t3022: Cap concurrent opus-4-6 dispatches at N per Anthropic account to prevent rate-limit cascade

[marcusquinn]

Problem

Single Anthropic account sustains many concurrent sonnet workers but only ~3-4 concurrent opus before triggering 429 rate limits. Currently no per-model concurrency cap exists in dispatch logic. When fill_floor + cascade-elevation both push opus-tier dispatches, all simultaneous opus workers hit 429 within seconds and become 20-min zombies (see GH#21570 follow-up).

Evidence: 3 opus-4-6 workers killed at the same minute with rate_limit (headless-runtime-metrics.jsonl ts=1777397345-1777397359). Sonnet workers in the same pool succeeded.

How

EDIT: .agents/scripts/pulse-dispatch-engine.sh — function _dff_compute_max_parallel (~line 712-727). Add a per-model concurrency check:

# Count in-flight workers by model from active worker process list
local opus_inflight
opus_inflight=$(pgrep -f 'opencode.*-m anthropic/claude-opus' | wc -l | tr -d ' ')
local opus_cap=${AIDEVOPS_OPUS_CONCURRENCY_CAP:-4}
if (( opus_inflight >= opus_cap )); then
    # Skip opus-tier candidates this cycle; sonnet candidates still dispatch
    return $opus_cap
fi

EDIT: .agents/scripts/pulse-dispatch-core.sh near line 1119 — when a candidate's resolved model is opus AND opus_inflight >= cap, defer with reason="opus_concurrency_cap" instead of dispatching.

NEW: per-model cap config in .agents/configs/dispatch-model-caps.conf:

opus-4-6=4
opus-4-7=4
sonnet-4-6=24

Reference pattern

shared-gh-wrappers.sh rate-limit threshold logic is conceptually similar: probe budget before action, defer when below threshold.

Verification

# Spawn 6 simulated opus workers (sleep loops with matching cmdline)
for i in 1 2 3 4 5 6; do (exec -a 'opencode -m anthropic/claude-opus-4-6' sleep 600 &); done
# Run pulse cycle
pulse-wrapper.sh --canary
# Log should show 'opus_concurrency_cap' deferrals for issues that would have used opus

Acceptance

• Concurrent opus workers never exceed the cap (default 4)
• Sonnet dispatch unaffected
• Cap is overridable via AIDEVOPS_OPUS_CONCURRENCY_CAP env
• Deferred candidates retry next cycle (not NMR'd)

Filed from interactive root-cause session (2026-04-28). Composes with the early-429-detection fix.

---

aidevops.sh v3.13.7 plugin for OpenCode v1.14.29 with claude-opus-4-7 spent 1h 12m and 19,339 tokens on this with the user in an interactive session.

[marcusquinn]

Large File Simplification Gate

This issue references file(s) exceeding 2000 lines: .agents/scripts/pulse-dispatch-engine.sh (2014 lines).

Workers dispatched against large files spend most of their context budget reading the file, leaving insufficient capacity for implementation.

Simplification issues: #21334 (existing)

Status: Held from dispatch until simplification completes. The needs-simplification label will be removed automatically when the target file(s) are below threshold.

Automated by _issue_targets_large_files() in pulse-wrapper.sh

---

aidevops.sh v3.13.8 automated scan.

[alex-solovyev]

CLAIM_RELEASED reason=worker_failed runner=alexey ts=2026-04-29T03:57:11Z

[alex-solovyev]

DISPATCH_CLAIM: Worker PID 2579657 dispatched at 2026-04-29T03:57:52Z by alex-solovyev (pulse-sweep, manual dispatch due to GraphQL exhaustion). Model: anthropic/claude-sonnet-4-6. GraphQL reset expected ~04:11 UTC.