@iam2r/pi-upstream-compat0.6.2

@iam2r/pi-upstream-compat

A Pi extension that fixes two upstream quirks Pi only handles for its own built-in providers — both decided by model id, so they work for any model served through a third-party gateway:

1. OpenCode session affinity — attach x-opencode-session / x-opencode-client to models whose id mentions opencode.
2. Thinking-mode replay — mark DeepSeek-family thinking models so Pi replays reasoning_content on assistant turns, which upstream requires.

The problems

1. OpenCode session headers

Pi only sends OpenCode's session headers when the selected model is served by its own built-in opencode / opencode-go providers, or when the request URL host is exactly opencode.ai (getSessionHeaders()).

Model ids also travel through third-party OpenAI-compatible gateways — for example builtin:opencode/deepseek-v4.1-flash-free served by a custom provider. Pi sees the gateway as the provider, so the gate can never match and no session header is sent at all. Upstream then either

• rejects the request outright — 400 MissingSessionID — or
• accepts it but derives a fresh session per request, destroying prompt-cache affinity (every turn pays full input-cache cost and latency).

2. Thinking-mode reasoning_content

DeepSeek-family thinking upstreams reject a replayed assistant turn that carries tool_calls but no reasoning_content:

400 The `reasoning_content` in the thinking mode must be passed back to the API.

Pi already knows how to satisfy that: a model with compat.requiresReasoningContentOnAssistantMessages gets reasoning_content: "" replayed on assistant messages. Pi sets that flag on its built-in opencode-go and deepseek models, and its other detection is URL-based (opencode.ai, api.deepseek.com) — never id-based. A model declared with a custom provider and a gateway base URL therefore never gets the flag, and every tool-using turn fails.

Install

pi install npm:@iam2r/pi-upstream-compat

For a single run without installing:

pi -e npm:@iam2r/pi-upstream-compat

Then restart Pi — extensions load at startup.

Renamed. This package was published as @iam2r/pi-opencode-session-header up to 0.3.3. Replace it with the name above; the old PI_OPENCODE_* environment variables keep working, but PI_UPSTREAM_COMPAT_* is the documented spelling now.

What it matches

Session headers

A request gets the session header when any of these hold:

Condition	Example
provider is opencode or opencode-go	Pi's built-in providers
baseUrl host is exactly opencode.ai	https://opencode.ai/inference/go/v1
model id contains opencode (case-insensitive)	builtin:opencode/deepseek-v4.1-flash-free

It appends x-opencode-session (the session id, see below) plus x-opencode-client: pi, and never overwrites a header another layer already set.

The id rule is the one that breaks in practice, because a gateway need not keep the upstream model name: deepseek-v4.1-flash-free, with no opencode anywhere in it, matches nothing above. Set PI_UPSTREAM_COMPAT_MODEL_MATCH to a comma-separated list of extra substrings (for example deepseek,v4.1) and those ids are recognised too.

Thinking-mode compat

An assistant turn is replayed with reasoning_content when the model id contains one of these fragments (case-insensitive):

Default fragments	Family
deepseek, kimi, glm, minimax, mimo	thinking upstreams that require the field back

Add more with PI_UPSTREAM_COMPAT_THINKING_MATCH (comma separated, appended to the defaults — setting one extra family cannot silently switch DeepSeek off), and skip individual ids with PI_UPSTREAM_COMPAT_THINKING_EXCLUDE. PI_UPSTREAM_COMPAT_NO_THINKING=1 disables this half of the extension entirely.

What the extension does for a matching model, right before Pi serializes the request:

• sets model.reasoning = true (unlocks Pi's thinking levels and is the gate Pi uses for the replay), unless the model already says reasoning: true
• sets compat.requiresReasoningContentOnAssistantMessages = true, merging into an existing compat object and keeping its other flags
• leaves models that Pi's own catalog already defines untouched — for those, Pi's curated metadata is the better answer (and guessing over it by id would turn, say, a non-thinking deepseek-chat into a reasoning model)
• never mutates a frozen model: it hands the implementation a shallow copy instead

reasoning: false cannot express an opt-out: Pi's provider composer fills that default for every models.json model that does not set it, so it is the normal state of the gateway models this extension exists for. Use PI_UPSTREAM_COMPAT_THINKING_EXCLUDE instead.

Where it applies

1. The session's model registry (session_start, model_select, before_agent_start, header event) — capabilities the UI reads live on these objects (AgentSession.supportsThinking() checks model.reasoning), so decorating the registry is what makes thinking levels and the model picker agree with the request builder. Decoration is scoped to the provider in use and skips every model Pi's catalog defines.
2. Before every request in both layers below — guarantees the model the implementation actually receives is marked, including models a gateway adds after startup, and regardless of what the registry holds.

The decision is per request, so it also covers models added to models.json after Pi started, and models whose ids a gateway rewrites.

pi --list-models prints the catalog without a session, so its thinking column still says no for gateway models: no session_start fires there, and Pi exposes no model-registry handle at extension-load time. Inside a session the model is marked before the first turn.

Marking a model as reasoning is not free — Pi offers thinking levels for it and may send reasoning parameters. The defaults stay narrow on purpose. If you add a family that is not actually a thinking model, expect upstream errors; drop it from PI_UPSTREAM_COMPAT_THINKING_MATCH or list the offending ids in PI_UPSTREAM_COMPAT_THINKING_EXCLUDE.

Two layers

Neither mechanism covers every request path on its own, so both are installed.

1. before_provider_headers — the official event. Emitted only from the coding-agent SDK's request-options builder, which covers the main agent loop but not callers that invoke ctx.modelRegistry.streamSimple(...) directly. Extensions do that for side requests (pi-subagents' pruned-fork summaries, for example), and those would reach the gateway with no session header.

2. An api-level wrapper — registerApiProvider() around the api implementation named by PI_UPSTREAM_COMPAT_API (default openai-completions). It injects for matching models, applies thinking compat, and delegates to the implementation it captured, so it sits below every caller and needs no knowledge of which provider serves the model.

Both layers are idempotent with each other: whichever sees the request first sets the headers or model flag, and the second one finds nothing left to do.

The registry decoration above is idempotent too — repeated session_starts and model switches only count models that still lack the flag.

Why layer 2 is re-asserted on every session

Two host behaviours shape this layer.

First, it cannot be a named provider registration:

// pi-coding-agent/dist/core/extensions/loader.js
registerProvider: (name, config, extensionPath) => {
	runtime.pendingProviderRegistrations.push({ name, config, extensionPath });
};

The host drains pendingProviderRegistrations once, at extension-load time (extensions/runner.js:bindCore()). A pi.registerProvider(...) call made after that — from a session_start handler, say — only appends to a queue nobody reads again, so it is silently dropped, even though the call returns normally. Wrapping the api implementation sidesteps that queue, needs no provider id, and covers every provider serving that api at once.

Second, Pi clears api overrides when a session reloads: AgentSession.reload() (core/agent-session.js), which serves /reload and session replacement, calls resetApiProviders() — every registration is dropped and only the built-ins are put back. So the wrapper is installed while the factory runs and re-asserted on every session_start. Re-installing is idempotent: the previous entry is dropped first, and the wrapper delegates to a captured base instead of to the registry, so installs cannot stack.

Why the session id is read lazily

The api wrapper runs outside any event context, so it needs a session id of its own. The cache is filled on session_start and refreshed on every before_provider_headers, and requests read the live session whenever that cache is empty — the session file name (<timestamp>_<id>.jsonl) is the last resort. A context captured before a reload counts as no answer instead of throwing.

This is hardening rather than a fix for a reproduced failure: at this extension's session_start the id is already available (measured), so the cache is normally populated before anything can ask. The fallback matters when it is not — an earlier handler that yields for a while delays every handler after it, since Pi awaits them in load order, and a session can be asked for a request while it starts.

The remaining boundary is inherent: a side request that arrives before any handler of this extension has run cannot be helped. The request carries no session information of its own, and Pi exposes no session id outside an event context (ExtensionAPI has no session getter). Requests after the first handler are covered.

Configuration

All optional, via environment variables. The pre-rename PI_OPENCODE_* names are still read as fallbacks.

Variable	Default	Meaning
PI_UPSTREAM_COMPAT_API	openai-completions	Api id(s) layer 2 wraps, comma separated. Every provider serving a model with one of these api ids is covered.
PI_UPSTREAM_COMPAT_SESSION_STYLE	raw	raw forwards Pi's session id verbatim. canonical derives a ses_… id instead.
PI_UPSTREAM_COMPAT_MODEL_MATCH	(unset)	Extra model-id substrings (comma separated) that also count as OpenCode, for gateways that rename them.
PI_UPSTREAM_COMPAT_NO_PROVIDER	(unset)	Set to 1 to disable layer 2 and use the event hook alone.
PI_UPSTREAM_COMPAT_THINKING_MATCH	(unset)	Extra model-id fragments that get the thinking-mode compat flag, appended to the defaults.
PI_UPSTREAM_COMPAT_THINKING_EXCLUDE	(unset)	Model-id fragments that keep their metadata as-is, e.g. a non-thinking model whose id mentions deepseek.
PI_UPSTREAM_COMPAT_NO_THINKING	(unset)	Set to 1 to leave model metadata untouched.
PI_UPSTREAM_COMPAT_DEBUG	(unset)	File path; appends one JSON line per injection, plus one per layer-2 install.

Set PI_UPSTREAM_COMPAT_API if your gateway speaks a different api (for example anthropic-messages or openai-responses). Unlike the event hook, layer 2 sees only the api ids it is told about.

raw vs canonical

• raw sends Pi's own session id (01a144cf-8eeb-72cd-aec1-3ee2789673ef). Correct when the gateway forwards the header unchanged to the upstream OpenCode deployment.
• canonical derives ses_… in the same descending-timestamp + base62 layout OpenCode uses, so a gateway that validates the shape of the header also accepts it. The derivation is deterministic — a stable Pi session always produces the same id — and idempotent, so an already-canonical value passes through.

Pick canonical if upstream rejects a raw id for being malformed; pick raw if upstream only requires a stable opaque key.

Verify

PI_UPSTREAM_COMPAT_DEBUG=/tmp/oc-headers.log pi -e npm:@iam2r/pi-upstream-compat

Each injected request appends a line naming the layer (event or api), the model, whether the thinking-mode compat applies to it, and the header values actually sent. A registry line appears when a session start or model switch marks models, with the models count. Other lines report the layer-2 install: api-installed (fresh wrapper, or a new base after the host replaced ours), api-reinstalled (re-asserted unchanged on session_start), api-missing, api-incomplete, api-failed.

A request is only ever modified when an injection line appears for it. Thinking-mode compat has no line of its own: check the thinking field of a request's line, then the compat flag on that model.

License

MIT