docs(agents): add the parallel-session claim protocol to the shared contract

.claude/agents/portfolio-surveyor.md

@@ -29,11 +29,30 @@ public and private — no per-repo loop needed to enumerate):
 
 1. **Open PRs (org-wide, one call):**
    `gh search prs --owner devantler-tech --archived=false --state open --limit 300 --json number,repository,title,author,isDraft,labels,updatedAt,url`
-2. **Open issues (org-wide, one call):**
-   `gh search issues --owner devantler-tech --archived=false --state open --limit 300 --json number,repository,title,labels,updatedAt,url`
+2. **Open issues (org-wide, one call) — include `assignees`, they are a CLAIM signal:**
+   `gh search issues --owner devantler-tech --archived=false --state open --limit 300 --json number,repository,title,labels,updatedAt,url,assignees`
    (`--archived=false` keeps archived repos' stale PRs/issues — e.g. `data-product`'s 2025 bot PRs —
    out of every survey; archived repos are read-only and carry no actionable signal.)
    (`gh search issues` returns issues only — not PRs; treat label-less issues as untriaged.)
+   Report assignee **logins**, not a count. Only a **`devantler`** assignment can be a claim: the
+   contract's *Claim protocol* lease is specifically the agent account's, and every instance assigns as
+   `devantler`, so that login means **an instance has claimed this** — never "the maintainer took it".
+   An issue assigned to **anyone else** (a human collaborator, `Copilot`) is **not** a claim even with a
+   leftover `claude/*-<issue>` branch present: reporting it as one would park actionable work behind an
+   unrelated person's assignment and time the lease off the wrong assignment event. Report those as
+   ordinary open issues, noting the assignee so the orchestrator can respect a human's in-progress work
+   on its own merits. Without these logins the orchestrator selects the oldest issue blind to live
+   claims and re-opens the duplicate-build race the protocol exists to close.
+2b. **Claim branches (one call per repo that has assigned-but-PR-less issues):**
+   `gh api repos/<o>/<r>/branches --paginate --jq '.[].name' | grep '^claude/'` — report any
+   `claude/*` branch that ends in `-<issue>`, ends in a **takeover suffix** (`-<issue>-2`, `-3`, …),
+   OR whose normalised stem matches an open issue's
+   title (strip `war-`/area prefixes and hyphens, and normalise `our`→`or` spelling) — legacy claims
+   predate the issue-number template and would otherwise be invisible during rollout — for an open
+   issue with **no** open PR, as
+   `CLAIMED <repo>#<issue> (branch, no PR)`. This is the only pre-PR claim signal that exists: before
+   a PR there is no body to grep, so the issue number in the branch name is what makes the claim
+   discoverable. Keep it bounded — skip the call for repos with no assigned-and-PR-less issues.
 3. **Short-circuit dependency automation, then deepen only actionable candidates.** An org-search PR
    whose author is the exact `renovate[bot]` or `dependabot[bot]` identity is an automation-owned
    dependency PR. Emit only `AUTOMATION-OWNED (NO-ACTION)` from the cheap search row; do **not** call
@@ -300,12 +319,21 @@ nothing_on_fire: <true|false>   # true only if NO CI red on main AND no actionab
 ### Advance
 - <repo>: roadmap-ready → #<n> "<title>" (<label>)
 - <repo>: NO roadmap yet → strategy-review candidate
+- <repo> #<n> "<title>" — CLAIMED: assignee=devantler, claim-branch=<name>, no open PR
 ```
 
 Digest rules:
 - **Classify, don't decide.** Surface signals; the **orchestrator** selects the work and overlays its
   own native-memory cadence cursors (`last_worked`, `weekly`, docs/roadmap) — **you do not read
   memory**, only live GitHub.
+- **Emit a `CLAIMED` row only when BOTH a `devantler` assignment and a matching claim branch exist**
+  (and no open PR). Match `claude/*-<issue>`, a takeover branch (`claude/*-<issue>-2`, `-3`, …), or a
+  legacy normalised stem. An assignment to **anyone but `devantler`** is not a claim at all, and a
+  `devantler` assignment with **no** branch is not a live claim under the contract's *Claim
+  protocol*, so reporting either as one would let a bare assignee park an issue — exactly what "a bare
+  assignee does not reserve an issue" forbids. Report that case as an ordinary open issue (mention
+  `assignees=<n>` if useful), never as skip reason (e). The orchestrator times the ~2h lease from the
+  issue's newest `assigned` timeline event; an assignee is an **instance** claim, never the maintainer.
 - **Never assert ownership of a `devantler` PR.** Routine-own vs maintainer-interactive is the
   orchestrator's creation-record call, not yours — report CI state + `headRefName` + disclosure as DATA
   and tag it `OWNERSHIP-UNVERIFIED`, never `MERGE-READY`/"own". (Bot-trusted authors have no ambiguity.)

.claude/skills/portfolio-maintenance/SKILL.md

@@ -368,8 +368,23 @@ actionable open issue**, and any new non-trivial find is **filed as an issue fir
 backlog. Use the [`product-engineering`](../product-engineering/SKILL.md) skill; in order:
 7. **Resolve the oldest actionable open issue** *(the default advance action)* — pick the **oldest**
    open issue that's actually startable; skip one only if it's blocked, too under-specified to begin, or
-   it already has an open PR. A **bare assignee does *not* reserve** an issue (only an open PR does), so
-   if nobody has opened one you may pick it up regardless of who's assigned. If it **already has an
+   it already has an open PR. A **bare `devantler` assignee does *not* reserve** an issue
+   **indefinitely** — a `devantler` assignment plus a **pushed branch** is a live claim for ~2h
+   (contract *Claim protocol*), and with no branch, or once that lapses with no PR, you may pick it up
+   (timed from the issue's newest `devantler` `assigned` timeline event, never a branch commit date).
+   **Only the agent account's assignment is a claim, and only it expires:** an issue assigned to a
+   **human collaborator** (or `Copilot`) is someone else's work-in-progress — respect it and pick a
+   different issue, never take it over on this window. **Claim
+   before you build:** self-assign + push the branch **with the issue number in its name** the moment
+   you select — and if `devantler` is ALREADY assigned (a stale bare assignment from an abandoned run),
+   **remove then re-add**, since adding an existing assignee is a no-op that would leave your lease
+   carrying the old timestamp. **The push decides the race:** put a real commit on the claim branch
+   (never a bare base pointer), push without force, then confirm `git ls-remote` shows YOUR sha — two
+   instances derive the same branch name, so a rejected push or someone else's tip means you lost;
+   stand down rather than force over them. Check open PRs, remote
+   `claude/*` branches AND assignees by **issue number, never literal branch name**. A live claim
+   (assigned + branched, in-window, no PR) is skip reason **(e)** — the only one that expires by
+   itself. If it **already has an
    actionable trusted-author, non-draft PR**, drive *that* to merge instead of duplicating; leave
    automation-owned dependency PRs to repository automation, **draft** PRs for the maintainer, and
    **external** PRs static-review-only (trust gate). Otherwise ship it: tests +
@@ -435,7 +450,9 @@ For each selected product:
    .claude/scripts/submodule-init.sh <path>   # init at the pinned commit + repair + probe (fail-closed)
    ```
 
-   Then `git -C <path> worktree add .claude/worktrees/maint-<runid> -b claude/<area>-<desc>` and work
+   Then `git -C <path> worktree add .claude/worktrees/maint-<runid> -b claude/<area>-<desc>-<issue>`
+   (issue-less hotfixes and trivial obvious fixes keep plain `claude/<area>-<desc>` — they go straight
+   to a PR, so no claim window applies) and work
    **in that worktree**. A stray `core.worktree` makes the worktree resolve back into
    `.git/modules/<name>`, silently collapsing every parallel session into one physical tree — so on any
    submodule you did **not** initialise through the wrapper (a tree someone else populated), **probe

.claude/skills/product-engineering/SKILL.md

@@ -103,8 +103,10 @@ Issues are the unit of work (contract *Issue-driven*) — this is where new work
 1. **Pick the oldest *actionable* open issue — and "big" is not a reason to skip it.** Prefer the
    **oldest** startable issue. Skip an older one **only** if (a) it already has an open PR, (b) it is
    blocked on a **named, live-verified** external dependency you can cite, (c) it is too
-   under-specified to begin, or (d) a delivered experiment is waiting for its named future measurement
-   date. Once that date arrives, measuring it is actionable. **Size, difficulty, a `roadmap`/`enhancement`/
+   under-specified to begin, (d) a delivered experiment is waiting for its named future measurement
+   date — once that date arrives, measuring it is actionable — or (e) another instance holds a **live
+   claim** (assigned **and** branched, within ~2h, no PR yet; contract *Claim protocol*), the only
+   skip reason that expires on its own. **Size, difficulty, a `roadmap`/`enhancement`/
    `security`/`repo-assist`/`automation` label, or a "maintainer-hot" feeling are NOT skip reasons** —
    when the oldest issue is large, **decompose it into a small first child and ship that increment**
    (`Fixes #child`; add `Part of #experiment` when the parent stays open) so the big thing advances
@@ -118,8 +120,19 @@ Issues are the unit of work (contract *Issue-driven*) — this is where new work
    **ship the decision as a draft PR** (he steers there) — never a passive "awaiting-maintainer" note, never
    the end-of-run report (he rarely reads it), never an **`@devantler` mention** (no notification). Re-verify
    any "gated" against live state before trusting it (memory goes stale) and name the
-   blocker in the report. A **bare assignee does *not* reserve** it (only an open PR does), so if nobody's opened
-   a PR you may take it regardless of who's assigned; if an **actionable trusted-author** PR already
+   blocker in the report. A **bare `devantler` assignee does *not* reserve** it **indefinitely** — a
+   `devantler` assignment plus a **pushed branch** is a live claim for ~2h (contract *Claim protocol*);
+   with no branch, or once that lapses with no PR, you may take it (timed from the issue's newest
+   `devantler` `assigned` timeline event, never a branch commit date). **Only the agent account's
+   assignment is a claim, and only it expires:** an issue assigned to a **human collaborator** (or
+   `Copilot`) is someone else's work-in-progress — respect it and pick a different issue. **Claim the lane before you build** (self-assign + push
+   the branch **with the issue number in its name**, then harden), check open PRs / remote `claude/*`
+   branches / assignees by **issue number rather than literal branch name**, and on a lost race
+   **abandon** — then diff your build against the winner's and post only findings you have verified:
+   execute the probe against a **trusted/routine-owned** winner's branch, but for an
+   **external-contributor** winner it is **static review only** (the trust gate is not relaxed by a
+   lost race). If an
+   **actionable trusted-author** PR already
    exists, drive *that* one instead of duplicating — a non-draft to merge, a **routine-owned draft**
    to genuine readiness → self-promotion → merge (contract *Autonomy*); leave
    automation-owned dependency PRs to repository automation, other authors' drafts to their owners,

.claude/skills/products/monorepo/SKILL.md

@@ -19,7 +19,7 @@ improve discovery, comprehension, adoption, and portfolio positioning through bo
 material refreshes. Validate with the `docs` build before any PR.
 
 ## Repo-specific conventions
-- **Branch** `claude/<area>-<desc>`. Submodule pointers often show as modified — **expected, not yours**; detect real dirtiness with `git status --porcelain --ignore-submodules=all`; never stage pointer bumps; never `git submodule update --remote`.
+- **Branch** `claude/<area>-<desc>-<issue>` (the issue number makes a pre-PR claim matchable — see the contract's *Claim protocol*; issue-less hotfixes and trivial obvious fixes keep plain `claude/<area>-<desc>`). Submodule pointers often show as modified — **expected, not yours**; detect real dirtiness with `git status --porcelain --ignore-submodules=all`; never stage pointer bumps; never `git submodule update --remote`.
 - **Validate:** before any `docs/` PR, `[ -d docs/node_modules ] || (cd docs && npm ci)` then `cd docs && npm run build` (use `npm --prefix docs …`). Workflows → `actionlint`. Never edit auto-generated files / `*.lock.yml`.
 - **Agent/review files:** `AGENTS.md` is the **single canonical** instruction file — what humans, agents, and **Copilot code review** all read ([since 2026-06-18](https://github.blog/changelog/2026-06-18-copilot-code-review-agents-md-support-and-ui-improvements/)); there is no separate `.github/copilot-instructions.md` (retired portfolio-wide). Keep `AGENTS.md`, the `.claude/` skills/cards, and any path-scoped `.github/instructions/` files from drifting apart (see `product-engineering` §7).
 - **Labels** (apply only from this set): `automation`, `documentation`, `ci`, `dependencies`, `submodules`, `bug`, `enhancement`, `question`, `duplicate`, `wontfix`, `needs triage`, `needs investigation`, `performance`, `refactor`, `security`, `repo-assist`, `agentic-workflows`, `good first issue`, `help wanted`, `spam`, `blocked`, `next`.