)
|
||||
issue = (board_state or {}).get("issue")
|
||||
self._log("operator").append(
|
||||
UserMessage(text=comment.body, source=f"gitea:comment:{comment.id}")
|
||||
UserMessage(
|
||||
text=comment.body, source=f"gitea:issue:{issue}:comment:{comment.id}"
|
||||
)
|
||||
)
|
||||
self._apply_gate_reply(
|
||||
sender=comment.author,
|
||||
@@ -916,7 +934,10 @@ class Orchestrator:
|
||||
# The comment enters the transcript, so the executor sees review
|
||||
# feedback as conversation rather than as an out-of-band instruction.
|
||||
self._log("operator").append(
|
||||
UserMessage(text=comment.body, source=f"gitea:comment:{comment.id}")
|
||||
UserMessage(
|
||||
text=comment.body,
|
||||
source=f"gitea:pr:{pr.get('number')}:comment:{comment.id}",
|
||||
)
|
||||
)
|
||||
try:
|
||||
with self._long_work():
|
||||
@@ -959,6 +980,7 @@ class Orchestrator:
|
||||
if sha
|
||||
else "no code change was needed for that comment"
|
||||
)
|
||||
self._fact(f"review comment {comment.id}: {note}")
|
||||
self._transport.send_message(note)
|
||||
if pr.get("number"):
|
||||
self._publisher.comment(int(pr["number"]), note)
|
||||
@@ -973,6 +995,7 @@ class Orchestrator:
|
||||
def _publish_milestone(self, milestone: int) -> None:
|
||||
if self._publisher is None:
|
||||
# No PR to wait on, but PR_OPEN's commands still lead on.
|
||||
self._fact(f"milestone {milestone} done; no publisher, so no PR")
|
||||
self._transition(PR_OPEN, cause=f"[stub] milestone {milestone} done")
|
||||
self._transport.send_message(
|
||||
f"[stub] milestone {milestone} done; no publisher, so no PR — "
|
||||
@@ -991,6 +1014,7 @@ class Orchestrator:
|
||||
# The intent log already records what was in flight, so boot
|
||||
# reconciliation can settle it against Gitea. Gate rather than
|
||||
# retry blindly into a possible duplicate.
|
||||
self._fact(f"publishing milestone {milestone} failed: {exc}")
|
||||
self._await_input(
|
||||
cause=f"publishing milestone {milestone} failed: {exc}",
|
||||
prompt=f"could not publish milestone {milestone}: {exc}",
|
||||
@@ -1000,6 +1024,7 @@ class Orchestrator:
|
||||
if result.pull_request is None:
|
||||
# Nothing left EXECUTING behind: the operator decides what next,
|
||||
# and answering returns to the still-approved plan.
|
||||
self._fact(f"milestone {milestone} not published: {result.note}")
|
||||
self._await_input(
|
||||
cause=f"milestone {milestone}: {result.note}",
|
||||
prompt=(
|
||||
@@ -1009,6 +1034,10 @@ class Orchestrator:
|
||||
pending={"kind": "milestone_stopped", "milestone": milestone},
|
||||
)
|
||||
return
|
||||
self._fact(
|
||||
f"milestone {milestone} published: {result.note} "
|
||||
f"(PR #{result.pull_request.number}, {result.pull_request.url})"
|
||||
)
|
||||
self._unwatch("plan") # the gate is behind us; the PR is the thread now
|
||||
self._data["pr"] = {
|
||||
"number": result.pull_request.number,
|
||||
@@ -1036,6 +1065,7 @@ class Orchestrator:
|
||||
pr = self._data.get("pr") or {}
|
||||
project = self._project
|
||||
self.stop_watching()
|
||||
self._fact(f"work item abandoned by {sender}")
|
||||
self._close_item("abandoned", f"operator: abandon ({sender})")
|
||||
self._data = {**_fresh_data()}
|
||||
self._project = None
|
||||
@@ -1111,7 +1141,14 @@ class Orchestrator:
|
||||
command = command_form(text)
|
||||
if command == "next milestone":
|
||||
self._unwatch("review")
|
||||
pr = self._data.get("pr") or {}
|
||||
self._data["milestone"] = int(self._data.get("milestone", 1)) + 1
|
||||
self._fact(
|
||||
f"PR #{pr.get('number')} left for review; planning milestone "
|
||||
f"{self._data['milestone']}"
|
||||
if pr.get("number")
|
||||
else f"planning milestone {self._data['milestone']}"
|
||||
)
|
||||
self._data["gate"] = None
|
||||
self._data["pr"] = None
|
||||
self._transition(MILESTONE_PLANNING, cause="operator: next milestone")
|
||||
@@ -1119,6 +1156,7 @@ class Orchestrator:
|
||||
elif command == "project complete":
|
||||
board_state = self._data.get("board") or {}
|
||||
self.stop_watching()
|
||||
self._fact(f"work item completed: {sender} said project complete")
|
||||
self._close_item("completed", "operator: project complete")
|
||||
self._data = {**_fresh_data()}
|
||||
self._project = None
|
||||
|
||||
@@ -33,6 +33,7 @@ from nisse_dev.canonical import (
|
||||
AssistantText,
|
||||
DelegationCompleted,
|
||||
DelegationStarted,
|
||||
PlanSubmitted,
|
||||
StoredEvent,
|
||||
ToolCall,
|
||||
ToolResult,
|
||||
@@ -328,6 +329,7 @@ class Delegation:
|
||||
touched: set[str] = set()
|
||||
did_work = False
|
||||
spent = 0
|
||||
after = _plan_scope(transcript)
|
||||
|
||||
for turn in range(self._caps.turns):
|
||||
if self._cancel.requested:
|
||||
@@ -350,7 +352,7 @@ class Delegation:
|
||||
)
|
||||
try:
|
||||
produced = self._client.complete(
|
||||
transcript.read(),
|
||||
transcript.view(after=after),
|
||||
tools=DELEGATE_TOOLS,
|
||||
model=self._model,
|
||||
system=system,
|
||||
@@ -580,6 +582,16 @@ def delegation_verdict(
|
||||
return "resume"
|
||||
|
||||
|
||||
def _plan_scope(transcript: Transcript) -> int:
|
||||
"""Where the helper's view begins: at the approved plan the work is under,
|
||||
so it sees the plan and this milestone, not earlier ones (L6)."""
|
||||
latest = 0
|
||||
for stored in transcript.read():
|
||||
if isinstance(stored.event, PlanSubmitted):
|
||||
latest = stored.id
|
||||
return max(latest - 1, 0)
|
||||
|
||||
|
||||
def _completed_after_start(events: list[StoredEvent]) -> bool:
|
||||
"""True when the last delegation event in the log is a completion."""
|
||||
for stored in reversed(events):
|
||||
|
||||
@@ -47,6 +47,8 @@ def _event_line(kind: str, payload: dict[str, Any], full: bool) -> str:
|
||||
case "plan_submitted":
|
||||
plan = payload["plan"]
|
||||
return f"PLAN ({payload.get('model', '')}, {len(plan)} chars): {_clip(plan, full)}"
|
||||
case "fact":
|
||||
return f"FACT: {_clip(payload['text'], full)}"
|
||||
case "gate_opened":
|
||||
return f"GATE {payload.get('gate', '')}: {payload.get('reason', '')}"
|
||||
case _:
|
||||
|
||||
@@ -128,7 +128,7 @@ class LLMExecutor:
|
||||
f"Milestone {milestone} is stuck. The local executor tried and its "
|
||||
f"commands failed {failures} times in a row.\n\n"
|
||||
f"Most recent failure:\n{last_failure[:2000]}\n\n"
|
||||
"The approved plan and the full transcript are in the conversation "
|
||||
"The approved plan and the milestone so far are in the conversation "
|
||||
"above. Finish this sub-task, staying inside the approved plan."
|
||||
)
|
||||
result = self._delegate.run(
|
||||
@@ -214,7 +214,7 @@ class LLMExecutor:
|
||||
return ExecutorOutcome("stopped", f"before turn {turn + 1}")
|
||||
try:
|
||||
produced = self._client.complete(
|
||||
transcript.read(after=after),
|
||||
transcript.view(after=after),
|
||||
tools=tools,
|
||||
model=model,
|
||||
system=system,
|
||||
|
||||
@@ -151,7 +151,7 @@ class LLMPlanningAgent:
|
||||
def _turn(
|
||||
self, transcript: Transcript, model: str, system: str
|
||||
) -> SubmitPlan | AskHuman | None:
|
||||
events = transcript.read()
|
||||
events = transcript.view()
|
||||
produced = self._client.complete(
|
||||
events, tools=PLANNER_TOOLS, model=model, system=system
|
||||
)
|
||||
|
||||
@@ -16,6 +16,7 @@ from nisse_dev.canonical import (
|
||||
Compaction,
|
||||
DelegationCompleted,
|
||||
DelegationStarted,
|
||||
Fact,
|
||||
GateOpened,
|
||||
PlanSubmitted,
|
||||
StoredEvent,
|
||||
@@ -121,6 +122,8 @@ class OpenAIChatRenderer:
|
||||
"role": "assistant",
|
||||
"content": f"[delegation handback]\n{handback}",
|
||||
}
|
||||
case Fact(text=text):
|
||||
return {"role": "user", "content": f"[orchestrator] {text}"}
|
||||
case Compaction(summary=summary):
|
||||
return {
|
||||
"role": "user",
|
||||
|
||||
@@ -23,6 +23,7 @@ from nisse_dev.canonical import (
|
||||
Compaction,
|
||||
DelegationCompleted,
|
||||
DelegationStarted,
|
||||
Fact,
|
||||
GateOpened,
|
||||
PlanSubmitted,
|
||||
StoredEvent,
|
||||
@@ -152,6 +153,8 @@ class AnthropicMessagesRenderer:
|
||||
{"type": "text", "text": f"[delegation handback]\n{handback}"},
|
||||
)
|
||||
]
|
||||
case Fact(text=text):
|
||||
return [("user", {"type": "text", "text": f"[orchestrator] {text}"})]
|
||||
case Compaction(summary=summary):
|
||||
return [
|
||||
(
|
||||
|
||||
@@ -28,6 +28,7 @@ from nisse_dev.canonical import (
|
||||
to_payload,
|
||||
)
|
||||
from nisse_dev.oplog import OpLog
|
||||
from nisse_dev.views import project
|
||||
|
||||
CURRENT_SCHEMA_VERSION = 4
|
||||
|
||||
@@ -619,8 +620,14 @@ class Transcript:
|
||||
return self._store._append(event, self.work_item, self.actor)
|
||||
|
||||
def read(self, after: int = 0) -> list[StoredEvent]:
|
||||
"""The log as recorded: every role's events, for the record."""
|
||||
return self._store._read(self.work_item, after)
|
||||
|
||||
def view(self, after: int = 0) -> list[StoredEvent]:
|
||||
"""The log as this actor is shown it: what a model's context is
|
||||
built from (L6). See ``views``."""
|
||||
return project(self.read(after), self.actor)
|
||||
|
||||
def as_actor(self, actor: str) -> "Transcript":
|
||||
"""The same work item, written by another role: e.g. the executor
|
||||
handing a sub-task to the helper."""
|
||||
|
||||
@@ -0,0 +1,129 @@
|
||||
"""Per-role views of a work item's transcript (architecture L6; Development
|
||||
Plan M1.2).
|
||||
|
||||
The log records every role's events in one work item. A model must not read
|
||||
another role's turns as its own: in the second live run the planner, shown
|
||||
the executor's tool calls as assistant turns, took that work for its own
|
||||
(#78). So what a role is shown is projected here, before rendering:
|
||||
|
||||
- its own events, unchanged;
|
||||
- what the operator said, as a user message labelled with where it came from;
|
||||
- the orchestrator's facts and gates, as labelled statements;
|
||||
- another model role's work only where it needs it, and only as a labelled
|
||||
report or narration, never as a turn.
|
||||
|
||||
The projection keeps each event's id, so cursors (``after``) and compaction
|
||||
ranges still line up with the log. Everything not the viewer's own becomes a
|
||||
``UserMessage``: from where the viewer sits, it is all input.
|
||||
"""
|
||||
|
||||
from dataclasses import replace
|
||||
|
||||
from nisse_dev.canonical import (
|
||||
AssistantText,
|
||||
DelegationCompleted,
|
||||
Fact,
|
||||
GateOpened,
|
||||
PlanSubmitted,
|
||||
StoredEvent,
|
||||
ToolCall,
|
||||
ToolResult,
|
||||
UserMessage,
|
||||
)
|
||||
|
||||
# How much of another role's tool result a narration line keeps.
|
||||
NARRATION_RESULT_CHARS = 300
|
||||
|
||||
# The roles whose work the viewer is shown, and how: "report" keeps only the
|
||||
# handback, "narrate" keeps a line per step. Anything not listed is not shown.
|
||||
_SHOWN: dict[str, dict[str, str]] = {
|
||||
"planner": {},
|
||||
"driver": {"helper": "report"},
|
||||
"executor": {"helper": "report"},
|
||||
"helper": {"executor": "narrate", "driver": "narrate"},
|
||||
}
|
||||
|
||||
|
||||
def operator_label(source: str) -> str:
|
||||
"""Where an operator message came from, for its label."""
|
||||
origin, _, idle = source.partition(" (sent while idle")
|
||||
if origin.startswith("transport:"):
|
||||
channel = "Matrix"
|
||||
elif origin.startswith("gitea:pr:"):
|
||||
number = origin.split(":")[2]
|
||||
channel = f"comment on PR #{number}"
|
||||
elif origin.startswith("gitea:issue:"):
|
||||
number = origin.split(":")[2]
|
||||
channel = f"comment on plan issue #{number}"
|
||||
elif origin.startswith("gitea:"):
|
||||
channel = "Gitea comment"
|
||||
else:
|
||||
channel = origin or "unknown"
|
||||
if idle:
|
||||
channel += ", sent while idle"
|
||||
return f"operator, {channel}"
|
||||
|
||||
|
||||
def _said(stored: StoredEvent, label: str, text: str) -> StoredEvent:
|
||||
return replace(
|
||||
stored,
|
||||
event=UserMessage(text=f"[{label}] {text}", source=f"view:{stored.actor}"),
|
||||
)
|
||||
|
||||
|
||||
def _narration(call: ToolCall, result: ToolResult | None) -> str:
|
||||
target = call.arguments.get("path") or call.arguments.get("command")
|
||||
step = f"{call.tool}({str(target)[:80]})" if target else call.tool
|
||||
if result is None:
|
||||
return f"{step} → no result recorded"
|
||||
outcome = "ok" if result.success else "failed"
|
||||
text = result.result.strip()
|
||||
if len(text) > NARRATION_RESULT_CHARS:
|
||||
text = text[:NARRATION_RESULT_CHARS] + " …"
|
||||
return f"{step} → {outcome}: {text}" if text else f"{step} → {outcome}"
|
||||
|
||||
|
||||
def project(events: list[StoredEvent], viewer: str) -> list[StoredEvent]:
|
||||
"""The events as ``viewer`` should be shown them, in log order."""
|
||||
shown = _SHOWN.get(viewer)
|
||||
if shown is None:
|
||||
# A role with no view defined (the expert is stateless, the
|
||||
# classifier sees only an action) reads the log as it is.
|
||||
return list(events)
|
||||
results = {
|
||||
e.event.call_id: e.event for e in events if isinstance(e.event, ToolResult)
|
||||
}
|
||||
out: list[StoredEvent] = []
|
||||
for stored in events:
|
||||
event, actor = stored.event, stored.actor
|
||||
if isinstance(event, PlanSubmitted):
|
||||
# A record for the executor's scope and the dump, not a turn:
|
||||
# the planner's own submit_plan call already carries the plan.
|
||||
if viewer == "helper":
|
||||
out.append(_said(stored, "approved plan", event.plan))
|
||||
continue
|
||||
if actor == viewer:
|
||||
out.append(stored)
|
||||
elif actor == "operator" and isinstance(event, UserMessage):
|
||||
out.append(_said(stored, operator_label(event.source), event.text))
|
||||
elif isinstance(event, Fact):
|
||||
out.append(_said(stored, "orchestrator", event.text))
|
||||
elif isinstance(event, GateOpened):
|
||||
out.append(
|
||||
_said(stored, "orchestrator", f"gate {event.gate}: {event.reason}")
|
||||
)
|
||||
elif shown.get(actor) == "report" and isinstance(event, DelegationCompleted):
|
||||
out.append(
|
||||
_said(
|
||||
stored,
|
||||
f"{actor} report, {event.exit_status}",
|
||||
event.handback,
|
||||
)
|
||||
)
|
||||
elif shown.get(actor) == "narrate":
|
||||
if isinstance(event, ToolCall):
|
||||
line = _narration(event, results.get(event.call_id))
|
||||
out.append(_said(stored, actor, line))
|
||||
elif isinstance(event, AssistantText) and event.text.strip():
|
||||
out.append(_said(stored, actor, event.text.strip()))
|
||||
return out
|
||||
Reference in new issue
Block a user