The bespoke Pi executor was 3,360 lines across 16 files. The deletion landed only after the generic bridge could carry the profile fields the executor had been responsible for and after the missing live observation surfaces were built. The result removed a second backend path, widened profile delivery, and kept a written list of what the generic wire still cannot do.
That sequence matters. Deleting a special case because it looks repetitive is cleanup theater. Deleting it after measuring its capabilities and building the missing ones is an architectural change a reader can inspect.
This is the second public post in Agent Runtime Infrastructure. The worker observability post covers the bridge change that made the deletion possible.
The old path had real differences
The first deletion proposal was rejected after a source review recorded four missing capabilities. The source review found four concrete differences between the bespoke executor and the generic bridge:
- The bridge had no
progress(), so a parent could not observe a running worker. - The bridge had no
traceSource(), so tool calls stayed in a result artifact instead of becoming spans. - The bridge ran Pi with
--print --mode jsonand closed stdin, so steering became process control rather than a live queue message. - A provider error could return HTTP 200 with an empty result and zero cost instead of throwing.
The review did not say the special case should live forever. It named the missing capabilities that had to exist before removal.
Commit 76db51d3 then added generic progress and trace input from the bridge stream.
The observability record shows how that code preserves unknown status and duration instead of manufacturing them.
The deletion
Commit fee88987 removed:
| Area | Change |
|---|---|
| Files changed | 16 |
| Lines added | 50 |
| Lines deleted | 3,360 |
| Bespoke executor | 880 lines deleted |
| Bespoke MCP adapter | 534 lines deleted |
| Executor tests | 685 lines deleted |
| MCP tests | 840 lines deleted |
| Public exports removed | 15 |
The generic bridge already routes a Pi model through the same createExecutor({ backend: 'bridge' }) path used by other workers.
After the deletion, a Pi worker is a bridge worker with a model name, not a second runtime family.
Four profile fields stopped being second-class
The commit body names the behavior that the generic path carries for Pi:
| Profile field | Generic bridge action |
|---|---|
reasoningEffort | lowered to Pi’s --thinking option |
tools | lowered to --exclude-tools |
resources | materialized into the run directory |
systemPrompt | passed as a native --system-prompt file |
The old path had been a narrow interpretation of a profile. The generic bridge holds Pi to the full profile contract instead. The profile contract post explains why that comparison must happen at the individual field level.
The losses stayed in the record
The deletion did not preserve every Pi-specific behavior. The public commits name the losses instead of hiding them behind the line-count reduction.
| Capability | After the deletion |
|---|---|
| mid-turn steering | a steer lands after the current --print --mode json run; interrupt is a process-tree kill |
per-worker args and env | no wire equivalent; configure those values on the bridge server |
| per-worker binary override | no wire equivalent |
WORKER_TRACE_PROPAGATION | the pi: true row is removed because the bridge has no worker environment channel |
| Pi MCP config file | the bridge owns MCP registration; agent-runtime no longer writes --mcp-config |
The earlier absence of progress() and traceSource() was addressed in 76db51d3 before the deletion.
The remaining differences are protocol limits, not silently dropped profile fields.
The merge cleanup at commit 43b11053 is part of the story.
A parallel branch edited the deleted file, a merge resurrected it, and the cleanup removed it again while keeping the generic observation code.
The line count is the easy part
The 3,360-line diff is useful because it tells us how much duplicate surface disappeared. It does not prove the generic path is equivalent. Equivalence comes from the capability list and the named losses.
That is also the runtime decision rule. Use one path when its profile contract, error behavior, observation fields, and control behavior match the worker’s needs. Keep a specialized path only for capabilities that the shared wire cannot express, and give that difference a test and a name.
The AI dev container guide describes the wider runtime boundary around commands, sessions, and review evidence. This post is the narrower version for backend code: remove the second path when the first can carry the same declared profile, then publish the losses that remain.
Release boundary
The deletion commit is public.
The npm package query checked on 2026-08-01 still returned @tangle-network/[email protected].
The changelog entry for 0.118.0 describes the deletion, but that release was not yet installable when this post was written.
Readers should inspect the commit when they need the source-level change and check npm when they need an installable version.
Reproduce the diff
The line counts and commit body are public.
git clone https://github.com/tangle-network/agent-runtime.git
cd agent-runtime
git show --stat --oneline fee88987ed5d495a765f2ea04bd0657ed9b71c8b
git show --format=fuller --no-patch fee88987ed5d495a765f2ea04bd0657ed9b71c8b
git show --stat --oneline 43b11053218a18c13dda31282b43477a90bc608aThe first command prints the 16-file, 3,360-line diffstat. The second prints the capability changes and named losses. The third shows the cleanup after the merge resurrected the deleted file.
FAQ
Why delete a backend-specific executor?
Because the generic bridge now carries the profile fields and observation surfaces that made the special path different. Keeping both would make every future fix land twice.
Did the deletion remove behavior?
Yes. Mid-turn steering degrades, per-worker argument and environment overrides have no wire equivalent, the Pi trace-propagation row is removed, and MCP setup moves to the bridge.
Did the generic path improve profile delivery?
Yes for the four fields named in the commit: reasoning effort, tools, resources, and system prompt. They now use the bridge’s native options or run-directory materialization instead of being silently ignored by the bespoke path.
Does 3,360 fewer lines prove the architecture is better?
No. The capability comparison and the named losses provide the evidence. The line count only shows how much duplicate code was removed.
Is the deletion in the published package?
The source commit is public, but the package query was 0.117.0 on 2026-08-01.
Check the package version before depending on the 0.118.0 changelog entry.
Sources checked 2026-08-01
The primary records are the generic observability change and the deletion.
The merge cleanup is recorded as commit 43b11053 in the reproduction command.
The installable package source is the agent-runtime repository.