Integrate #2023 #1928 #2064 #2069 #2076

.github/workflows/debt-ratchet.yml

@@ -1,47 +1,84 @@
-name: Debt ratchet
+name: Merge queue
 
 on:
-  pull_request:
-    branches: [main]
+  status:
+  merge_group:
+    types: [checks_requested]
   workflow_dispatch:
     inputs:
-      base:
-        required: true
-        type: string
-      head:
-        required: true
-        type: string
+      enter_queue:
+        description: Enqueue the already reviewed head after successful checks
+        type: boolean
+        default: false
 
 permissions:
   contents: read
 
 concurrency:
-  group: debt-ratchet-${{ github.ref }}
-  cancel-in-progress: true
+  group: merge-queue-${{ github.event.sha || github.sha }}
+  cancel-in-progress: false
 
 jobs:
-  debt-ratchet:
-    name: Debt ratchet
-    runs-on: ubuntu-latest
+  dispatch-reviewed:
+    if: github.event_name == 'status' && github.event.state == 'success'
+    runs-on: [self-hosted, Linux, X64, ts-gaming, agent-comms-merge]
     timeout-minutes: 5
     steps:
-      - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-          submodules: recursive
-          ref: ${{ github.event.pull_request.head.sha || inputs.head }}
-      - uses: actions/setup-python@v5
-        with:
-          python-version-file: pyproject.toml
-      - name: Install development dependencies
-        run: python -m pip install --disable-pip-version-check -e src/agent_comms/skills/refactor-audit '.[all,dev]'
-      - name: Compare touched production files
+      - name: Dispatch checks on the approved PR head, not the default-branch status event
         env:
-          PR_BASE: ${{ github.event.pull_request.base.sha || inputs.base }}
-          PR_HEAD: ${{ github.event.pull_request.head.sha || inputs.head }}
-        run: python tools/debt_ratchet.py --root src/agent_comms --base "$PR_BASE" --head "$PR_HEAD"
-      - name: Run every marked surface guard
-        if: ${{ github.event_name == 'workflow_dispatch' }}
+          RUN_ID: ${{ github.run_id }}
+          REVIEWED_HEAD: ${{ github.event.sha }}
+          REVIEW_CONTEXT: ${{ github.event.context }}
+        run: |
+          set -euo pipefail
+          request=$(python3 -c 'import shlex, sys; print(shlex.join(sys.argv[1:]))' \
+            "$RUN_ID" dispatch "$REVIEWED_HEAD" "$REVIEW_CONTEXT")
+          ssh -o BatchMode=yes -o StrictHostKeyChecking=yes \
+            -i "$AGENT_COMMS_GATE_KEY" "$AGENT_COMMS_GATE_ENDPOINT" "$request"
+
+  ratchet:
+    if: github.event_name != 'status'
+    name: debt-${{ matrix.kind }}
+    runs-on: [self-hosted, Linux, X64, ts-gaming, agent-comms-merge]
+    timeout-minutes: 15
+    strategy:
+      fail-fast: false
+      matrix:
+        kind: [src, tools]
+    steps:
+      - name: Check the exact GitHub run through the restricted host entrypoint
+        env:
+          RUN_ID: ${{ github.run_id }}
+          CHECK_KIND: ${{ matrix.kind }}
+        run: |
+          set -euo pipefail
+          ssh -o BatchMode=yes -o StrictHostKeyChecking=yes \
+            -i "$AGENT_COMMS_GATE_KEY" "$AGENT_COMMS_GATE_ENDPOINT" "$RUN_ID $CHECK_KIND"
+
+  practice-install:
+    if: github.event_name != 'status'
+    name: install-rehearsal
+    runs-on: [self-hosted, Linux, X64, ts-gaming, agent-comms-merge]
+    timeout-minutes: 60
+    steps:
+      - name: Practice install; private buses and logs never enter the runner sandbox
+        env:
+          RUN_ID: ${{ github.run_id }}
+        run: |
+          set -euo pipefail
+          ssh -o BatchMode=yes -o StrictHostKeyChecking=yes \
+            -i "$AGENT_COMMS_GATE_KEY" "$AGENT_COMMS_GATE_ENDPOINT" "$RUN_ID install"
+
+  enqueue-approved:
+    if: github.event_name == 'workflow_dispatch' && inputs.enter_queue
+    needs: [ratchet, practice-install]
+    runs-on: [self-hosted, Linux, X64, ts-gaming, agent-comms-merge]
+    timeout-minutes: 5
+    steps:
+      - name: Enter GitHub's queue after approval and successful checks
         env:
-          PYTEST_DISABLE_PLUGIN_AUTOLOAD: '1'
-        run: python -m pytest -p pytest_asyncio.plugin -o addopts='' --strict-markers -m refactor_guard -q
+          RUN_ID: ${{ github.run_id }}
+        run: |
+          set -euo pipefail
+          ssh -o BatchMode=yes -o StrictHostKeyChecking=yes \
+            -i "$AGENT_COMMS_GATE_KEY" "$AGENT_COMMS_GATE_ENDPOINT" "$RUN_ID enqueue"

.pi/APPEND_SYSTEM.md

@@ -56,6 +56,11 @@ You are one of several agents coordinating through agent-comms. Tristan owns pro
 
 ## Ownership, code, and delivery
 
+- Do not merge main into feature branches. When your own branch conflicts with main, rebase it onto origin/main that turn, keep main's side where main deleted or moved code, push the owned branch, and report its new head. Do not ask again for that routine authorization. Do not make formatting-only or re-budget-only commits.
+- Fresh-environment practice installs belong to the merger, on the pushed head. Workers prove the affected behavior on a copy of the bus using existing dependencies and report precisely which installation proof the merger still owes.
+- Line growth is a measurement, not a defect or merge gate. Do not request a line budget. Fix duplicated authorities, type filters and strings replacing declared identities; the existing measure owner decides which findings block.
+- P269 practice installation uses a retained bus tail with its production indexes, bounded to 3 GiB and measured against actual free space. It does not require a 26 GiB whole-desktop copy or GitHub configuration changes. Preserve original sessions, inputs and stores; never turn short storage into a skipped success.
+
 - Check active owners, open PRs, and claims before starting overlapping work. Coordinate with the owner of an existing feature; do not edit their worktree or start a competing implementation. Delegate genuinely independent work when it speeds delivery.
 - At the next safe checkpoint, when you find a concrete defect, check that a named owner is fixing it. If none is, take or assign ownership, provide the reproducer and acceptance check, and follow through to a tested fix or a precise blocker without duplicating work.
 - Implement working end-to-end behavior rather than a dormant path. Enable features by default only when safe and authorized; preserve opt-in or default-off safeguards for trust-sensitive features. Put tested work in a draft PR promptly and keep its full working version visible. Work only in an isolated worktree; do not reset, clean, or commit in the live main checkout or change installed packages without review.

docs/merge-queue.md

@@ -0,0 +1,92 @@
+# Automatic merge queue
+
+Owner: pi-cn-load, P269, draft PR #2023. Not enabled or accepted.
+The takeover preserved pi-install-gate's newer published 307e5955e, a descendant
+of the ordered f0dc3df0c checkpoint, and merged current main without rewriting
+history. First takeover push: fb6074f3e. P714/#1947 is merged as the review-owner
+dependency, not copied into another verdict implementation.
+
+GitHub owns queue order, proposed merge identity and failed-entry removal.
+QueueCheck and RunEvent members own host execution and admission. The existing
+workflow runs debt-src, debt-tools and install-rehearsal on merge_group.
+No commander's merge scheduler is involved after bootstrap.
+
+## Review and installer ownership
+
+ReviewedHeadVerdict.context is the sole review-context declaration. Its existing
+publisher and the queue reader both use it. The status event passes its context
+as data; the host compares it with that declaration. No repository variable
+mirrors the context. No native GitHub approval is required: the contributors all
+use the same GitHub identity and cannot approve their own PRs.
+
+The queue invokes PracticeGateCommand. PracticeCopy uses the existing CheckRoot
+owner to retain the bus tail, bounded saved-session tails and their production
+indexes; it does not snapshot the whole desktop bus. PracticeCopy owns the 3 GiB
+cap and checks measured bytes against current free space. It never admits an
+empty substitute or a skipped installation. The gate checks out the proposed
+revision, then rehearse invokes target/tools/upgrade_bus.py, including --help and
+the actual rehearsal. It does not invoke the live checkout's installer. GateResult/MachineResult carry
+typed failure reasons back to the queue. CheckFailed carries the actual completed command and delegates its public
+summary to the QueueCheck member. Ratchet summaries decode introduced_lines
+and print Measure file:line. Installer summaries return the declared
+UpgradeRefused message through the existing typed gate result. Raw child
+stdout/stderr, successful rehearsal steps and bus data remain private.
+
+## Isolation and lifetime
+
+The existing runner ts-gaming-agent-comms-merge (ID 21) is registered but offline.
+Labels: self-hosted, Linux, X64, ts-gaming, agent-comms-merge. Its bubblewrap
+namespace exposes vendor runtime, work/log directories, system binaries and a
+sealed runner home, not host homes, buses, personal tokens or desktop sockets.
+The SSH key admits only the fixed host entrypoint. GitHub run identity and the
+workflow path are verified before checks. No generic hosted runner reads a bus.
+
+The host derives home/user from its environment and OS identity. There is no audit path on the wrapper's PYTHONPATH: debt_ratchet derives it
+from the repository submodule itself. Host.fetch initializes/updates the pinned
+submodules. No unit checkout is a dependency of production gates.
+There is no shared host wrapper timeout. Ratchet members own 600 seconds,
+install owns 3000, and enqueue/dispatch share their command capability's 120.
+YAML ceilings are 15, 60 and 5 minutes respectively. Local development ratchets
+retain the ordered timeout 300 and faulthandler.
+
+Only explicit gate failure reasons and source annotations cross the runner
+boundary. Full child diagnostics, copied state and successful rehearsal results
+remain on the host. A storage or allocation refusal fails the gate; it never
+turns into a skipped or successful practice install.
+
+## Settings Tristan must change (no settings changed by this worker)
+
+- Settings → Rules → Rulesets → Automatic merge queue (25439636): change Enforcement from Disabled to Active for main, after reviewed bootstrap code is on main.
+- In that ruleset → Merge queue: Merge method Merge; maximum group/build/merge size 1; require debt-src, debt-tools and install-rehearsal from GitHub Actions.
+- In that ruleset → Require a pull request before merging: set required approvals to 0; the staged rule currently incorrectly requires 1 native approval.
+- Settings → Actions → Runners: reuse registered runner ts-gaming-agent-comms-merge (ID 21), with the labels above; do not register a duplicate. If it must be replaced, use New self-hosted runner → Linux → x64 with those same custom labels.
+
+Runner registration does not start its host process. The previous owner confirms
+that its restricted key was never added to authorized_keys and no runner
+process/service has started. An authorized operator must install the reviewed
+fixed-command, loopback-only key restriction and start the reviewed isolated
+launcher from main, retaining the existing private runner credentials. Do not
+launch the previous owner's queue-qualified/ scripts; those are old qualification
+evidence, not installed entrypoints. No runner launch, authorized_keys edit,
+queue enablement, native approval, production install or restart was performed
+by this worker.
+
+## Practice proof and later activation
+
+The latest owner ruling replaces real-queue proof with a practice installation
+on the bounded retained-tail copy. Report that command's exit code and the pushed
+revision; unit checks alone are not the practice-install proof. Both source and
+tools ratchets must exit zero. SourceLines reports actual growth without a
+budget or blocking the merge; structural findings still block through their
+owning Measure implementations.
+
+The disabled queue, offline runner and unauthorized key are not prerequisites
+for this proof. They remain untouched; later production activation requires the
+operator actions above. No deliberate failing GitHub branch is claimed here.
+The copy does not need room for a 26 GiB desktop snapshot. Its admission uses
+actual copied bytes and actual remaining space, not a second queue policy.
+
+Worker checkouts and receipts stay under /home/ts/wt/pi-cn-load--P269 and
+/home/ts/.cache/agent-scratch/pi-cn-load, below the combined 3 GB limit. The old
+runner state remains with its existing owner; its credentials have not been
+copied or deleted.

plans/commander-native-plan.md

@@ -206,4 +206,4 @@ Toad is a view of Core, so the commander's page is one more view over Core's ans
 
 ## Order and dependencies
 
-The phases and what each waits for are declared once, in `src/agent_comms/plan_phases.py` (the `PlanPhase` family). An order names the phase it serves with a `Serves: <phase>` line, stored as `ProblemOrder.serves`; a phase is done when every order serving it is closed. Each rotation prints `PLAN IDLE` for every phase whose dependencies are done and that has no open order (`PlanIdle` in `src/agent_comms/standing_rules.py`).
+The phases and what each waits for are declared once, in `src/agent_comms/plan_phases.py` (the `PlanPhase` family). An order names the phase it serves with a `Serves: <phase>` line, stored as `ProblemOrder.serves`. A phase is done when its explicitly declared full-phase checked task has accepted installed evidence; closing related subtasks does not complete the phase. Earlier delivered phases record their delivery on their declarations. Each rotation prints `PLAN IDLE` for every phase whose dependencies are done and that has no open order (`PlanIdle` in `src/agent_comms/standing_rules.py`).

src/agent_comms/check_root.py

@@ -17,7 +17,6 @@
 from .comms import Comms
 from .fresh_private_session import create_fresh_private_session
 from .goals import NoGoal
-from .importing import ImportLimits
 from .launch_runtime import RestartEnvironment, ThreadLaunch
 from .native_entries import ImportedSessionWriter, NativeEntry
 from .read_basis import MemberActivityDisplayScope
@@ -66,8 +65,8 @@ def open(self) -> Iterator[Comms]:
         root.mkdir(mode=0o700, parents=True)  # Refuse an existing directory, including a symlink.
         try:
             copied = Comms(root)
+            self._populate(copied)
             try:
-                self._populate(copied)
                 yield copied
             finally:
                 for thread in copied.registry.snapshot().visible_threads(
@@ -142,7 +141,6 @@ def _populate(self, copied: Comms) -> None:
     def _session(self, original: Thread, directory: Path, project: Path) -> str:
         session = create_fresh_private_session(directory, worktree=project)
         source = original.runtime.history_source(self.source.root, original.incarnation)
-        limits = ImportLimits(messages=self.max_records, characters=self.max_bytes)
         with source.reading() as reader, session.path.open("r+") as output:
             writer = ImportedSessionWriter(output, NativeEntry.read(output.readline().encode()))
             output.seek(0, os.SEEK_END)
@@ -158,16 +156,13 @@ def _session(self, original: Thread, directory: Path, project: Path) -> str:
                 with source.projection(
                     source, TranscriptRoutes(self.source.root)
                 ) as project_records:
-                    for record, events in project_records(
+                    for _record, events in project_records(
                         iter(reversed(selected)),
                         max_records=self.max_records,
                         max_bytes=self.max_bytes,
                     ):
-                        record.project(
-                            lambda current, events=events: writer.import_record(
-                                current.entry, events, limits
-                            )
-                        )
+                        for event in events:
+                            writer.append_recorded(event)
         return session.session_file
 
     def run(self, command: Sequence[str]) -> CheckRootResult:

src/agent_comms/commander/judge.py

@@ -20,7 +20,7 @@
 from dataclasses import asdict, dataclass
 
 from ..comms import Comms
-from ..todos import AcceptedVerdict, PeerAcceptedVerdict, BluntText, OrderVerdict, SendBackVerdict, StrikeReason, Todo, Slop, Derivation, SupersededVerdict
+from ..todos import AcceptedVerdict, PeerAcceptedVerdict, BluntText, ToolCallVerdict, SendBackVerdict, StrikeReason, Todo, Slop, Derivation, SupersededVerdict
 from ..tools import CommsUnjudgedTool, CommsVerdictTool
 from ..field_codec import FieldCodec, option
 from ..tool_caller import ThreadCaller
@@ -146,7 +146,7 @@ class VerdictVerb(JudgeVerb, RuledVerb):
         "--line", action="append", metavar="LINE", help="an exact diff line; repeat for each")
 
     @abstractmethod
-    def operation(self, ctx: order.Commander, todo: Todo, edit_id: str) -> OrderVerdict: ...
+    def operation(self, ctx: order.Commander, todo: Todo, edit_id: str) -> ToolCallVerdict: ...
 
     @abstractmethod
     def check_edits(self, edits: list[Edit]) -> None: ...

src/agent_comms/commander/rules/always-fork-subagents.md

@@ -0,0 +1,13 @@
+---
+name: always-fork-subagents
+description: "Owner rule: dispatch subagents as forks (subagent_type \"fork\"), not fresh agents, unless the brief fully teaches correct maintenance and the paper"
+type: feedback
+---
+
+Always fork (Agent with `subagent_type: "fork"`) instead of starting a fresh general-purpose agent. Owner, 2026-10-18: "always fork. stop it, fork and let the fork take over", then "unless you already schooled it on correct maintenance and the paper".
+
+**Why:** a fresh agent knows only its brief. My briefs carried the owner's rules (multiple inheritance, no back-compat) but not the paper's definition of slop and the results behind it, so a fresh agent judges shapes by dogma. A fork inherits the conversation, including that understanding (, ).
+
+**How to apply:** default to forks for refactor and design work. A fresh agent is acceptable only when its brief teaches correct maintenance itself (slop = a restated required fact a complete mechanism derives; multiple parents derive crossings; mirrors add disagreeing states; passing tests cannot reveal outside supply) and points at `.claude/verify/PAPER-RESULTS.md`. Supersedes the older habit of fresh agents with written briefs.
+
+**Turn limit (2026-10-18):** owner said "don't use that [turn limit] on ur subagents". Forks stop at ~200 turns and can't be given a limit; they have no definition file (docs: code.claude.com/docs/en/sub-agents.md, `maxTurns` frontmatter). For long work use the user-level agent `correct-maintainer` (`~/.claude/agents/correct-maintainer.md`, `maxTurns: 100000`, system prompt teaches the standard). Definitions load at session start only. If a fork hits the cap anyway, resume it via SendMessage right away.

src/agent_comms/commander/rules/claude-usage-budget.md

@@ -0,0 +1,15 @@
+---
+name: claude-usage-budget
+description: Claude subagent usage is the scarce budget; keep background forks to the critical path until Pi (gpt-6.1-sol) threads work on the bus
+type: feedback
+---
+
+2026-10-19: I had 7 background forks running at once: runtime, carry-order, strings, index-lock, hooks, fast-ratchet and fast-board. Owner: "don't launch too many background agents, I'm trying to move to Pi Codex threads to save on usage, so please be mindful. If it's going to take another 5 hours to get Pi GPT working in comms we can't just blow through usage."
+
+**Why:** the move to Pi threads exists to save Claude usage. Spending Claude forks freely while that move is blocked defeats it.
+
+**How to apply:**
+- Only work on the path to working Pi threads gets a Claude fork; aim for one or two at a time.
+- Everything else (perf, cleanup, slop hunts) becomes a written order queued for Pi units, not a fork.
+- Briefs say "finish and report in N lines", with no exploration beyond the task.
+- This tempers [[slow-is-a-defect]] and [[structural-fixes-over-pushing-through]]: dispatch still happens, but to Pi units once they work. See [[codex-units-need-dictation]].