Found while running the #14122 reader program. Three separate dev seats reported the same contract conflict in their delivery reports, each resolving it differently and each flagging the disagreement rather than resolving it silently — which is the right behaviour and is why it is filed rather than patched over.
The conflict
Two rules, both currently true, that cannot both be followed:
| Source |
Says |
AGENTS.md:450-456 (Prime Directive; CLAUDE.md inlines it verbatim) |
"Assign it to yourself … as the first action of the task — before the worktree, before the first read." and "Already assigned to someone else? It is taken — pick another or ask; never reassign it to yourself." |
.claude/skills/pm-dispatch/SKILL.md:516 |
"assignee 字段归 PM:原子对 step 1 设,dev 席恒不写它(os-dev 规则 2 同句)" |
A dev seat dispatched onto a card therefore reads: assign yourself first (AGENTS.md) and never write the assignee field (its own seat contract). Both are binding on it. The seats that hit this took the seat contract and recorded the disagreement; that is the correct precedence, but nothing in AGENTS.md says so.
Why it just got sharper
Until today the conflict was mostly latent, because the dispatched cards were arriving unassigned — the PM had not executed step 1 of the atomic pair, so the dev seat saw an open invitation and only the "assign yourself" half was live. That was a PM-side half-state (five cards of one batch: #15005, #15006, #15007, #15229, #15232 — all pm:queue, all unassigned, all with an in-flight dispatch), and it has now been corrected: assigned, and the label swapped to pm:dispatched.
Correcting it activates the other half of the conflict, and this one is worse. A dev seat dispatched onto one of those cards now reads AGENTS.md:454 — "already assigned to someone else? It is taken — pick another" — and yields a card it was just dispatched to. Yielding a dispatched card is more expensive than double-claiming one: the PM waits on a seat that has already walked away, and nothing on the card records why.
The shape of the fix
Not a weakening of "claim before you write code" — that rule is load-bearing and its failure story (one issue implemented twice in one morning) is real. The missing sentence is that the assignee field can arrive pre-set by the dispatching PM on the seat's behalf, and how to tell that case from a genuine prior claim:
- a claim is still two acts, and the dev seat still performs the second one (the claim comment carrying its session ID and branch);
- when a dispatch order names the seat and the card is already assigned to the dispatching account, that assignment is step 1, already done — the seat does not re-write it and does not yield;
- the discriminator stays exactly where
AGENTS.md:457-460 already puts it: re-read the comments. An earlier claim comment with a different session ID or branch means taken. A bare assignee with no such comment does not.
That keeps one rule ("the comments are what tell you whether a claim is yours") and removes the contradiction rather than adding an exception to it.
Surfaces to touch
AGENTS.md — the Prime Directive paragraph at :450.CLAUDE.md — it inlines the same paragraph and must not drift from it.- possibly
.claude/skills/pm-dispatch/SKILL.md:516 and the os-dev seat definition, so the two sides cite each other rather than each asserting alone.
⛔ Governed surfaces (AGENTS.md, CLAUDE.md, .claude/**) — no auto-merge; a human reviewer decides.
Found while running the #14122 reader program. Three separate dev seats reported the same contract conflict in their delivery reports, each resolving it differently and each flagging the disagreement rather than resolving it silently — which is the right behaviour and is why it is filed rather than patched over.
The conflict
Two rules, both currently true, that cannot both be followed:
AGENTS.md:450-456(Prime Directive;CLAUDE.mdinlines it verbatim).claude/skills/pm-dispatch/SKILL.md:516A dev seat dispatched onto a card therefore reads: assign yourself first (AGENTS.md) and never write the assignee field (its own seat contract). Both are binding on it. The seats that hit this took the seat contract and recorded the disagreement; that is the correct precedence, but nothing in
AGENTS.mdsays so.Why it just got sharper
Until today the conflict was mostly latent, because the dispatched cards were arriving unassigned — the PM had not executed step 1 of the atomic pair, so the dev seat saw an open invitation and only the "assign yourself" half was live. That was a PM-side half-state (five cards of one batch: #15005, #15006, #15007, #15229, #15232 — all
pm:queue, all unassigned, all with an in-flight dispatch), and it has now been corrected: assigned, and the label swapped topm:dispatched.Correcting it activates the other half of the conflict, and this one is worse. A dev seat dispatched onto one of those cards now reads
AGENTS.md:454— "already assigned to someone else? It is taken — pick another" — and yields a card it was just dispatched to. Yielding a dispatched card is more expensive than double-claiming one: the PM waits on a seat that has already walked away, and nothing on the card records why.The shape of the fix
Not a weakening of "claim before you write code" — that rule is load-bearing and its failure story (one issue implemented twice in one morning) is real. The missing sentence is that the assignee field can arrive pre-set by the dispatching PM on the seat's behalf, and how to tell that case from a genuine prior claim:
AGENTS.md:457-460already puts it: re-read the comments. An earlier claim comment with a different session ID or branch means taken. A bare assignee with no such comment does not.That keeps one rule ("the comments are what tell you whether a claim is yours") and removes the contradiction rather than adding an exception to it.
Surfaces to touch
AGENTS.md— the Prime Directive paragraph at:450.CLAUDE.md— it inlines the same paragraph and must not drift from it..claude/skills/pm-dispatch/SKILL.md:516and theos-devseat definition, so the two sides cite each other rather than each asserting alone.⛔ Governed surfaces (
AGENTS.md,CLAUDE.md,.claude/**) — no auto-merge; a human reviewer decides.