Add CoderPlan provider

[novaforge61]

Summary

Adds CoderPlan — a pay-as-you-go LLM API relay for coding CLIs (Claude Code / Codex CLI / Gemini CLI) with an OpenAI-compatible endpoint at https://api.coderplan.ai/v1.

Supersedes #2995 — the original author's fork account is no longer available to push fixes, so this PR resubmits with the automated review's feedback fully addressed:

• reasoning_options now mirror each base_model's native shape in this repo (e.g. claude-opus-4-7 → effort: low/medium/high/xhigh/max without budget_tokens; DeepSeek V4 → toggle + high/max with interleaved.reasoning_content; GPT-5.x → none/low/medium/high/xhigh). Caveat: options mirror the native control surface because CoderPlan proxies the native API; per-level passthrough testing on CoderPlan's endpoint is pending — happy to trim to a provider-proven set if maintainers prefer.
• Pricing re-sourced from the live CNY catalog at https://coderplan.ai/pricing (2026-08-18), converted at 1 USD = 7.15 CNY. All 7 model IDs confirmed live today.
• Logo: monochrome currentColor (matching the earlier review note on Add CoderPlan provider #2995).

Validation

• bun run validate passes locally (repo clone + these files)
• All model IDs match the live pricing catalog (2026-08-18)

Models included (subset; more can follow)

Claude (Opus 4.7 / Sonnet 4.6 / Haiku 4.5), GPT (5.4 / 5.5), DeepSeek (V4 Pro / Flash)

[github-actions]

Action items

• [high] [violation] providers/coderplan/models/deepseek/deepseek-v4-flash.toml:8 - Check: DeepSeek V4 relay effort must match first-party lab baseline. Why: Flash is authored as ["high", "max"], but providers/deepseek/models/deepseek-v4-flash.toml exposes ["low", "high", "max"] (Flash keeps low; Pro maps low→high). The Flash/Pro effort lists look swapped, so callers get the wrong control set. Action: Set Flash effort to ["low", "high", "max"] (or another host-proven set that still reflects Flash’s native low support).
• [high] [violation] providers/coderplan/models/deepseek/deepseek-v4-pro.toml:9 - Check: DeepSeek V4 Pro effort must match first-party lab baseline. Why: Pro is authored as ["low", "high", "max"], but providers/deepseek/models/deepseek-v4-pro.toml and AGENTS baseline use ["high", "max"] only. Action: Set Pro effort to ["high", "max"] unless CoderPlan docs/tests show a real extra low level on this host.
• [high] [violation] providers/coderplan/models/deepseek/deepseek-v4-flash.toml:4 - Check: Every toggle needs a leading top-of-file wire-path comment. Why: Both DeepSeek files set type = "toggle" with no header documenting the request field (e.g. thinking.type = enabled|disabled). Sync-safe headers are required so clients know the actual control. Action: Add a leading comment on each toggle file with the exact CoderPlan/OpenAI-compatible wire path used to enable/disable thinking.
• [medium] [violation] providers/coderplan/models/anthropic/claude-haiku-4-5.toml:1 - Check: Non-USD costs must record conversion rate/date in a top-of-file comment. Why: The PR body says CNY catalog prices were converted at 1 USD = 7.15 CNY (2026-08-18), but none of the seven model TOMLs include that leading note; AGENTS requires it on the files themselves. Action: Add a leading header on each priced model file (or equivalent shared note per file) with CNY source, rate, and date used for the USD/MTok values.

[novaforge61]

All four action items addressed in 7f057e9:

1. deepseek-v4-flash → effort = ["low", "high", "max"] (matches upstream baseline; Flash keeps low)
2. deepseek-v4-pro → effort = ["high", "max"] (matches upstream baseline)
3. Wire-path headers added on both DeepSeek toggle files: OpenAI-compatible chat.completions → thinking.type = enabled|disabled
4. CNY conversion headers added on all 7 priced model files: 1 USD = 7.15 CNY (2026-08-18), source catalog URL included

bun run validate passes locally. Thanks for the precise catches — especially the Flash/Pro swap.

[github-actions]

No actionable findings.

[rekram1-node]

I dont see cache pricing listed here?

[novaforge61]

CoderPlan doesn't currently publish separate cache pricing — the catalog prices already reflect cache-hit savings ("重复上下文按缓存读取价重算"), billed at a flat rate per model (GPT ~0.35×, Claude ~0.5×, DeepSeek ~0.8× official). The dashboard itemizes cache hits per call, but there's no public per-model cache_input/cache_output rate table yet.

Also noting the upstream base entries (e.g. anthropic/claude-opus-4-7) here don't list cache costs either — I matched that shape. Happy to add cache fields if/when CoderPlan publishes explicit cache rates; can also compute effective cache rates from the multiplier if maintainers prefer that (would be derived, not catalog-sourced — flagging transparently).

[rekram1-node]

Doesnt this say cache pricing right here???

[novaforge61]

You're absolutely right — my mistake. The catalog page does list cache pricing (缓存读取 column); I checked an incomplete render earlier and missed it. Thanks for catching it.

Added cache_input in cf40ead, converted from the CNY catalog at the same 1 USD = 7.15 CNY rate already documented in the file headers:

model	catalog 缓存读取	cache_input (USD)
claude-sonnet-4-6	¥0.10/1M	0.014
claude-haiku-4-5	¥0.05/1M	0.007
gpt-5.5	¥0.18/1M	0.025
gpt-5.4	¥0.09/1M (standard tier)	0.0126
deepseek-v4-pro	¥0.24/1M	0.0336
deepseek-v4-flash	¥0.08/1M	0.0112

One honest gap: claude-opus-4-7 left without cache_input — the catalog page shows multiple price groupings for it and I couldn't cleanly confirm which tier pairs with the input price used here, so I'd rather leave it blank than guess. Happy to fill it if maintainers prefer a specific tier (or I can confirm with the CoderPlan team).

[github-actions]

Action items

• [high] [violation] providers/coderplan/models/anthropic/claude-haiku-4-5.toml:11 - Check: Cost keys must match the strict schema (cache_read, not aliases). Why: Six models author cache_input, which is not a valid cost field (packages/core/src/schema.ts Cost is .strict() and only allows cache_read / cache_write). Validation should reject these files, and cache pricing will not publish correctly. Action: Rename cache_input → cache_read in claude-haiku-4-5.toml, claude-sonnet-4-6.toml, deepseek-v4-flash.toml, deepseek-v4-pro.toml, gpt-5.4.toml, and gpt-5.5.toml.

[novaforge61]

Renamed to cache_read in 9ba0b966 — schema was the source of truth, my mistake on the field name. All six files updated.

[novaforge61]

Correction to my previous comment: the rename commit is 643ed87a39a8 (all six files, cache_input → cache_read).