[trissim]

Waste flags itself on the board now instead of waiting for someone to read #comms. 2 commits, 7 files, +310/−25.

SEQUENCE → CONFLICT, only when Git says so

Before: SharedFileBranches emitted one SEQUENCE … who goes first? row per shared file between two held orders. No answer could clear it, because sharing a file is ordinary parallel work. Tonight's board carried 12 of these across 8 order pairs, and units learned to skim the board.

After: ConflictingBranches emits one CONFLICT row per order pair, and only when git merge-tree --write-tree of the two branch heads conflicts. The row names both orders, both branches and the conflicting paths, and says "merge one into the other and resolve". The merge clears it.

• HolderBranch.parallel_to owns the pair filter (different holders, same repository, different branches, shared files), and both PreparedBranch and HeldBranch inherit it. merge-tree runs only on candidates.
• merge-tree runs in prepare, before the cut. HeldBranchRule.prepare hands each member the prepared branches via prepare_captured (a no-op except on ConflictingBranches). The answer is captured in ProgressRead under the unordered commit pair, and derive reads it with require_branch_conflicts. The scan opens no Git, as test_held_branch_rules_consume_the_complete_cut_after_worktree_mutation requires.
• Repo.merge_conflicts owns merge-tree's exit contract: 0 merges cleanly, 1 lists the conflicting paths, anything else is a FailedGitCommand.
• DriftOverlapRulesPhase evidence now names ConflictingBranches, and the fleet-replay seed is renamed to match.

Replay on tonight's actual heads (read-only):

pair	merge-tree	row
P927 × P957, P927 × P924, P954 × P958, P954 × P924, P957 × P958	clean	none
P957 × P918	3 conflicts (commander_hud.py, order_tools.py, turn_digests.py)	CONFLICT
P918 × P958	4 conflicts (bus.py, goal_scheduler.py, maintenance_barrier.py, …)	CONFLICT
P918 × P924	5 conflicts (codex_protocol.py, external_transcript.py, harness_delivery_states.py, …)	CONFLICT

12 rows → 3. All three are P918, which is already decided to land last.

BlockedTalk (rules/no-blocking.md)

A commanded unit's published message that announces a stop gets a named row citing its message id. The phrases are declared once in BlockedTalk.words ("waiting on", "awaiting", "blocked on", "frozen for", "until root", …). A message is skipped when it names an owner (an @thread or a recorded mention), carries a file:line, or negates the phrase ("not waiting on").

Replay of the last 6 hours: 14 rows, for example:

• claude-nra, three in a row at 13:46–13:48: "Still waiting on the judge's verdict for Test the public fields that pinned Toad builds read from agent-comms #7 and debt_ratchet.py is only the CI entry; the pinned audit owns every measure #3606."
• install-speed, 14:17: "Source delivered; awaiting root review."
• commander-layout and hud-repair-pi, 14:35–14:36: "frozen for your review".
• slot-activity, 17:32: "ban stays in force until root supplies qualification".

One false positive: board-commander quoting the phrase "'waiting on the run' labels". Bare "freeze"/"preserving"/"pending" were dropped after the first replay, which hit "the author freeze is superseded" and "preserving your … behaviour".

Ceremony

A unit holding pending work that published more than Ceremony.per_commit = 6000 characters per commit over the last hour gets a row naming both numbers. The rule reads the bus sent index and the held orders' existing progress sources; no new store.

Replay over 6 hours: 0 rows for units holding work. The busiest was claude-nra, 19,166 characters over 7 commits. board-commander, which holds no order and is forbidden to code, is excluded by the pending-work condition.

Shape

UnitTalk is one window over a unit's observation, and UnitTalkRule is a FamilyScan over the heartbeat's existing UnitState observations. Its members BlockedTalk (dated by the message) and Ceremony (undated) are discovered via the agent_comms.commander package. No kind strings, and no new reads beyond the bus sent index.

Proof

• tests/test_unit_talk.py: 10 passed. It covers unowned waits flagging; owned, located, negated and ordinary messages not flagging; and the ceremony threshold, including the no-pending case.
• The fleet-replay seed for ConflictingBranches uses an add/add conflict on shared.py.

Assumed: merge-tree --write-tree leaves unreachable tree and blob objects in the order's repository during prepare, which git gc prunes; no ref moves. The replay above read the bus with sqlite3 -readonly and computed commits with git log in each unit's worktrees, mirroring what the rules read through Core. Row counts on the live board appear only after install.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[github-actions]

debt-src at 2951e35, every finding:

grew BooleanChainTerms:src/agent_comms/standing_rules.py +4
::warning file=src/agent_comms/commander/unit_talk.py::LintFindings +1 (warns, never blocks)
grew LongBooleanChain +1
::warning file=src/agent_comms/commander/unit_talk.py::Unformatted +1 (warns, never blocks)
found BooleanChainTerms:src/agent_comms/standing_rules.py at src/agent_comms/standing_rules.py:422 src/agent_comms/standing_rules.py:422 src/agent_comms/standing_rules.py:422 src/agent_comms/standing_rules.py:422
found LongBooleanChain at src/agent_comms/standing_rules.py:422 src/agent_comms/todos.py:4133