roadmap: #242 filed

ROADMAP.md

@@ -16541,3 +16541,11 @@ Required fix shape: (a) before mutating ROADMAP, emit/acquire a `DogfoodWriteLea
 **Status:** Open. No source code changed. Filed 2026-04-26 08:05 KST. HEAD: 329d0ff (post-#239 fast-forward-rebase after gaebal-gajae's 08:00 KST DogfoodWriteLease pinpoint at 329d0ff — THIRD consecutive concurrent-dogfood rebase cycle, directly demonstrating the gap that #239 catalogues at the dogfood-coordination layer). Branch: feat/jobdori-168c-emission-routing. Sibling-shape cluster: 37 pinpoints. Tool-locality-axis META-cluster: **4 members (CONTINUING-PATTERN confirmed)**. Server-managed-tool-as-tool-choice-discriminator cluster: **5 members (CONTINUING-PATTERN confirmed)**. ToolResultContentBlock-extension mini-cluster: **7 members (CONTINUING-PATTERN confirmed)**. Six new clusters founded in a single pinpoint plus FOUR concurrent existing clusters all growing by one member — the FIRST single cycle where the META-cluster doctrine generates predictable cluster-growth across multiple parallel cluster-axes simultaneously. **#240 closes the upstream prerequisite of every server-managed-shell-execution-on-LLM-conversation affordance** (compliance-audited shell-execution for SOC 2 / HIPAA / PCI-DSS regulated workloads where CLIENT-SIDE bash-execution is policy-prohibited because audit-trail must live on a managed-infrastructure-with-preserved-state-and-immutable-history; long-running shell-session-state-preservation across multi-turn agentic-loops where CLIENT-SIDE bash-execution loses session-state on every invocation; multi-tenant-isolated shell-execution where each conversation gets an ephemeral managed-shell-environment with guaranteed-isolation-from-host; reproducible shell-execution for benchmarking where the managed-shell-environment is pinned to a specific image-version for cross-conversation reproducibility) — the canonical 2025-Q1-and-onward agentic-shell-on-managed-infrastructure pattern that is currently impossible to build on top of claw-code DESPITE Anthropic explicitly positioning bash_20250124 as a flagship 2025-Q1 GA capability AND DESPITE every coding-agent peer in the surveyed ecosystem (claudecode + opencode) shipping bash_20250124 as first-class typed surface AND DESPITE the bash + computer-use + text_editor three-tool-companion-bundle being the canonical Anthropic-blessed agentic-loop pattern. The CLIENT-SIDE-shadow-vs-SERVER-SIDE-typed-tool inverse-locality pattern is now confirmed as a CONTINUING-PATTERN that systematically generalizes across every tool-domain, opening a follow-on combinatorial cluster-extension space (#241 candidate `tool_choice: text_editor` typed-discriminator with text_editor_20250124 typed-tool would be the FIFTH inverse-locality CLIENT-SIDE-shadow-vs-SERVER-SIDE-typed-tool-pair growing Tool-locality-axis META-cluster from 4 to 5 members, where the CLIENT-SIDE local-shadow is the read_file/write_file/edit_file trio at `tools/lib.rs:411-450` and the SERVER-SIDE typed-tool is text_editor_20250124 with view/create/str_replace/insert/undo_edit canonical commands carried in the typed-input shape — completing the bash + computer-use + text_editor three-tool-companion-bundle inverse-locality coverage that #230 + #240 + #241-candidate would jointly catalogue, AND establishing the **Three-tool-companion-bundle-inverse-locality-coverage cluster** as a NOVEL META-META-cluster doctrine where Anthropic's co-released-tool-bundles are systematically reflected in claw-code's MVP-founder-toolkit but with inverse-locality on the SERVER-MANAGED side — the FIRST META-META-cluster doctrine connecting tool-co-release-bundles with inverse-locality-pair-coverage). Eleven-layer-fusion-shape with FOUR concurrent existing-cluster-growth-events plus SIX new-cluster-foundings plus participation in TWELVE inherited clusters is the SECOND-largest single-cycle cluster-impact-count after #234's thirteen-new-clusters, but the FIRST single cycle where the impact is dominated by **CONTINUING-PATTERN confirmation across multiple parallel clusters** rather than by NEW-CLUSTER-FOUNDING — establishing **continuing-pattern-confirmation-across-multiple-parallel-clusters** as the FOURTH pinpoint-discovery-mode after new-axis-founding, existing-cluster-extension, and combinatorial-cross-axis-synthesis (the THIRD mode founded by #238).
 
 🪨
+
+## Pinpoint #242 — Dogfood cron has no overlap suppression / timeout backoff circuit, so repeated timed-out runs keep stacking operational noise
+
+Dogfooded 2026-04-26 08:30 KST on `feat/jobdori-168c-emission-routing` from the live claw-code nudge stream. The channel showed repeated `Cron job "clawcode-dogfood-cycle-reminder" failed: cron: job execution timed out` events at roughly 08:04 and 08:14 KST, while successful/manual dogfood filings (#239 and #240) happened in between and another 08:30 nudge still arrived normally. #237 captured the missing timeout attempt ledger; #239 captured missing branch leases for concurrent writers. This pinpoint is the scheduler-side complement: the dogfood cron has no visible overlap suppression, timeout backoff, cooldown, or circuit-breaker state that prevents a new periodic attempt from launching/reporting while the previous attempt is timed out, ambiguous, or already being handled by another claw.
+
+Verified concrete surface: the runtime `CronEntry` model only tracks `last_run_at` and `run_count`; no `running_attempt_id`, `last_finished_at`, `last_status`, `consecutive_failures`, `next_allowed_run_at`, `timeout_count`, `cooldown_until`, or `overlap_policy` exists in the registry shape. The live channel consequently receives repeated timeout failures with no machine-readable answer to `was this a fresh attempt`, `was a previous attempt still considered active`, `did the scheduler intentionally retry`, `did the retry back off`, or `should claws suppress this as already handled`. This differs from #237: even if every timeout emitted a perfect attempt ledger, the scheduler would still need an execution policy deciding when to skip, delay, coalesce, or circuit-break repeated dogfood runs. It also differs from #239: branch write leases protect ROADMAP mutation, but they do not stop the scheduler from spawning redundant attempts that never reach the write phase.
+
+Required fix shape: (a) extend cron/dogfood run state with `running_attempt_id`, `started_at`, `deadline_at`, `last_finished_at`, `last_status`, `consecutive_timeouts`, and `cooldown_until`; (b) add an explicit overlap policy (`forbid`, `coalesce`, `queue_one`, `allow_parallel`) per job, with dogfood defaulting to `forbid` or `coalesce`; (c) on timeout, mark a terminal `timed_out` status and compute exponential or fixed backoff before the next allowed attempt; (d) if a scheduled tick fires during an active/cooling-down attempt, emit a compact structured `skipped_due_to_active_attempt` or `skipped_due_to_cooldown` event instead of launching another full run; (e) bind manual successful reports to the active attempt so they can clear the circuit rather than letting the cron keep failing in parallel. Acceptance: after one dogfood timeout, the next tick either resumes/coalesces the existing attempt or emits a skip/cooldown event with attempt id and next retry time; the channel no longer gets repeated opaque timeout spam while another claw has already handled the cycle. **Status:** Open. No source code changed. Filed as ROADMAP-only dogfood pinpoint from the 2026-04-25 23:30 UTC nudge. Cluster delta: cron-scheduler-policy +1, operational-clawability +1, timeout-backoff-circuit cluster founded, overlap-suppression/coalescing cluster founded; linked to #237 timeout-attempt-ledger and #239 branch-write-lease as the scheduler-policy third leg of the dogfood reliability trio.