changes.md

changes.md — ai

Follow Groq Qwen catalog replacement during generation (2026-08-04)

What changed

• scripts/generate-models.ts: moved the Groq Qwen reasoning-level compatibility override from the removed qwen/qwen3-32b catalog entry to the active multimodal qwen/qwen3.6-27b replacement.
• test/openai-completions-tool-choice.test.ts: moved the focused reasoning_effort request regression to the same generated model ID.
• src/providers/data/*.json: refreshed the reviewed live provider snapshots so strict release generation and checked-in model-ID types agree.

Why

• models.dev removed Qwen 3.2 after Groq's first-party model endpoint replaced it with Qwen 3.6. The release generator therefore removed the old typed ID, while the compatibility regression still referenced it, causing the 2026.8.4-2 release to fail during root TypeScript validation before any commit or tag was created.
• Groq documents Qwen 3.6 thinking mode as reasoning_effort: "default" and non-thinking mode as "none", so the existing compatibility mapping remains required on the replacement model.

Why this cannot be expressed as an extension

• Model inventory and provider-specific reasoning metadata are generated before the coding-agent extension runtime loads, and the typed built-in model IDs are consumed by the AI package itself.

Modified upstream files

• scripts/generate-models.ts
• test/openai-completions-tool-choice.test.ts
• src/providers/data/*.json

Expected merge conflict zones

• MEDIUM: the Groq branch of applyThinkingLevelMetadata() when upstream changes Qwen reasoning controls.
• MEDIUM: generated provider JSON whenever models.dev, OpenRouter, or OpenCode metadata changes again.

OpenAI compatibility resolver merge repair (2026-08-01)

What changed

• Restored getOpenAICompletionsCompat as the single compatibility resolver used and exported by the OpenAI Completions adapter.
• Ported upstream Z.AI max_tokens selection into the shared resolver.
• Preserved both automatically detected and explicitly configured toolSchemaFlavor values, with focused Moonshot coverage.

Why

• The merge restored an upstream-local adapter resolver beside the fork's shared browser-safe resolver. The duplicate omitted Moonshot schema flavor selection, so wire-bound tool schemas retained an unsupported root anyOf wrapper.
• Keeping one resolver prevents API and browser-safe compatibility decisions from diverging again.

Why this cannot be expressed externally

• Provider compatibility selection and final wire-payload schema normalization occur inside the provider adapter before extension hooks can safely compensate.

Expected merge conflict zones

• src/api/openai-completions.ts, src/utils/prompt-cache-ttl.ts, OpenAI compatibility types, and tool-schema/prompt-cache tests.

Require explicit opt-in before probing Ollama in stream tests (2026-07-31)

What changed

• test/live-api-gates.ts: owns Ollama discovery behind a gate that short-circuits before probing unless PI_ENABLE_LOCAL_LLM=1 or PI_ENABLE_LIVE_API_TESTS=1, using where ollama on Windows and which ollama elsewhere.
• test/live-api-gates.test.ts: mocks the command boundary and covers the default no-probe behavior, both explicit opt-in paths, and both platform-specific lookup commands.
• test/stream.test.ts: uses the gated Ollama discovery function instead of treating the absence of PI_NO_LOCAL_LLM as permission to probe and run the live suite.
• ../../test.sh: clears the two opt-in flags instead of exporting the retired PI_NO_LOCAL_LLM opt-out flag.

Why

• A normal npm test on a machine with Ollama installed could enter the live suite, pull gpt-oss:20b, start a local server, and load a large model without explicit consent. Default workspace tests must not probe or start local model infrastructure.

Why extension system couldn't handle this

• This behavior occurs during packages/ai Vitest discovery and setup, before the coding-agent extension surface is involved.

Modified upstream files

• test/live-api-gates.test.ts
• test/live-api-gates.ts
• test/stream.test.ts
• ../../test.sh

Expected merge conflict zones

• LOW: test/live-api-gates.ts and its tests may conflict if upstream changes live-test activation helpers.
• MEDIUM: the Ollama discovery and setup block in test/stream.test.ts may conflict if upstream changes how the local OpenAI-compatible test server is detected or started.
• LOW: ../../test.sh may conflict if upstream changes its isolated live-test environment variables.

Shared reasoning-tier capability detection (2026-07-30)

What changed

• Shared xhigh / max model-family constants are hoisted in models.ts, and map-less inference now uses one case-normalized, boundary-aware family matcher instead of unbounded substring checks.
• getSupportedThinkingLevels delegates extended-tier precedence to the exported supportsXhigh and supportsMax predicates rather than duplicating their map-omission and null-veto rules.

Why

• Capability inference for custom map-less models should reject unrelated ids and case-normalize legitimate aliases while keeping one precedence implementation. Generated catalog models retain their explicit maps, so behavior for real catalog models is intentionally unchanged.

Browser-safe prompt-cache TTL resolver (2026-07-28)

What changed

• src/utils/prompt-cache-ttl.ts (new): resolvePromptCacheTtlSeconds(model, env?) -> number | undefined plus PROMPT_CACHE_TTL_SHORT_SECONDS (300) / PROMPT_CACHE_TTL_LONG_SECONDS (3600). It mirrors EACH target API's own resolveCacheRetention precedence verbatim rather than inventing a unified one: anthropic-messages falls back to "long" and honors the bare process.env.PI_CACHE_RETENTION set-but-not-long branch; openai-completions / openai-responses / bedrock fall back to "short"; pi-messages returns undefined (backend default). Retention "none" and every API with unknown cache semantics (google, mistral, pi-messages, unknown) resolve to undefined.
• The pure compat predicates the resolver needs moved INTO that browser-safe utility and the API modules now import them from there and re-export for their existing consumers: getAnthropicCompat + isAnthropicApiBaseUrl (from src/api/anthropic-messages.ts), the resolved-compat getter (from src/api/openai-completions.ts), and supportsPromptCaching (from src/api/bedrock-converse-stream.ts).
• src/index.ts exports the new module from the browser-safe root surface.

Why

• senpi sizes how long its bash tool and omo's task tool may block in the foreground on the active model's prompt-cache lifetime. That lifetime is already decided per provider inside this package, so one shared resolver here is the single source of truth instead of a table duplicated in every consumer.

Why the compat predicates had to move rather than be imported

• The root surface is browser-safe. Importing supportsPromptCaching directly from src/api/bedrock-converse-stream.ts pulled the AWS SDK (@smithy/node-http-handler, agent-base, http-proxy-agent) into the browser bundle and broke npm run check:browser-smoke with 18 unresolved node:* errors. Moving the pure predicates into the utility and re-exporting from the API modules keeps one definition with no divergence risk, and keeps the root import graph free of Node-only dependencies.

Modified upstream files

• src/api/anthropic-messages.ts
• src/api/bedrock-converse-stream.ts
• src/api/openai-completions.ts
• src/index.ts

Expected merge conflict zones

• MEDIUM: each API module's resolveCacheRetention / compat-getter region, where the local definition became an import + re-export. If upstream edits those predicates, port the edit into src/utils/prompt-cache-ttl.ts so the resolver and the adapters stay in agreement.

Cover Claude Opus 5 in Anthropic adaptive-thinking metadata (2026-07-25)

What changed

• scripts/generate-models.ts: isAnthropicAdaptiveThinkingModel and isAnthropicTemperatureUnsupportedModel now match Opus 5 ids, and Opus 5 joins the native xhigh/max effort ladder alongside Opus 4.7/4.8 and Sonnet 5.
• src/api/anthropic-messages.ts: ADAPTIVE_THINKING_MODEL_MARKERS gained opus-4-8 and opus-5, and mapThinkingLevelToEffort maps Opus 5 xhigh/max to native efforts instead of collapsing them to high.
• src/providers/data/*.json: regenerated so every provider that serves Opus 5 (anthropic, github-copilot, opencode, vercel-ai-gateway, openrouter, amazon-bedrock) carries forceAdaptiveThinking, supportsTemperature: false, and the xhigh/max thinking level map.

Why

• Opus 5 is adaptive-thinking only. Sending it the legacy thinking: { type: "enabled", budget_tokens } payload is accepted by the API but produces a thinking block with no thinking text, so the model answers as if reasoning were disabled. Measured against the live API: legacy payload returned 0 thinking characters, while thinking: { type: "adaptive" } on the same prompt returned real thinking content.
• Without markers or catalog metadata, supportsAdaptiveThinking() fell through to the legacy branch for every provider whose Opus 5 entry had no compat, including proxy providers.
• Opus 5 also honors native xhigh and max effort, and they scale reasoning materially (measured on one prompt: high 849 thinking chars, xhigh 1123, max 3217). Mapping both down to high silently capped the model.

Why extension system couldn't handle this

• Adaptive-thinking detection and effort mapping happen while building the Anthropic Messages payload inside packages/ai, below any extension-visible surface, and the model catalog is generated build-time data.

Modified upstream files

• scripts/generate-models.ts
• src/api/anthropic-messages.ts
• src/providers/data/*.json

Expected merge conflict zones

• LOW: marker/predicate lists are append-only additions next to existing Opus/Sonnet entries.
• MEDIUM: regenerated provider data files conflict textually whenever upstream regenerates the same catalogs.

Carry non-enumerable context provenance through Responses conversion (2026-07-24)

What changed

• src/context-provenance.ts: added request-local, non-enumerable message/item provenance tokens.
• src/api/openai-responses-shared.ts: preserves those tokens while converting messages to Responses input items.
• src/types.ts and src/index.ts: expose the typed provenance helpers needed by coding-agent's replay boundary.
• src/utils/openai-codex-auth.ts: centralizes browser-safe ChatGPT account-ID extraction so normal Codex requests and remote compaction canonicalize the same wire tenant across bearer-token refreshes.

Why

• Provider-wire value equality cannot distinguish duplicated messages after filtering or injection. Replay slicing now requires the exact checkpoint-origin identities to survive the canonical context pipeline.

Why extension system couldn't handle this

• The provenance must survive conversion inside packages/ai, below extension-visible provider payloads.

Modified upstream files

• src/api/openai-responses-shared.ts
• src/index.ts
• src/types.ts
• src/utils/openai-codex-auth.ts

Expected merge conflict zones

• MEDIUM: Responses message conversion and shared public types.

Export canonical OpenAI Responses message conversion (2026-07-24)

What changed

• src/index.ts: exports convertResponsesMessages from the browser-safe root so coding-agent remote-compaction replay can locate checkpoint boundaries with the exact conversion semantics used by the real provider pipeline.

Why

• Counting checkpoint items with a separate converter could drop or duplicate the current prompt when errored/aborted assistants, orphaned tool results, empty users, or provider-native blocks changed item cardinality.

Why extension system couldn't handle this

• The boundary is defined by the provider wire conversion in packages/ai, below the coding-agent extension layer.

Modified upstream files

• src/index.ts

Expected merge conflict zones

• LOW: root exports if upstream reorganizes OpenAI Responses helpers.

Commit generated model catalog data for reproducible builds (2026-07-18)

What changed

• ../../.gitignore: removed the packages/ai/src/providers/data/ ignore rule so generated catalog JSON is committed, reviewed source, matching src/models.generated.ts.
• package.json: the ordinary build no longer runs generate-models; it compiles, restores the CLI executable bit, and copies the committed src/providers/data/ into dist. Networked regeneration stays explicit via the generate-models script, the root generate:models workflow, release tooling, and prepublishOnly.
• ../../scripts/build-all.test.mjs: the AI build config regression now asserts the ordinary build skips networked generation, keeps the committed-data copy step, retains the explicit generator workflow, and leaves catalog JSON unignored.
• README.md: model-generation guidance now describes src/providers/data/ as committed generated values.

Why

• The ordinary AI build fetched models.dev and provider APIs to regenerate ignored JSON catalog data, so a build could emit an unreviewed or different catalog and failed entirely offline. The committed .models.ts shards import the JSON at compile time, so the catalog must be committed generated source for the build to be reproducible.

Why extension system couldn't handle this

• Model inventory is generated before the coding-agent extension runtime is loaded, and package build scripts run before any extension hook exists.

Modified upstream files

• package.json
• README.md
• ../../.gitignore
• ../../scripts/build-all.test.mjs

Expected merge conflict zones

• LOW: AI package build scripts if upstream changes the compiler command or bin generation flow.

Preserve stable Kimi Coding model IDs during catalog generation (2026-07-17)

What changed

• scripts/generate-models.ts: added fallback metadata for kimi-for-coding and kimi-k2-thinking that live models.dev metadata can override but cannot silently remove.

Why

• Senpi's public model catalog and provider regressions still support these IDs. A transient upstream catalog omission caused release-time model regeneration to remove them and fail static validation.

Why extension system couldn't handle this

• Model inventory is generated before the coding-agent extension runtime is loaded.

Modified upstream files

• scripts/generate-models.ts

Expected merge conflict zones

• MEDIUM: the Kimi Coding generation block if upstream changes alias or fallback handling.

Preserve the generated CLI executable bit during builds (2026-07-13)

What changed

• package.json: ordinary AI builds now restore the executable bit on dist/cli.js, matching the existing publish-only safeguard.
• ../../scripts/build-all.test.mjs: added a regression assertion for the executable-bit build step.

Why

• TypeScript can rewrite dist/cli.js with mode 0644 when AI sources change. The release workflow runs an ordinary build before staging its release commit, so that rewrite could silently reverse the tracked executable mode.

Why extension system couldn't handle this

• This is package build and release behavior that runs before the coding-agent extension system is loaded.

Modified upstream files

• package.json
• ../../scripts/build-all.test.mjs

Expected merge conflict zones

• LOW: AI package build scripts if upstream changes the compiler command or bin generation flow.

Upstream model generation and test sync (2026-07-02)

What changed

• scripts/generate-models.ts: accepted upstream removal of stale model metadata fallbacks, including Copilot Sonnet 5 fallback cleanup.
• Updated focused AI regression tests covering Fireworks model routing, GitHub Copilot OAuth, delayed device-code polling, and OpenAI Codex stream request-body handling.

Why

• The fork should now rely on live/generated model metadata instead of keeping stale fallback entries, while preserving coverage for provider behavior touched by the upstream sync.

Why extension system couldn't handle this

• Model generation is a build-time catalog script, and the changed tests assert provider/library behavior outside the coding-agent extension runtime.

Modified upstream files

• scripts/generate-models.ts
• test/fireworks-models.test.ts
• test/github-copilot-oauth.test.ts
• test/oauth-device-code.test.ts
• test/openai-codex-stream.test.ts

Expected merge conflict zones

• MEDIUM: scripts/generate-models.ts if upstream changes provider metadata fetch or fallback handling again.
• LOW: focused provider tests if upstream changes request decoding, OAuth polling timing, or Fireworks model mappings.

Explicit live API opt-in for ambient credentials (2026-05-12)

What changed

• test/live-api-gates.ts: Added shared live-test gate helpers. Ambient provider keys and local model probes are ignored unless PI_ENABLE_LIVE_API_TESTS=1 or the provider-specific flag is set.
• test/oauth.ts: OAuth tokens from ~/.pi/agent/auth.json now resolve only for explicitly enabled live OAuth test providers.
• OpenRouter live suites in image, streaming, context-overflow, total-token, and thinking-disable tests now require PI_ENABLE_OPENROUTER_LIVE=1 in addition to a key.
• Local context-overflow suites now require PI_ENABLE_LOCAL_LLM=1, matching the existing fork policy that local model servers must be explicit opt-in.

Why

• npm test --workspaces --if-present must pass in developer environments that contain stale or unrelated credentials and local model daemons. An invalid ambient OPENROUTER_API_KEY, stale Anthropic OAuth token, and empty LM Studio server caused live suites to run and fail for reasons unrelated to the code under test.

Why extension system couldn't handle this

• These are packages/ai integration-test activation rules. Extension hooks are not involved in test discovery or live provider credential resolution.

Modified upstream files

• test/oauth.ts
• test/context-overflow.test.ts
• test/google-thinking-disable.test.ts
• test/image-tool-result.test.ts
• test/images.test.ts
• test/live-api-gates.test.ts
• test/live-api-gates.ts
• test/stream.test.ts
• test/total-tokens.test.ts

Expected merge conflict zones

• Upstream currently gates many live suites directly on credential presence. Rebase conflicts are likely in any live provider test that changes describe.skipIf(!process.env.<KEY>) conditions or OAuth token bootstrapping.

Live API test gating fixes (2026-04-09)

What changed

• test/tool-call-id-normalization.test.ts: the OpenRouter gpt-5.2-codex cases now pass reasoning: "high" so the live regression test still exercises tool-call ID normalization against the endpoint's current reasoning requirement.
• test/cross-provider-handoff.test.ts: the minimum-fixture assertion now exits early when fewer than two live fixtures are actually generated, so the suite skips gracefully in environments without enough working provider credentials.
• test/bedrock-utils.ts: Bedrock live tests now require both credentials and an explicit AWS region before enabling.
• test/context-overflow.test.ts: the OpenRouter Anthropic overflow case now accepts the provider's current managed-overflow behavior, and LM Studio overflow tests only auto-enable when PI_ENABLE_LOCAL_LLM=1.
• test/openrouter-cache-write-repro.test.ts: the narrow OpenRouter cache-write regression is now explicit opt-in via PI_ENABLE_OPENROUTER_CACHE_WRITE_REPRO=1.
• test/total-tokens.test.ts: the unstable OpenRouter deepseek/deepseek-chat total-token regression is now explicit opt-in via PI_ENABLE_OPENROUTER_DEEPSEEK_TOTAL_TOKENS=1.

Why

• OpenRouter now rejects openai/gpt-5.2-codex requests when reasoning is omitted or disabled, which broke the normalization regression for reasons unrelated to tool-call ID handling.
• The cross-provider handoff suite assumes multiple working live providers, but npm test --workspaces --if-present must pass even when the environment has no valid API keys (or only a partial/invalid live setup).
• Ambient Bedrock tokens without a region and auto-detected local model servers were causing unrelated live E2E suites to run in non-reproducible environments.
• A few narrow OpenRouter regressions are currently backend-specific and unstable in shared environments, so they now require explicit opt-in instead of making the default workspace test command flaky.

Why extension system couldn't handle this

These failures are in upstream packages/ai live integration tests, not in the coding-agent extension surface. Fixing them required targeted test-only updates in packages/ai/test/.

Modified upstream files

• test/tool-call-id-normalization.test.ts
• test/cross-provider-handoff.test.ts
• test/bedrock-utils.ts
• test/context-overflow.test.ts
• test/openrouter-cache-write-repro.test.ts
• test/total-tokens.test.ts

Expected merge conflict zones

• test/tool-call-id-normalization.test.ts: OpenRouter live test options may need re-merging if upstream changes the regression coverage or request options.
• test/cross-provider-handoff.test.ts: fixture-count gating may need re-merging if upstream restructures the live handoff bootstrap assertions.
• test/bedrock-utils.ts: credential gating may need re-merging if upstream changes how Bedrock test auth is detected.
• test/context-overflow.test.ts: OpenRouter overflow handling and local-LM opt-in logic may need re-merging if upstream revises those E2E expectations.
• test/openrouter-cache-write-repro.test.ts and test/total-tokens.test.ts: explicit opt-in guards may need re-merging if the affected OpenRouter backends become stable again.

TypeScript native tsc migration (2026-08-02)

What changed

• Replaced the tsgo compiler invocation with tsc in the build, build:offline, dev, dev:tsc, and prepublishOnly scripts; all flags and arguments remain unchanged.
• Bumped the root typescript pin from 6.0.3 to 7.0.2.
• Dropped the @typescript/native-preview toolchain dependency.
• Added @typescript/typescript6@6.0.2 (Microsoft's official TypeScript-6 API bridge) so scripts/check-ts-relative-imports.mjs keeps working: TypeScript 7 removed the classic programmatic JS API it imported.
• Added @typescript/native: npm:typescript@7.0.2 as a scoped alias. The typescript6 package publicly depends on @typescript/old (typescript 6.x), and npm hoists it; alphabetically @typescript/old beats typescript for the node_modules/.bin/tsc link, which would make every bare tsc invocation (root check and all package builds) silently run the TypeScript 6 compiler. The alias sorts after @typescript/old, so it deterministically wins the .bin/tsc link to the 7.0.2 native compiler. It is a bin-ownership pin, not an import target.

Why

• Adopt a stable-first toolchain policy: use the released typescript@7.0.2 native compiler for package builds and typechecks instead of the experimental tsgo dev build.
• The native-preview compiler has been retired upstream in favor of typescript@next.

Why this cannot be expressed externally

• Build scripts and devDependencies are package infrastructure, not runtime behavior; extensions cannot rewrite another package's manifest scripts or compiler selection.

Expected merge conflict zones

• package.json scripts blocks and devDependencies anywhere upstream still references tsgo or @typescript/native-preview.