Several suites read and write ~/.pi/agent/{auth,qoder-models-cache}.json.
Vitest runs files in parallel, so one file could truncate a file another
was reading, and a run also clobbered the developer's own ~/.pi state.
Give each test file its own HOME.
Isolation exposed a latent coupling in the stream suite: with no resolved
identity on disk, streamQoder fetches /userinfo first and consumes the
single mocked Response, so the chat read failed on a locked stream. The
suite only passed because another file left credentials behind. Stub the
identity so the mock only ever serves the chat request.
Co-authored-by: Simon Shi <simonsmh@users.noreply.github.com>
pi-coding-agent 0.84 no longer exports AuthStorage, so the named import
failed to compile even though the call site already guarded for its
absence. Read it off the module namespace instead, keeping the auth-file
fallback when the export is missing.
Co-authored-by: Simon Shi <simonsmh@users.noreply.github.com>
CI was red at the lint step: an unused providerID parameter in
createQoderOAuth, an unused getCachedCredentials import in stream.ts,
plus unsorted imports and stale formatting in stream.ts and pat.ts.
Co-authored-by: Simon Shi <simonsmh@users.noreply.github.com>
Aliyun Model Studio (the upstream behind Qoder's CN catalog) documents
Max Output Length = 131072 for every model we expose, in both normal and
thinking modes (thinking chain alone goes to 262144):
qwen3.8-max/flash, qwen3.7-max/plus/flash — all 131072.
The previous 32768 default (copied from qodercli's conservative 32e3) used
only 1/4 of the supported output budget. That truncates reasoning chains in
thinking mode (where the chain counts against output) and risks cutting long
generations / compaction summaries. qodercli caps its own UI at 65536, still
below the documented ceiling.
- models.ts: introduce MAX_OUTPUT_TOKENS = 131072 as the single source of
truth; all 24 staticModels entries + the lazy-loaded CN entry now reference
it. Remove the dead max_output_tokens field from the CN placeholder entry.
- stream.ts: maxTokens defaults to MAX_OUTPUT_TOKENS; still capped lower by
options.maxTokens (e.g. pi compaction passes 40000).
Verified against the real CN endpoint: max_tokens=131072 returns 200 / stop
(accepted, not rejected); an xhigh reasoning run completes with
finish_reason=stop (not truncated). Field placement + reasoning_effort send
also confirmed working in the same request.
Qoder's /model/list catalog never returns max_output_tokens (the field does
not exist on any entry — only max_input_tokens is present), so the per-model
lookup `entry.max_output_tokens || 32768` was dead code that always fell back
to the 32768 literal.
- models.ts: remove max_output_tokens from QoderModelEntry; maxTokens is now a
flat 32768 (matches qodercli's 32e3 fallback).
- stream.ts: remove the now-unused maxOutputTokens variable and its always-true
branch; maxTokens = 32768, capped lower only by options.maxTokens (e.g. pi
compaction passes 40000). pi-ai's maxTokens setting still has an outlet.
- stream.ts fallback modelConfig: drop the dead max_output_tokens: 32768 field.
qodercli does conditionally send max_tokens (user setting → generation config
→ 32K fallback), but the field is not required by the upstream API. Keeping it
gives pi-ai's options.maxTokens a place to flow through.
The picker was fixed to show thinking levels, but the chosen level was never
forwarded: stream.ts only used options.reasoning to toggle the local thinking
parser, so every request let the model default-decide its reasoning effort.
Reverse-engineered @qoder-ai/qodercli (bundle/qodercli.js) to confirm the
request shape: the chat body's `parameters` object carries `reasoning_effort`
("none"|"low"|"medium"|"high"|"xhigh"|"max") and `enable_thinking` (bool)
alongside `max_tokens`. qodercli's Zlr() sets reasoning_effort = zN(req)
(reasoning.enabled===false ? "none" : reasoning.effort) and enable_thinking.
Implementation follows the pi-ai OpenAI provider pattern:
- clampThinkingLevel(model, options.reasoning) clamps the requested level to
what the model advertises via thinkingLevelMap (returns "off" when
unsupported or disabled);
- map the clamped level to the upstream effort name via thinkingLevelMap,
falling back to the pi level name (Qoder effort names already match pi);
- effort-based models (thinking_config.enabled.efforts present, e.g.
qwen3.8-max) get reasoning_effort + enable_thinking=true;
- toggle-only models (qwen3.7-max: enabled/disabled, no efforts) get
enable_thinking=true only — no reasoning_effort, since they accept no level;
- no level / clamped to off: enable_thinking=false to explicitly disable.
Qoder has no "minimal" effort; clampThinkingLevel handles stepping it down
based on the model's thinkingLevelMap (minimal is pinned to null for effort
models, so it clamps to the next supported level).
The catalog API returns each model's supported thinking efforts
(thinking_config.enabled.efforts, e.g. low/medium/xhigh for qwen3.8), but the
provider never populated Model.thinkingLevelMap, so pi's getSupportedThinkingLevels
hid xhigh/max (they require a non-undefined map entry) and the picker topped out
at "high" even for models that support xhigh.
Parse thinking_config at catalog fetch time and build the map:
- effort-based models (qwen3.8-max/flash, DeepSeek, GLM): each supported
effort maps to its same-named pi level; unsupported levels pin to null
(hidden). xhigh is now exposed. "off" is selectable when the catalog also
advertises a "disabled" option.
- toggle-only models (qwen3.7-max/plus: enabled/disabled, no efforts):
every pi level is exposed and maps to "enabled" so picking any level turns
thinking on; "off" selectable when "disabled" is advertised.
- models without thinking_config (qwen3.7-flash): no map, pi falls back to its
default level set.
Qoder effort keys are already pi level names (low/medium/high/xhigh/max), so
the mapping is by identity — no translation table.
This is display-layer only: it fixes which thinking levels the picker shows.
Sending the chosen effort upstream is a separate change that needs the exact
request field from a qoder-cli capture; the map values ("low"/"xhigh"/
"enabled"/"disabled") are placeholders that will be mapped to the request
shape once confirmed.
max_input_tokens (180K) was only used to seed the context window before
context_config options overrode it. Every live CN model exposes selectable
contexts (200K/400K/1M) whose smallest option already exceeds 180K, so the
seed only ever acted as a redundant floor and was never the final value.
- updateQoderModelsCache: derive contextWindow from the largest
context_config option directly; fall back to 200K when no context_config is
present (was 180K via max_input_tokens).
- staticCnModels: the Auto entry (the no-context fallback) now uses 200K too,
matching the dynamic fallback; its description no longer references the
stale 180K live-catalog note.
The CN model picker used to translate Qoder's internal keys (qfmodel,
qmodel_latest, ...) into hardcoded "friendly" ids (qwen3.6-flash, ...) via a
manual mapping table, with the display label hardcoded alongside. Both
sides rotted independently of upstream: when Qoder upgraded qfmodel to
Qwen3.8-Flash the label stayed "Qwen 3.6 Flash", and new keys (qmodel_38max,
q37fmodel) had no entry so their raw key leaked into the picker as the id.
The catalog API already returns every key with its display_name, so the
mapping table duplicates data the server provides and can only ever be stale.
Rework: use the upstream display_name (whitespace-stripped) directly as the
pi-visible Model.id, and read the request-time upstream key back from the
cached model config at send time. No key<->friendlyId table is needed.
- cosy.ts: remove qoderCNFriendlyModels, getQoderCNDirectModel,
getQoderCNFriendlyModelInfo, toQoderCNFriendlyModel, prettifyQoderCNModelName.
Add toQoderCNModelId(display) = display with whitespace stripped.
- models.ts: updateQoderModelsCache now sets CN model id =
toQoderCNModelId(display_name); the original key is still stored in configs
(keyed by both key and id) and read back via getCachedModelConfig. Simplify
the CN fallback in getCachedModelConfig (drop the stale reasoning set).
staticCnModels ids/names rewritten to the display_name form.
- stream.ts: the upstream key comes from modelConfig.key (the cached entry)
instead of a reverse-lookup table; falls back to model.id for the
international site, which already uses the key as id.
- index.ts: modelsForProvider no longer re-maps CN models through a friendly
transform — the cached/static ids are already the display name.
Trade-off: a renamed/upgraded model changes its id, so a previously saved
default model may no longer resolve after an upstream rename and the user
reselects. This is expected: users should learn about new models. This is
accepted.
Tests updated for the new contract; e2e against the live CN catalog confirms
id=Qwen3.8-Flash resolves to send key=qfmodel with no mapping table.
Two independent bugs made Qoder turns hang or fail outright.
1. The SSE read loop never terminated on the sentinel. `data: [DONE]`
broke only the inner line loop, so the outer loop kept awaiting
reader.read() until the socket closed. Qoder's gateway holds the body
open after the sentinel, so a reply that had already streamed in full
produced no `done` event: the agent sat there with the complete answer
buffered and no error, until something else timed the connection out.
The sentinel also arrives wrapped in an envelope (`body: "[DONE]"`),
which likewise only did `continue`, so both forms now end the loop.
The reader is cancelled once the reply is complete instead of being
left holding the connection.
2. Aborting a turn produced an invalid request. Assistant messages with
stopReason error/aborted are skipped, but the tool result that followed
was still emitted, leaving a `tool` message whose tool_call_id had no
matching tool_calls entry. Upstreams reject that with "tool must follow
a message with tool_calls" -- the same failure the null-content
workaround above it already guards against -- so after one cancelled
tool call, every later request in that session failed. Tool calls
belonging to a dropped turn are now recorded and their results dropped
with them.
Both fixes are covered by regression tests that fail on the current code:
the two stream tests time out at the 8s limit, and the transform tests
produce an orphaned tool message.
- Resolve the real Qoder identity (userID/email/name/machineID) from the
job token at request time. OMP keeps login credentials in its own
agent.db, not in ~/.pi/agent/auth.json, so a cache miss sent uid
"qoder-user" and Qoder CN rejected it with "Login expired" (105).
- Normalize the system prompt to a plain string. OMP supplies it as a
single-element content array which Qoder's MessagesInputDto#content
(String) rejects with 400 "Execution failed: set property" on qwen models.
- Drop the modifyModels hook. OMP (Bun) structuredClones the whole model
catalog before invoking it, and its bundled catalog is not cloneable,
which removed qoder from `omp models`.
- Make the build script cross-platform: use esbuild on PATH instead of a
Unix-style ./node_modules path that fails on Windows cmd.exe.
The "files" field in package.json only includes "dist" and "README.md",
so src/index.ts is not available in the published npm package.
This was introduced in commit 1b110a6 ("fix index") which changed
the path from the correct "./dist/index.js" to "./src/index.ts".
A tool call whose arguments never arrive — a no-argument tool, or a model that
sends id and name and then stops — was accumulated into toolCallsState but
never added to the message, because the content block was only created inside
`if (tc.function?.arguments)`.
The finalizer then checked `toolCallsState.length > 0` and set
stopReason "toolUse" on a message containing no tool call at all. pi's agent
loop continues only on tool calls, so it had nothing to execute and the turn
ended: mid-task, no tool run, no error, nothing in the transcript except the
assistant saying what it was about to do. Observed against a real session where
the model announced "let me consult the advisor" four times in a row and the
run then stopped dead; its final message on disk is stopReason "toolUse" with a
single text block and no toolCall.
Two changes:
- open the block as soon as the call is identifiable (id or name), and keep
its id/name in step if they arrive in a later delta, since the finalizer
only rewrites `arguments`;
- guard the stopReason on blocks that actually reached the message rather
than on the state array being non-empty, so a genuinely malformed stream
reports "stop" instead of an unactionable "toolUse".
Tests: no-argument call, identity arriving after the block is open, and a
malformed delta that must not claim toolUse.
A tool result can carry images. pi's `read` tool returns a text note plus an
`image` block for png/jpg/gif/webp/bmp, and screenshot tools do the same.
transformMessagesForQoder() built the `tool` message with getContentText(),
which maps every non-text block to "". The images were dropped silently: the
TUI rendered the picture, the model received only "Read image file [image/png]"
and answered that it could not see images — then fell back to scraping the DOM.
Nothing reported a problem, and the model catalogue advertises input
["text", "image"], so the capability check upstream of this had no reason to
warn either.
The OpenAI-shaped `tool` role has nowhere to put them: its content is a plain
string. So the images now follow as a separate user message, using the same
image_url part the user branch already builds, with a short label so the model
does not read a bare image as something the human just sent.
Tool results without images are unchanged and still produce exactly one
message.
The "files" field in package.json only includes "dist" and "README.md",
so src/index.ts is not available in the published npm package.
This was introduced in commit 1b110a6 ("fix index") which changed
the path from the correct "./dist/index.js" to "./src/index.ts".
Follow the fix in stream.ts: prompt_tokens (42) includes cached_tokens (5),
so input = 42 - 5 - 10 = 27. Use cache_write_tokens (not cacheable_tokens)
for cacheWrite, and keep cacheable_tokens:99 in the payload to prove it is
no longer mapped to cacheWrite.
pi-core computes promptTokens = input + cacheRead + cacheWrite (Anthropic
convention: input EXCLUDES cached/written tokens), but qoder follows OpenAI
semantics where prompt_tokens INCLUDES cached_tokens. pi-ai's own OpenAI
provider subtracts them; this provider didn't, which under-reported the
cache rate (45% vs the true ~83%) and tripped pi-core's miss detector on
every cached turn.
- input = max(0, prompt_tokens - cacheRead - cache_write_tokens)
- cacheWrite reads cache_write_tokens, not cacheable_tokens (a capacity
metric that is 0 even on first-turn writes, not a write count)
Qoder's backend sometimes routes a literal <thinking> opener into
reasoning_content while the matching closer lands in the content
stream, leaving tag fragments in the thinking block. Strip them during
streaming and add live repro scripts.
A fresh random session_id on every request meant the Qoder server could
never match a previous cache entry, so cacheRead was ~0% in all agentic
sessions. Ablation testing showed a stable session_id alone raises the
turn-2 cache hit rate from 0% to ~99.6% of prompt tokens (affinity
headers alone contribute nothing). The previously observed "406 Session
blocked" no longer reproduces with a per-pi-session stable id.
When assistant messages contain only tool_calls (no text/thinking), transform.ts
was setting content to null. Qoder's gateway drops these null-content assistant
messages, orphaning the following tool_result and causing upstream rejections:
- dmodel (OpenAI): 'tool must follow a message with tool_calls'
- ultimate (Anthropic): 'unexpected tool_use_id in tool_result blocks'
Analysis of 28 historical sessions showed 100% failure rate on dmodel and 75%
on ultimate for this pattern, while auto/GLM5.2 (GLM upstream) tolerated it.
Fix: inject a single-space placeholder when tool_calls are present but content
is empty, so the gateway preserves the assistant message and tool pairing stays
valid.
Also fixes pre-existing lint formatting in stream.ts and package.json.
Adds regression test for the full assistant+toolResult pairing scenario.
Verified via behavior test against Qoder API (ultimate model):
- system field with a directive: model ignores it (prompt_tokens=15, normal reply)
- same directive in messages[0] role:system: model obeys (prompt_tokens=97, marker reply)
The server's top-level `system` field in the request body is never fed to
the model, so pi's systemPrompt (language prefs, tool rules, agent instructions)
was silently dropped. Inject it as a leading role:system message instead, which
the server honors.
Three stability fixes for the streaming layer. Together they eliminate
the "agent silently stops mid-task" failure mode users hit during agentic
loops.
1. Surface upstream errors instead of swallowing them
The SSE parsing loop wrapped every line in try { ... } catch {}, which
silently dropped the throw raised when Qoder returned an error envelope
(e.g. 406 Session blocked, provider_error). The stream then completed
with empty content and stopReason "stop", so pi saw a clean stop instead
of an error and never retried or informed the user.
The catch now only skips genuinely malformed SSE lines (SyntaxError) and
re-throws everything else so the outer handler can set stopReason "error"
+ errorMessage. A QODER_DEBUG env var logs skipped lines for diagnosis.
2. Stop overwriting the real finish_reason with "stop"
At end-of-stream the code unconditionally did:
output.stopReason = toolCalls.length ? "toolUse" : "stop"
clobbering any meaningful finish_reason ("length", "content_filter", ...)
captured earlier. Now only toolUse is forced when tool calls are present;
otherwise the upstream finish_reason is preserved (defaults to "stop").
The done event reason cast now includes "length".
3. Give each request a fresh server-side session id
session_id was stableHash("qoder-session", userID, qoderModel) —
constant across all requests for a given user+model. Qoder enforces a
per-session lock and replies 406 Session blocked when a second request
arrives before the previous one fully drains server-side, which happens
routinely in tight agentic tool loops. Appending crypto.randomUUID()
makes every request use a fresh session. We send full conversation
history on each request, so losing server-side session continuity is
safe.
4. Capture upstream usage, responseId and responseModel
The provider reported all-zero usage and no responseId/responseModel.
Now the finish chunk usage (prompt/completion/total tokens, cache
read/write) is mapped into output.usage, and id/model into
responseId/responseModel, so pi sees real token accounting and can
correlate responses.
Add stream.test.ts with regression coverage for the 406 path, the
finish_reason preservation, the toolUse stop reason, and usage/response
metadata capture.