In your existing $HERMES_HOME/config.yaml (default ~/.hermes/config.yaml), merge these keys into the existing sections. The example provider is named redwatts; use your existing named RedWatts connection. Add the example affinity key only if missing; preserve an existing native guard or custom affinity setting. Keep its URL, private key configuration, model controls, other profiles, permissions and tools unchanged.

agent:
  api_max_retries: 1
  auto_recovery_cycles: 0
providers:
  redwatts:
    context_length: 262144
    session_affinity_header: X-RedWatts-Turn-ID

Preserve custom timeout and native role controls. Do not add ignored named-provider request_timeout_seconds or stale_timeout_seconds declarations: this CLI resolves runtime provider custom and uses supported environment/default lookup. Read/write and decoded-chunk stale guards concern inactivity, not a whole-task cap; discarded SSE comments do not reset decoded progress. api_max_retries: 1 means one main-loop attempt, and auto_recovery_cycles: 0 disables recovery loops. These alone do not stop Hermes’s separate HTTP 5xx streaming-to-nonstreaming fallback. Preserve the supported affinity header: native session identity is stable through retries, tools and compression, and changes for a fresh session. A previously dispatched matching generation returns guarded HTTP 409 without another generation or cached-answer replay, even if transport mode changes. After failure inspect partial output and tool history before a deliberate fresh turn/session. Current serving capacity is 262,144 total / 65,536 output; preserve explicit model/reasoning and separate compression settings.

Tested with OpenCode 1.18.20. The helper declares the scoped plugin’s missing @ai-sdk/provider SDK dependency in .opencode/package.json; it preserves existing dependency versions/ranges and custom fields, and does not run npm install. Complete normal official OpenCode dependency initialization as consumer setup outside the customer’s candidate/code sandbox, not as a model Bash validation workaround. Missing dependencies can cause a scoped plugin to be skipped even while built-in plugins load. Check non-secret startup diagnostics and ensure the scoped plugin actually loads before relying on its turn, replay or compaction controls; the file’s presence or a helper “installed” message does not prove consumption. From your project, with Node.js available, set REDWATTS_API_KEY privately. Download the readable helper and inspect it before executing the separate second block.

REDWATTS_HELPER_DIR="$(mktemp -d)" &&
curl --fail --proto '=https' --output "$REDWATTS_HELPER_DIR/connect-opencode.mjs" https://redwatts.com/connect-opencode.mjs &&
cat "$REDWATTS_HELPER_DIR/connect-opencode.mjs"

After inspection, write the project provider fragment, scoped hook and absent SDK dependency declaration. Complete normal native dependency setup and verify the hook loads before relying on this scoped launch:

node "$REDWATTS_HELPER_DIR/connect-opencode.mjs" "$PWD" --model deepseek/deepseek-v4.1-flash &&
OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX=65536 OPENCODE_CONFIG="$PWD/.opencode/redwatts-provider.json" opencode --model redwatts/deepseek/deepseek-v4.1-flash

Existing credentials, permissions and unrelated configuration remain unchanged. The reviewed provider fragment uses current serving limits: 262,144 total context and 65,536 output tokens. Canonical API model metadata remains the physical 1,048,576 / 393,216 bounds, not an available-capacity promise. OpenCode 1.18.20 otherwise caps outgoing output at 32,000; the supported override above requests 65,536 and leaves at most 196,608 input tokens before framing. Do not request the physical maximum: it is currently incompatible.

For an existing helper-owned connection, inspect the current .opencode/redwatts-provider.json, scoped hook and consumer dependency manifest first. The helper will not silently replace recognized older generated catalogs, limits or hooks, including the prior 1.4.21 branded fragments and old 128,000-context/8,192-output connection, or switch an existing generated profile. After reviewing the new behavior, explicitly request migration:

node "$REDWATTS_HELPER_DIR/connect-opencode.mjs" "$PWD" --model deepseek/deepseek-v4.1-flash --migrate-owned-limits

Migration is limited to exact recognized generated settings and matching known old hooks, with an explicit migration flag and backup of the prior bytes. Unknown hooks, customized or differing files and shared/symlinked setup paths refuse rather than being overwritten. Existing dependency versions/ranges, credentials, permissions, other profiles and unrelated provider settings stay unchanged. Then complete normal consumer dependency setup and use the scoped launch above with the reviewed fragment.

The launch omits effort for default High; optional --variant low and --variant high preserve explicit effort. None/disabled and Max are currently incompatible, not auto-downgraded. Output includes reasoning, answer text and tool arguments. Keep streaming, compression and one active API request. UTF-8 JSON requests remain bounded at 8 MiB; messages and function-tool definitions count toward framing and total serving context. Preserve your permissions and validate complete tool calls before execution.

When actually loaded, the scoped hook supplies the native turn identity and marks a genuine guarded HTTP 409 nonretryable. It adds X-RedWatts-Output-Policy: fit-context only to the actual native compaction role; main and title requests remain strict. The compactor’s requested output maximum becomes a ceiling over all physical context room left after its whole input and framing are tokenized, without a separate summary quota, input trimming or forced non-thinking. This role behavior alone is not proof of recovery from every maximum-size history. A matching generation stays protected for at least 30 days and beyond, even if only streaming mode changed; the gateway does not replay the old answer. Changed tool history/new turns remain distinct. After failure, inspect local output/tool history before deliberately sending a new turn. Unfinished text may be discarded by the client; do not assume a visible partial answer was saved or an exit code means success.

Claude Code 2.1.220 uses the bounded Messages adapter with the selected DeepSeek V4.1 Flash profile, not a Claude model. API identity remains canonical DeepSeek. Set REDWATTS_API_KEY privately and launch from your project using the origin without /v1. These subshell settings do not rewrite global authentication/configuration or existing permissions, deny rules, hooks, skills or MCP tools. Check non-secret model/connection information in /status; managed settings may request unsupported controls. The launch uses current serving context 262,144 and standard output 65,536. Physical API metadata remains 1,048,576 / 393,216, not current capacity. Claude’s model/cost display is its own estimate, not verified RedWatts capability.

(
  export ANTHROPIC_BASE_URL='https://redwatts.com'
  unset ANTHROPIC_AUTH_TOKEN
  export ANTHROPIC_API_KEY="$REDWATTS_API_KEY"
  export ANTHROPIC_MODEL='deepseek/deepseek-v4.1-flash'
  export ANTHROPIC_DEFAULT_SONNET_MODEL="$ANTHROPIC_MODEL" ANTHROPIC_DEFAULT_OPUS_MODEL="$ANTHROPIC_MODEL"
  export ANTHROPIC_DEFAULT_HAIKU_MODEL="$ANTHROPIC_MODEL" CLAUDE_CODE_SUBAGENT_MODEL="$ANTHROPIC_MODEL"
  export CLAUDE_CODE_MAX_CONTEXT_TOKENS=262144 CLAUDE_CODE_MAX_OUTPUT_TOKENS=65536
  export CLAUDE_CODE_AUTO_COMPACT_WINDOW=196608
  export CLAUDE_STREAM_IDLE_TIMEOUT_MS=2147483647
  export CLAUDE_CODE_MAX_RETRIES=0
  export CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1
  export MAX_THINKING_TOKENS=0
  export CLAUDE_CODE_EXTRA_BODY='{"max_tokens":65536,"reasoning":{"exclude":true}}'
  export ANTHROPIC_CUSTOM_HEADERS='x-redwatts-output-policy: fit-context'
  claude --model "$ANTHROPIC_MODEL"
)

MAX_THINKING_TOKENS=0 suppresses Claude’s unsupported adaptive codec/budget request; it does not disable DeepSeek’s default High reasoning. The reasoning.exclude response filter hides returned thinking without changing effort or total output accounting; set it to false to display genuine reasoning. If managed configuration requests unsupported server-side context editing, add only "context_management":null to that extra-body object as an explicit opt-out. Actual schema requests are currently incompatible; capability announcements alone do not make them supported. Keep one active request and matching results for every local tool call.

Retry safety: this scoped launch disables automatic SDK retries and the separate streaming-to-nonstreaming fallback. A stream can fail after returning genuine output; an automatic fallback can change its output ceiling and start another generation. Inspect partial output and tool history before deliberately starting a new turn. A visible partial reply may not have been saved by the client, and a failed response is not a completed tool action.

Native output and compression: use the 65,536 allowance and 262,144 total serving context above; the physical 393,216 output / 1,048,576 context bounds are not currently available. The all-role fit-context header explicitly treats output as an at-most ceiling within the model context; it does not expand current serving capacity. It never trims input, forces non-thinking or creates a fixed summary quota. Preserve native main/title/compaction roles and inspect failed or partial output before retrying.

Native decoded-progress watchdog: CLAUDE_STREAM_IDLE_TIMEOUT_MS=2147483647 was actually parsed and scheduled by installed Claude Code 2.1.220. It is a resettable idle boundary of about 24.85 days, not an HTTP/network/task timeout, retry policy, token quota or iteration quota; genuine decoded progress can reset it. Controlled native streaming completed a 331,474 ms decoded-progress gap with this setting, not a 24.85-day duration test or an uptime guarantee. Do not set it to the Source policy’s 31-day ceiling (2678400000 milliseconds): the installed client accepted that number but aborted actual main/title streams within milliseconds. The native client cannot represent every possible silence interval within the full Source policy ceiling. This limitation does not shorten Source’s deadline or model budget, and an output-derived generation estimate alone does not guarantee paced delivery behavior.

Reasoning selection: native --effort low and --effort high select DeepSeek effort labels, not Anthropic ratios or independent thinking-token quotas. Omitted effort uses default High. Explicit reasoning intent takes precedence over reasoning_effort, then native output_config.effort; an exclude-only filter does not replace effort. minimal maps to Low and medium/xhigh to High. None/disabled and Max remain explicit intent but cannot currently be served; they return an actionable error rather than a downgrade.

Current capacity supports text, local function tools with automatic or no tool choice, Low/High/default reasoning and unsigned genuine thinking. Strict-schema/other response formats, required/named tool choice, images, server tools/search, adaptive thinking/budgets, server-side context editing and context-1m are not currently runnable. Token counting is not implemented; the gateway still checks context. A length-truncated response is incomplete: review it and never execute partial tool arguments.

These controls change copied examples only, not existing client configuration. Omitted effort uses High. Low and High are effort choices, not hard thinking quotas. None/disabled and Max are currently incompatible and are not silently replaced.

Default omits max_tokens; explicit allocation requests 65,536. Input, framing and reserved output must fit current 262,144-token serving context. Physical model metadata does not imply that 1M context or 393,216 output can currently be served. Explicit allocations are preserved, not silently clamped to available capacity.

Actionable errors: 400 unsupported_model_request means current capacity cannot support your controls; explicitly choose compatible controls or another service. 400 context_length_exceeded means verified context overflow requires deliberately shortening/compressing input or revising the allocation. Both are nonretryable. 503 source_capacity_unavailable is temporary; follow Retry-After: 3. Never automatically replay an accepted or partially returned generation; inspect local tool history first.

These are connection values, not a universal client configuration file. Add your key in the client’s private credential field; do not put it in a repository.

This request offers one read_file function to inspect at most 60 lines of one file before proposing a patch. It uses the selected DeepSeek V4.1 Flash profile. Set REDWATTS_API_KEY privately in your shell first. Running it uses your plan’s access.

The API returns tool calls; it does not read your files or run tools. A tool-capable client must collect a complete call, validate its JSON and permitted path, run the read-only function locally, then send the assistant’s tool_calls plus a role: "tool" result with the matching tool_call_id in the next request. Keep the same selected model. The curl example shows the first request only; review any proposed patch before applying it.

DeepSeek uses native tool calling: omit parallel_tool_calls from your client request. Explicit values are rejected, not silently ignored. A response may contain multiple tool calls; validate each complete call and return a matching result for every call. One active API request does not mean one tool call per response.