feat(kiro): add Kiro provider (supersedes #2559)

[NachoFLizaur]

Supersedes #2559 by @jeffpardy (auto-closed as stale on 2026-09-04). This branch continues that work with the review fixes applied and keeps both authors' commits.

What

• Adds providers/kiro/ with provider.toml and 19 models: auto, Claude Sonnet 4 / 4.5 / 4.6 / 5, Claude Opus 4.5 / 4.6 / 4.7 / 4.8 / 5, Claude Haiku 4.5, GPT-5.6 Sol / Terra / Luna, DeepSeek 3.2, MiniMax M2.1 / M2.5, GLM-5, Qwen3 Coder Next.
• Every non-Kiro-native model uses base_model and is override-only; only auto is defined inline.
• reasoning_options are authored per model against Kiro's actual effort control (low-max for Claude reasoners; none-max for GPT-5.6; [] where Kiro exposes no caller control, e.g. GLM-5, MiniMax, DeepSeek).
• cost is 0/0 throughout: Kiro bills in credits per request, not per token, so there is no USD/MTok price to publish.
• Context/output overrides come from kiro-cli chat --list-models --format json (context_window_tokens) and Kiro docs.

Sources:

• Models: https://kiro.dev/docs/cli/models/
• Effort control: https://kiro.dev/docs/cli/chat/effort/
• ACP: https://kiro.dev/docs/cli/acp/
• Image input: https://kiro.dev/docs/cli/chat/images/
• Pricing (credits): https://kiro.dev/pricing/

Review findings

• Logo: intentionally not added; requesting a maintainer waiver (cc @rekram1-node). The official Kiro mark collapses to a solid square under the currentColor single-color renderer, and a modified mark would violate AWS branding requirements. The generic fallback is used deliberately rather than shipping a non-compliant or altered mark.
• qwen3-coder-next: was a full inline definition. Upstream now has models/alibaba/qwen3-coder-next.toml, so the Kiro entry is now base_model = "alibaba/qwen3-coder-next" with only cost 0/0 and limit.context = 256_000 (Kiro reports 256000 vs the lab's 262_144).
• GPT-5.6 limit.context = 272_000 with base_model_omit = ["limit.input"]: this is what Kiro serves ("Experimental preview of OpenAI GPT 5.6 ... with 272k context window" in kiro-cli chat --list-models, confirmed by the docs), not the lab's 1_050_000 / 922_000 input split, hence the omit.
• Output limits: aligned with Kiro's documented max_tokens table (https://kiro.dev/docs/models/effort/): Opus 4.6 / 4.7 and Sonnet 4.6 are capped at 64000 and carry an override; Opus 4.8 / Opus 5 / Sonnet 5 / GPT-5.6 are 128000 and inherit the lab value; GLM-5, MiniMax and Qwen have no published Kiro cap and inherit the lab value.
• DeepSeek 3.2: base_model is now deepseek/deepseek-v3.2 (Kiro serves "Experimental preview of DeepSeek V3.2"); only limit.context = 164_000 (from kiro-cli) and reasoning_options = [] are kept.

Consumer

kiro-acp-ai-provider and the opencode-kiro plugin (npm) resolve models through this provider entry.

Validation

• bun install --frozen-lockfile - no changes
• bun run validate - exit 0, no warnings; kiro resolves 19 models, all base_model targets exist
• bun test - 307 pass / 4 fail; the same 4 failures (open-weight weights links, DeepInfra modalities, LLM Gateway case variant, Cloudflare AI Gateway reasoning issue) reproduce on pristine anomalyco/dev and do not touch providers/kiro

Co-authored-by: Jeff Brown 375987+jeffpardy@users.noreply.github.com

[github-actions]

Action items

• [critical] [violation] providers/kiro/ - Check: New providers must ship a compliant logo.svg (currentColor, no fixed size/hardcoded colors; preferably square viewBox). Why: This PR adds provider kiro with no logo.svg. AGENTS.md and the review policy treat that as a merge blocker; the generic fallback is not a substitute. A maintainer waiver request in the PR body does not remove the requirement until an explicit waiver is granted. Action: Add a compliant providers/kiro/logo.svg, or get an explicit maintainer waiver recorded on the PR before merge.
• [medium] [possible mistake] providers/kiro/models/gpt-5.6-sol.toml:10 - Check: Provider limit.context must reflect this host’s real context window, not an unrelated pricing threshold. Why: All three GPT-5.6 entries set context = 272_000 while the lab bases use 1_050_000 context / 922_000 input; 272_000 matches the common GPT-5.6 cost-tier boundary, which is a frequent mix-up. Action: Confirm from kiro-cli chat --list-models / Kiro docs that Sol/Terra/Luna are actually capped at 272k on Kiro; if not, set the real context (and keep base_model_omit = ["limit.input"] only if input remains inapplicable).
• [medium] [possible mistake] providers/kiro/models/claude-opus-4.6.toml:14 - Check: Output limits should be consistent with Kiro’s reported caps across the catalog. Why: Opus 4.6 alone overrides limit.output = 64_000 (lab is 128k), while Opus 4.7/4.8/5 and Sonnet 5 keep inherited 128_000, and GPT-5.6 also inherits 128_000. If Kiro’s CLI/docs apply a shared lower output cap, several models are overstated. Action: Align limit.output (and any missing context overrides) with the same kiro-cli / docs source used for Opus 4.6, model by model.
• [medium] [possible mistake] providers/kiro/models/deepseek-3.2.toml:1 - Check: Non-lab hosts must base_model the correct underlying lab identity. Why: The file is id/name “DeepSeek 3.2” but points at deepseek/deepseek-chat (non-reasoning chat metadata, 1M context). Lab/peers usually map V3.2 to deepseek/deepseek-v3.2 (reasoning hybrid, different limits). Current overrides (name, description, 164k context) paper over a possible wrong base. Action: Verify which model Kiro serves; use deepseek/deepseek-v3.2 if it is V3.2 (and author reasoning_options if reasoning = true), or keep deepseek-chat only if Kiro’s non-thinking chat endpoint is intentional—and drop redundant fields that then match the base.
• [low] [possible mistake] providers/kiro/models/claude-haiku-4.5.toml:8 - Check: Modality overrides must be real host deltas. Why: Multiple Claude/GPT entries replace lab input = ["text", "image", "pdf"] with ["text", "image"] only. That is valid only if Kiro actually rejects PDF. Action: Confirm against Kiro image/docs/CLI capabilities; restore pdf (or omit the modalities block) where PDF is supported.

[NachoFLizaur]

GPT-5.6 context (272_000 + base_model_omit = ["limit.input"])

Verified against the runtime. kiro-cli chat --list-models --format json (kiro-cli 2.21.0) returns:

• gpt-5.6-sol: "description": "Experimental preview of OpenAI GPT 5.6 Sol with 272k context window", "context_window_tokens": 272000
• gpt-5.6-terra: "description": "Experimental preview of OpenAI GPT 5.6 Terra with 272k context window", "context_window_tokens": 272000
• gpt-5.6-luna: "description": "Experimental preview of OpenAI GPT 5.6 Luna with 272k context window", "context_window_tokens": 272000

The Kiro docs say the same: the quick-comparison table at https://kiro.dev/docs/models/ lists all three at 272K, and https://kiro.dev/docs/models/available-models/ states "All three GPT-5.6 tiers have a 272K context window."

272k is the real per-host context on Kiro, not a pricing tier. limit.input is omitted because Kiro reports one window and publishes no separate input cap; the lab's 922_000 input figure does not apply to a 272k host.

Output limits

Kiro publishes output caps, but only for some models. https://kiro.dev/docs/models/effort/ has a "max_tokens limits per model" table (Minimum / Maximum):

• Claude Opus 4.8: 1024 / 128000
• Claude Opus 4.7: 1024 / 64000
• Claude Opus 4.6: 1024 / 64000
• Claude Sonnet 5: 1024 / 128000
• Claude Sonnet 4.6: 1024 / 64000
• GPT-5.6 Terra / Sol / Luna: 1024 / 128000

kiro-cli chat --list-models reports no output field (only context_window_tokens), so the docs table is the only host source. Applied model by model and deferred the models that don't appear to their lab defaults.

• Auto: not in Kiro's table and has no lab base. It routes to Claude models, so 64_000 is kept as the lowest cap among its targets, now with a source comment.

DeepSeek base

Upstream models/deepseek/deepseek-v3.2.toml exists. Switched base_model to deepseek/deepseek-v3.2. Kept limit.context = 164_000 (Kiro's reported window vs. the lab's 128_000) and reasoning_options = [] (Kiro exposes no caller-side control for this model).

PDF

No change because Kiro rejects PDF input. Evidence:

• https://kiro.dev/docs/cli/chat/images/ lists the accepted attachment formats as "JPEG/JPG, PNG, GIF, and WebP" only.
• https://kiro.dev/docs/cli/acp/ documents the ACP agent capabilities as loadSession: true and promptCapabilities.image: true, nothing for documents.
• Live handshake with kiro-cli 2.21.0 (kiro-cli acp, JSON-RPC initialize) returns "promptCapabilities": {"image": true, "audio": false, "embeddedContext": false}.

[github-actions]

Action items

• [critical] [violation] providers/kiro/logo.svg - Check: New providers must ship a compliant logo.svg (currentColor, no fixed size/hardcoded colors; square viewBox preferred). Why: providers/kiro/ is a new provider with no logo; AGENTS.md treats this as a merge blocker. A maintainer waiver request in the PR body does not satisfy the requirement. Action: Add a compliant providers/kiro/logo.svg, or obtain an explicit maintainer exception recorded in-repo; do not ship without one of those.
• [high] [violation] providers/kiro/models/claude-opus-4.8.toml:1 - Check: Provider limit.output must reflect this host’s max output, not an inflated lab default. Why: Lab anthropic/claude-opus-4-8 has output = 128_000. Sibling Kiro entries (claude-opus-4.6, claude-opus-4.7) override to 64_000 with comments that Kiro caps Claude max_tokens at 64000, but claude-opus-4.8 never got that override after the base_model refactor, so it would publish 128k. Action: Add [limit] output = 64_000 (and the same source comment) if Kiro caps Opus 4.8 at 64k; otherwise document why this model differs.
• [high] [violation] providers/kiro/models/claude-opus-5.toml:1 - Check: Same host limit.output rule for Claude reasoners. Why: Lab anthropic/claude-opus-5 is output = 128_000; this file has no output override while the PR’s Kiro max_tokens alignment only fixed 4.6/4.7. Action: Set limit.output to Kiro’s documented cap for Opus 5 (likely 64_000), or cite why inheritance of 128k is correct on this host.
• [high] [violation] providers/kiro/models/claude-sonnet-5.toml:1 - Check: Same host limit.output rule for Claude reasoners. Why: Lab anthropic/claude-sonnet-5 is output = 128_000 with no Kiro override, so the catalog would overstate max tokens versus the 64k Claude caps applied elsewhere in this PR. Action: Add the Kiro limit.output override (or evidence that Sonnet 5 is uncapped / different on Kiro).
• [high] [possible mistake] providers/kiro/models/gpt-5.6-sol.toml:12 - Check: limit.context must be the host context window, not another product’s pricing breakpoint. Why: Lab GPT-5.6 context is 1_050_000 (input = 922_000); OpenAI first-party pricing uses a 272_000 tier threshold. Setting Kiro context = 272_000 on sol/terra/luna matches that tier size and is easy to mis-copy from pricing rather than context_window_tokens. Action: Confirm against kiro-cli chat --list-models --format json (or Kiro model docs) and keep 272k only if that is Kiro’s reported window; otherwise use the real context (and keep base_model_omit = ["limit.input"] only if still needed).
• [medium] [possible mistake] providers/kiro/models/qwen3-coder-next.toml:8 - Check: Output limits must match Kiro’s documented max_tokens and stay consistent with the PR claims. Why: Patch 15 removed limit.output = 64_000 here (and on glm/minimax) while the PR body still says qwen uses limit.output = 64_000. Resolved output now inherits lab 65_536 (glm/minimax inherit 131_072), which may overstate Kiro if a 64k host cap exists. Action: Reconcile each model’s limit.output with Kiro’s max_tokens table / CLI data, restore overrides where the host is lower, and fix the PR body so it matches the files.

[NachoFLizaur]

PR body fixed. The qwen3-coder-next / GLM-5 / MiniMax entries now inherit the lab limit.output (65_536 / 131_072 / 131_072) because Kiro publishes no cap for them; the body previously still said 64_000.

Logo: unchanged, see the explanation above.

Claude Opus 4.8 / Opus 5 / Sonnet 5 output limits: the reported violations are not correct. Kiro's max_tokens table (https://kiro.dev/docs/models/effort/) lists 128000 for these three models. Only Opus 4.6, Opus 4.7 and Sonnet 4.6 are capped at 64000.

GPT-5.6 context 272_000: already answered in the previous comment with the verbatim kiro-cli chat --list-models output (context_window_tokens: 272000 for Sol, Terra and Luna) and the Kiro docs.