[AnthonyRonning]

The first full "catch up to upstream Pi" review of the pi-* crates, done as crates/UPSTREAM.md describes. It adopts 22 upstream commits, including the three that mid-task model switching needs. It does not switch models or touch Maple's per-task model lock; that is the next PR.

Range

• From Pi v1.0.4 (7c10bd433, 2026-10-05), the reference PR #1104 ported, to the v1.4.0 release (4a080252f, 2026-10-22). Agent core in Rust after Pi: pi-ai, pi-agent-core, pi-coding-agent #1104's parity review had also checked 1.1.0 (6fb2e78); its notes kept items it had already matched from being redone.
• Paths: packages/ai/src, packages/agent/src, packages/coding-agent/src/core (78 non-merge commits) and packages/coding-agent/src/utils (6 more). Every one of the 84 is listed below.
• Result: 22 adopted, 7 already matched, 55 skipped. Out of scope, as Agent core in Rust after Pi: pi-ai, pi-agent-core, pi-coding-agent #1104 decided: providers other than Chat Completions, the catalog, OAuth, telemetry, durations, codemode, Pi's MCP, packages and the terminal UI.
• Nothing on main past v1.4.0 was ported.

The three must-adopt changes

1. Another model's thinking is dropped (#11000, 911309429). In pi-ai/src/openai.rs convert_messages, a reply from another provider, API or model keeps its text and tool calls. Its reasoning is no longer sent as visible text. The same model still gets its own reasoning back in the field it came in.
2. Summaries leave thinking out (15e7bdc14). serialize_conversation no longer writes [Assistant thinking]. This applies to compaction, split-turn and branch summaries.
3. No fixed output limit on summary requests (#11036, 6fcb48670). Before, compaction was capped at 0.8 × reserve, the turn prefix at 0.5 × reserve and branch summaries at 4096 tokens. Summary requests now send max_tokens: None, so the provider fits the model's own limit to the request, as it does for turns. reserve_tokens only decides when to compact.

Commits

Each commit names its upstream commits and says whether it adopts an upstream fix or a feature. Each one builds and passes the pi crates' tests on its own.

1. pi-ai: drop another model's thinking (#11000)
2. pi-ai: repair \u without hex digits (4b994d5e5), and report malformed final tool-call arguments (#10944). This adds ToolCall::arguments_error and parse_tool_call_arguments; struct literals in maple-agent tests and store-bench gain the field.
3. pi-ai: whitespace-only tool output counts as none (4dd2b149d)
4. pi-ai: retry and overflow texts of 1.4.0, and any HTTP 413 as overflow (9 commits)
5. pi-coding-agent: summaries leave thinking out (15e7bdc14)
6. pi-coding-agent: no fixed summary output limit (#11036), and /compact instructions reach the split-turn summary (adfc14461)
7. pi-coding-agent: recovery omits a cut-off call's tool results with it (02726086d)
8. pi-coding-agent: before_agent_start for custom-message runs, and an abort while a run starts (#10739)
9. pi-coding-agent: keep empty quoted template arguments (9fe3f3716), and load a worktree's symlinked context file (f1b2e77f5)
10. pi-coding-agent: Windows shell lookup of 1.4.0 (61060ddf8, 28bea006b)
11. docs: the UPSTREAM.md reference revision and review log

Behavioural changes to review

These are the riskiest, roughly in order:

• Custom-message runs (send_custom_message, an extension's triggered send_message) now start as prompts do:
  • compaction first when needed, then before_agent_start, the run's prompt options, the next-turn messages and the handlers' messages;
  • Maple's MCP, computer-use and read_image handlers now run for them, so their prompt sections no longer drop out mid-run (Pi #10267: a prompt change and a cache break);
  • BeforeAgentStart now carries trigger: RunTrigger::{User { prompt, images }, Custom(CustomMessage)} in place of prompt/images. Maple's handlers read only options;
  • an abort requested while the run starts now aborts it at agent_start. Before, run_prompt cleared the flag and the run went ahead;
  • messages added meanwhile are flushed before the run's own messages, as they would be while idle.
• Summary requests are bigger. Without the fixed cap they can use the model's whole output limit, fitted to its window, and reasoning no longer starves the summary. A summary request is still one request, so history larger than the summarizing model's window still fails (see follow-ups).
• Malformed tool-call arguments are no longer cut back to a parsed prefix. The call fails validation with Invalid arguments for tool "…": <error>\n\nReceived arguments:\n<text>, and the model sees what it sent.
• Recovery after a cut-off reply now writes a context edit for each tool result of the omitted call, not only for the reply. Maple's timeline already hides omitted entries generically.

New differences from Pi

No new adaptation or Maple policy is added to the crates.

• Tool-call end events. A call closes when the next block starts, a difference Agent core in Rust after Pi: pi-ai, pi-agent-core, pi-coding-agent #1104 recorded. So the strict parse runs when the response finishes, as Pi's stream end does. A ToolCallEnd sent earlier, for a call that can still receive fragments, carries the arguments as they stream.
• Error text. arguments_error carries serde_json's message, where Pi's carries JSON.parse's.
• Status-code boundary. Rust's regex has no lookaround, so the status-code pattern spells Pi's boundary as (?:^|[^a-z0-9])…(?:[^a-z0-9]|$). Agent core in Rust after Pi: pi-ai, pi-agent-core, pi-coding-agent #1104 had recorded that ours matched whole numbers and Pi did not; that difference is now gone.
• Windows paths. The WSL launcher check normalizes separators, repeated separators and case, not every win32.normalize rule.
• Shared on purpose. Agent core in Rust after Pi: pi-ai, pi-agent-core, pi-coding-agent #1104 had kept "a custom message that starts a turn skips before_agent_start" because Pi did the same. Pi changed it, and so does this PR.

Storage adaptations revisited

Upstream's session manager and tool output did not change between these tags, so all three adaptations stay. bd3c9b823 is the one change they block. It activates extension tools registered after a branch recorded its tools, and needs every tool the branch ever named; a partially loaded SessionBase carries only the current snapshot. Maple sets its tools itself and never moves in the tree, so it is skipped and noted in UPSTREAM.md.

Validation

All runs were in apps/maple-agent's pinned Nix shell, with a private build directory. The build log shows Compiling pi-ai … (/Users/admin/workspaces/pi14/…).

• just ci — passes.
  • fmt, all four clippy runs (-D warnings) and the dependency checks pass. ci-extra passes: the -D warnings build of every target, and 352 headless app tests.
  • ci-test: 2,464 tests in 27 test binaries pass, 8 ignored. (A macOS-only failure in a fix(agent): bound shell execution and show quiet command progress #1378 shell test, from /tmp being a symlink, is fixed on master by PR #1398; this branch no longer carries its own fix.)
• Pre-commit hook — just ci-lint and just ci-test passed on every commit.
• Each commit on its own — tests for pi-ai, pi-agent-core and pi-coding-agent pass, and cargo check --all-targets -p maple-agent -p store-bench is clean with -D warnings.
• New tests — the pi crates go from 424 to 442 tests, all against the scripted provider or a scripted transport. They follow upstream's own tests where it has them:
  • cross-model and alternating-model thinking;
  • \u repair, strict final arguments, a cut-off final call, and validation through the agent loop;
  • whitespace-only output;
  • the new retry and 413 texts, and digits inside ids;
  • summaries without thinking or an output cap, and the split-turn focus with and without history;
  • recovery of a resumed cut-off call (Pi #11227);
  • custom-message runs keeping one system prompt, a custom message that joins the starting run, a message added during before_agent_start, and an abort while the run starts (Pi #10739);
  • empty quoted arguments, a symlinked worktree context file, and the WSL launcher in %SystemRoot%.

Real app, local stack

Maple Work was built from this branch (just build) and run against mde smoke pi14, signed in as the local Pro fixture. The task used Kimi K3, with Web off, in a scratch project under /tmp.

1. A reply with thinking, expanded:

   

2. A turn with a tool call, which read notes.txt and answered with its first line:

   

3. Manual /compact. At the default 20,000-token keep window a two-turn task reports "Nothing to compact yet", so the task first read two ~50 KB files. Compaction then succeeded on the real model. The app shows a notice, not the summary text:

   

4. After compaction, a tool-free question about the summarized turns gets a correct answer, so the summary reached the model:

   

Upgrade check: tasks from master reopened on this branch

The same store was used by a master build and then by this branch's build. Seven tasks were created on master: thinking (Kimi K3), read/shell/write/edit tools, /compact with a follow-up, a stopped reply, an image on a vision model, an image on GLM 5.3 through read_image, and eight turns. All seven reopened on this branch with their transcripts intact, and a follow-up in each that depended on earlier context answered correctly. A new task on this branch also passed, and so did a Claude Code background result delivered back into its task (the custom-message run path this PR changes). Neither build logged errors or warnings.

Follow-ups (not in this PR)

• Compaction on a smaller-window model (next project). A summary request is one request, here and in Pi.
  • History larger than the summarizing model's window still fails as an overflow of the summary request.
  • Without the fixed cap, output is the model's limit fitted to what is left of its window, which can be small on a tight window.
  • The model-switch work should decide where it summarizes, for example on the outgoing model before switching.
• Model switching. Another model's thinking is now dropped, and a model gets its own back when the conversation returns to it, matched by provider, API and model id.
  • check_compaction's overflow and cut-off-reply checks still apply only to the current model's own last reply, as in Pi. After a switch, the threshold check is what compacts before the next prompt.
  • Maple's per-task model lock is untouched.
• Noted on upstream main after 1.4.0, for the next review:
  • #11482: branch summaries think at the session's level. Agent core in Rust after Pi: pi-ai, pi-agent-core, pi-coding-agent #1104 noted Pi never read it, and ours sends none either.
  • #11494: LiteLLM's ContextWindowExceededError counts as overflow.
  • #11572: Zhipu (GLM) usage-limit errors are not retried.
  • f715a15c1: streamed tool-call arguments are parsed incrementally.
• Skipped here, worth revisiting.
  • 6f2f164ab/bd3c9b823 (restoring tools on resume and tree moves) would need defaultTools and the full branch's tool history in SessionBase.
  • Windows shell lookup is covered only by unit tests on macOS and by Windows CI.

Adopt/skip table

All 84 upstream commits in the reviewed paths, v1.0.4..v1.4.0.

Adopted (22)

Upstream	Area	Decision	Reason
911309429 (#11000)	pi-ai provider	Adopt (must)	Another model's thinking is dropped instead of sent as text; its text and tool calls stay.
4b994d5e5	pi-ai JSON repair	Adopt	\u without four hex digits (\usepackage) stays literal instead of dropping the field and everything after it.
87f9a3641 (#10944)	pi-ai tool calls	Adopt	A finished call's arguments are parsed strictly; malformed text leaves arguments empty with arguments_error, which validation reports to the model.
4dd2b149d	pi-ai provider	Adopt	Whitespace-only tool output is sent as (no tool output).
6e491cfa9	pi-ai retry	Adopt	Status codes match only as standalone numbers; ours used \b, now Pi's boundary.
7bd01fb92	pi-ai retry	Adopt	Retry text: "The upstream provider returned an error".
7efebff18	pi-ai retry	Adopt	Retry text: "Unable to forward upstream stream".
550f1f3ec	pi-ai retry	Adopt	Retry text: "The service is temporarily unavailable".
0fbdbd3f8	pi-ai retry	Adopt	Retry text: request to … failed, reason:.
698a2d23b	pi-ai retry	Adopt	Retry texts for more gateway upstream failures.
29057ece3	pi-ai retry	Adopt	Retry text: "stream disconnected before completion".
cedefc532	pi-ai retry	Adopt	ChatGPT Business workspace usage limit is never retried (shared account-limit list).
73f8de80b	pi-ai overflow	Adopt	Any HTTP 413 counts as context overflow, so the session compacts and retries.
15e7bdc14	compaction	Adopt (must)	Summaries and branch summaries leave thinking out of the serialized conversation.
6fcb48670 (#11036)	compaction	Adopt (must)	Summary requests set no output limit; the provider fits the model's own. reserve_tokens only decides when to compact.
adfc14461	compaction	Adopt	/compact instructions also reach the split-turn prefix summary.
02726086d	session	Adopt	Omitting a failed or cut-off attempt also omits the tool results that answer its calls.
a447ba11f (#10739)	session, extensions	Adopt	Custom-message runs go through compaction, before_agent_start (now with a RunTrigger), next-turn messages and run prompt options; an abort while a run starts stops it.
9fe3f3716	prompt templates	Adopt	Empty quoted arguments keep their place.
f1b2e77f5	resources	Adopt	A worktree's AGENTS.md that links to the main checkout's file loads.
61060ddf8	tools (Windows shell)	Adopt	PowerShell found in its install folders; the WSL launcher check follows %SystemRoot%.
28bea006b	tools (Windows shell)	Adopt	Microsoft Store shell aliases count as existing shells.

Already matched (7)

Upstream	Area	Decision	Reason
8c4e87389	tools (grep)	Already holds	Native search reads no ripgrep config.
63035734f	pi-ai provider	Already matches	Ours already applies reasoning before content in one chunk.
503c60552	session events	Already matches / skip	agent_settled.aborted ported in #1104; OSC 7501 is terminal UI.
27c7b6ff4	user bash	Already matches	Split colour codes already held back (#1104).
ce8972a0e	classifiers	Already matches / skip	Classifier images ported in #1104; the OpenAI Decisions classifier is not ported (#1104).
27075fe07	pi-ai estimate	Already matches	Ported in #1104 (3.5 characters per token).
8b5708dbb	pi-ai retry	Already matches	server_busy texts already retried (#1104).

Skipped (55)

Upstream	Area	Decision	Reason
4a080252f	release	Skip	Release commit.
7f225f25d	classifiers	Skip	Per-service image wire forms (TypeSafe, Cloudflare, OpenRouter); our System One follows Maple's server, which takes data URLs (#1104).
bd3c9b823	session tools	Skip	Needs every tool a branch ever named, which a partially loaded SessionBase does not carry; its resume half follows defaultTools (not ported). Maple sets tools itself and does not move in the tree.
04a93bad8	client identity	Skip	User-Agent and provider attribution headers; the host transport sets headers.
adc5fa9f8	client identity	Skip	As above.
e5fe9777b	catalog	Skip	Catalog not ported.
22dc6c26f	extension loader	Skip	TypeScript extension loading.
26293e74f (#11168)	HTML export	Skip	Not ported.
530015d93	catalog	Skip	Catalog not ported.
caa6967b6 (#11142)	Responses tools	Skip	Responses/Kimi in-place tool declarations are not ported; Chat Completions sends the current tools.
2ec55d808	telemetry	Skip	Telemetry spans and metrics are not ported.
8efd314fa (#11038)	docs	Skip	Generated models.json reference.
95364a54e (#10910)	defaults	Skip	Pi's default model per provider is not ported (#1104).
90d863f7c	defaults	Skip	As above.
5ab7fe74c	Anthropic	Skip	Provider not ported.
a3f265b1c	Kimi Code	Skip	Provider catalog/config.
6a1f799b6	Kimi Code	Skip	Revert of the above.
dc939913f	Anthropic	Skip	Provider not ported.
b06c1f2cf	Kimi Code	Skip	Provider catalog/config.
910149912 (#10938)	rendering	Skip	TUI/HTML read labels.
b39979799	pi-ai frames	Skip	Removes an API we never ported.
1692bba63	Google	Skip	Provider not ported.
d7f7e0dfe (#10913)	model resolver	Skip	CLI model references; models come from the host registry (#1104).
60ac3bcc3 (#10930)	Responses	Skip	GPT-6 mid-conversation effort on the Responses API, not ported.
195120a91	OpenCode	Skip	Provider catalog/config.
6f2f164ab (#10887)	session tools	Skip	Resume restores tools only when they come from defaultTools/--tools, not ported; hosts choose tools at start (#1104).
34a8bb38b	Anthropic	Skip	Provider not ported.
cea257439	share	Skip	/share and HTML export are not ported.
d4d53612b	Anthropic	Skip	Provider not ported.
e6fb2581d	auth storage	Skip	auth.json locking; credentials come from the host.
014873d7d	codemode	Skip	Codemode not ported (#1104).
6a51a39d4	package settings	Skip	!/+/- resource entries are not ported (#1104).
ee409a85f (#10672)	OpenRouter catalog	Skip	Catalog not ported.
24382bedd	codemode/MCP	Skip	Pi's MCP and codemode are not ported (#1104).
0f4b1465a	Codex	Skip	Provider not ported.
6b07b4e57	Mistral	Skip	Provider not ported.
7f9e1198f (#9880)	settings schema	Skip	Publishes JSON schemas and moves types; defaults unchanged.
f76c1db66	catalog	Skip	Model catalog not ported.
7fb59f995	Mistral	Skip	Mistral's own finish-reason text; our Chat Completions reports its own.
8d8ae2fc2	Anthropic OAuth	Skip	OAuth not ported.
6b5854454	classifiers	Skip	OpenAI classifier catalog typing; not ported.
fe11328b0	faux provider	Skip	Pi's faux cache-usage estimate; usage numbers unchanged and ours estimates its own way (#1104).
f284a2460	durations	Skip	Timing internals and streamProxy, not ported.
3ba22ce17	pi-ai types	Skip	TypeScript stream typing.
43d376399	Radius	Skip	Catalog not ported.
36a686ee8	durations	Skip	Response and tool durations are not ported (#1104).
2989eb581 (#10142)	Bedrock	Skip	Provider not ported.
18336987a (#10557)	rendering	Skip	TUI outputPad.
ddaa0a034	CLI tools	Skip	--tools +name/-name is CLI and defaultTools (not ported).
0cf65d2bf (#10429)	Codex	Skip	Provider not ported.
9ad083102 (#10433)	OpenAI login	Skip	OAuth not ported.
3e477eab5	clipboard	Skip	Terminal UI.
592fb57b7	clipboard	Skip	Terminal UI.
b30a6dd77	image resize	Skip	Node worker channel; ours resizes with the image crate.
f8ab362cc	clipboard	Skip	Terminal UI.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[AnthonyRonning]

Review note, recorded here and deliberately not fixed:

P3: a broken shell link can be mistaken for a Microsoft Store alias on Windows. is_store_alias accepts any path where following the link fails but reading the link itself succeeds. An ordinary dangling symlink passes that too. A stale bash.exe or pwsh.exe link early on PATH could then be picked over a working shell later on PATH. This needs an uncommon setup and hasn't been reproduced on Windows.

We're ignoring it on purpose. Pi 1.4.0 behaves the same way: it checks a shell with accessSync(F_OK), which on Windows reads the link's own attributes, and it takes only the first where result. The Pi crates stay with upstream unless something is a blocker, and this one isn't. It can go if upstream changes it.