Skip to content

Add fx chats over ACP with reasoning effort and full tool results - #141

Open
molefrog wants to merge 38 commits into
mainfrom
claude/sweet-faraday-fy7od4
Open

Add fx chats over ACP with reasoning effort and full tool results#141
molefrog wants to merge 38 commits into
mainfrom
claude/sweet-faraday-fy7od4

Conversation

@molefrog

@molefrog molefrog commented Sep 11, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner

moi now offers fx workspaces (fx 0.0.9 or later, verified on 0.0.11) with streaming chats, images, model selection and reasoning effort over ACP. ACP clips every fx tool result to 200 bytes, so moi restores command output, exit codes and fx's own line diffs from fx session --id <id> --json, unwraps fx's model-facing envelopes, and turns fx's operational messages (provider failures, restarts, stops) into notices instead of replies.

fx 0.0.8                      → unavailable: Run fx upgrade
Run command  $ seq 1 400      → 400 lines, 3 broadcasts, "Exit code 0 · 0.1 s"
Edit  src/greet.ts            → fx's diff with context, "+2 −2"
Run subagent  count files…    → markdown report (was an escaped JSON string)
Model menu                    → "claude-opus-5.5" under Anthropic (was 256 raw ids)
cold reload, tab left open    → one session_reloaded refetch, no scrambled turns

Review closely:

  • server/harness/acp/session.ts: replayed turns are no longer broadcast; a cold load sends session_reloaded and clients refetch. fx chats released when idle or by an env change keep their live transcript and reuse it only when the replay has the same number of user turns. Hermes's env and stop hooks now act only on Hermes chats; they used to discard fx's.
  • server/harness/fx/history.ts: fx saves at most 4,096 bytes of a result, so a long shell envelope arrives cut mid-JSON. Saved shell output only fills rows with no output of their own; a streamed row is never replaced.
  • server/harness/acp/discovery.ts: fx's chat list uses one warm process per workspace (10-minute idle release) instead of spawning per refresh.

Verification:

  • bun test: 1,923 pass, 0 fail; typecheck, lint (existing warnings only) and format pass. PROBE_EFFORT=high bun scripts/probe-fx-acp.ts --fake: replay matches the live transcript.
  • fx 0.0.11 on the real Gateway with Kimi K3, GLM 5.3, Opus 5.5 and GPT-6 Sol: thinking probe (High confirmed, all answered 20/51); a browser matrix that found 22 defects (19 in moi, fixed; 3 in fx, documented); a 14-check re-run per model, whose cold-reload and listing findings were fixed and confirmed by a spot check after a restart (7/7 cold checks, a 2,000-line output kept intact, 2 listing processes in 9 minutes).
  • Evidence and upstream limits: fx notes, ACP addendum.

🤖 Generated with Claude Code

https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy

@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Sep 11, 2026 •
edited
Loading

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-09-11T12:03:58.917887Z c966218 PR opened
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@molefrog molefrog changed the title Verify fx session/load replay fix via ACP layer probe Record fx structured session/load replay and add an fx ACP probe Sep 11, 2026
chatgpt-codex-connector[bot]
chatgpt-codex-connector Bot reviewed Sep 11, 2026
View reviewed changes

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c96621852f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment thread scripts/probe-fx-acp.ts
Comment thread scripts/probe-fx-acp.ts
Comment thread scripts/probe-fx-acp.ts Outdated
@molefrog molefrog changed the title Record fx structured session/load replay and add an fx ACP probe Add fx chats over ACP with resumable tools and reasoning effort Sep 11, 2026
@molefrog molefrog changed the title Add fx chats over ACP with resumable tools and reasoning effort Add fx chats over ACP with tool history and reasoning effort Sep 11, 2026
@molefrog molefrog changed the title Add fx chats over ACP with tool history and reasoning effort Add fx 0.0.9 chats over ACP with reasoning effort Sep 12, 2026
claude and others added 24 commits September 25, 2026 12:16
@claude
Record fx structured session/load replay and add an fx ACP probe
6980204 
vercel-labs/fx#788 (merged to fx main) fixes the lossy ACP `session/load`
that the acp notes tracked as finding #1 (fx#624): replay now emits a
`tool_call` plus terminal `tool_call_update` per recorded call instead of
folding the turn into one "Previous tool execution" assistant chunk.

`scripts/probe-fx-acp.ts` drives `fx acp` through moi's real ACP client,
session and adapter, runs one tool-calling turn, kills the process, and
compares the cold-loaded view with the live one. Verified on 2026-09-10:
a fx main build (6fdbe1f83029) replays the tool card identically with no
adapter change; the 0.0.8 release (43c11dcc34a9) predates the fix and is
still lossy. Both runs used fx's fake-gateway shape; the script uses the
real Vercel AI Gateway when AI_GATEWAY_API_KEY is set.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HNrsnBvXC2cHkFL73BSwN1
@claude
Re-check the fx ACP findings on main against the real gateway
38217a9 
The #788 replay fix now has real Vercel AI Gateway evidence
(anthropic/claude-sonnet-5) on both the 0.0.8 release and a main build,
matching the fake-gateway runs. The other fx findings in the acp notes
were re-checked on main over raw JSON-RPC: one active session per
process, mode reset after load, generic titles, invalid model acceptance,
the missing effort option, resume without replay, and replay without
timestamps or usage all still hold; thought replay could not be
triggered live and is ruled out by the replay planner's shape.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HNrsnBvXC2cHkFL73BSwN1
@molefrog @claude
Add fx chats over ACP with resumable tools and reasoning effort
583e068 
Use isolated chat and discovery processes, preserve structured tool replay, validate session settings, and harden cancellation and transport ownership. Retain model-specific effort choices across cold and warm chats and hide discovery-owned empty sessions.

Verified official fx dev f4ea28b23764 with real and fake gateway probes, 16 live lifecycle checks, browser workspace creation and cold replay, and the complete suite (1640 pass, 4 skip). Preserve the invoking Bun runtime so the dev workspace form uses the correct bundler.
@molefrog @claude
Make ACP pipe-failure tests portable across Bun versions
7dcce19 
Linux Bun 1.3.14 retains the fixture stdin pipe after closeSync(0), so the old test paused input instead of producing a broken pipe. Inject write and flush failures on a real child sink and assert both pending requests reject and the child terminates. Verified on the exact CI runtime and with the full local suite (1641 pass, 4 skip).
@molefrog @claude
Preserve fx thinking beside operational warnings
84c0b2f
@molefrog @claude
Verify fx 0.0.9 effort and generated chat titles
6bf148e
@molefrog @claude
Require fx 0.0.9 and preserve native chat settings
bee37f9
@claude
Restore full fx tool results and classify fx diagnostics
7d5a2f6 
fx clips every ACP tool result to a 200-byte preview, live and on
session/load. After a history load and after each prompt, moi now reads
fx's supported `fx session --id <id> --json` history and replaces the
previews with the saved results: full command output with its exit code,
full tool text, and fx's line diff for committed writes and edits.

Backgrounded shell commands (fx 0.0.11 yields long runs and streams late
output onto the finished call) keep their successful outcome and show only
real output; status lines are never concatenated with it. Provider
failures, the response-restart marker and replayed interrupted outcomes
become notices instead of assistant replies.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Render fx file results, edits and tool actions clearly
b3028bf 
fx file reads now show highlighted source without fx's path/content wrapper
and line gutter, writes show the written file, and edits render as a diff:
fx's own committed diff when its history is available, else the replaced
and new text while the run is live. Tool rows name what fx actually did
(waiting on or stopping a backgrounded command, reading a skill, fetching
a page, messaging a subagent) instead of fx's generic titles. The thinking
probe accepts PROBE_EFFORT for every model.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Keep queued fx messages after the running reply and summarize commands
09f4d3f 
ACP backends (fx, Hermes) hold a message sent mid-run until the current
run ends. The chat inserted it into the transcript immediately, so the
running reply landed after it and folded into the next run's work log.
Such sends now wait in a dimmed "Queued" bubble after the live reply; the
server's echo places the turn in order when it is dispatched. Codex,
OpenClaw and Claude Code keep their current behavior.

fx command rows also end with how the command finished (exit code and
duration, stopped, or timed out), and tool rows expose aria-expanded.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Title an untitled ACP chat from its first message
fef51e6 
Until fx generates a title (at the end of the first run), its session
list entry has none, and moi replaced the client's prompt-based title
with a generic "Untitled session" for the whole first turn. A chat that
is open in this server now uses its first user message instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Offer effort for a model before its first fx chat
3b67bba 
fx advertises a model's effort levels only for a session's current model,
so selecting a model moi had not used hid the effort control until that
model's first run. When the picker selects such a model, moi now switches
a throwaway discovery session to it and caches the advertised options
(`GET /agent?model=…`); the probe session is hidden like discovery's.
Models known to have no effort report it, so they are never re-probed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Show fx tool failures and replayed images readably
2d6fdac 
fx reports failed tools as `{"error":{…}}` JSON; rows now show the message
(with reviewer advice for a held action, or "Stopped before it finished."
for a cancelled call) and keep the envelope as raw output. A replayed
message with images no longer shows fx's `[Image #1]` marker lines. Tool
output wraps at word boundaries instead of splitting identifiers.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Coalesce streaming tool updates sent to chat clients
1b9660b 
fx streams command output as one update per line, and every broadcast
carries the whole accumulated tool turn, so a 400-line `seq` sent 405
frames (about 1 MB) and longer output grew quadratically. Progress that
keeps a tool in the same state now updates the view immediately but
reaches clients at most every 120 ms; a tool's appearance and every state
change are still sent at once.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Keep one copy of streamed fx shell output per update
8033f0e 
The shell normalizer kept the accumulated output a second time in
rawOutput, so every broadcast carried it twice more than needed. It now
continues from the row's own text and recognizes its status lines by an
exact pattern instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Keep following the reply after sending from a scrolled chat
6c00627 
Sending scrolls to the bottom before the new message renders. The scroll
event could arrive after the content had grown again, measure a large
distance from the new bottom, and un-pin the chat, so the streaming reply
stayed below the fold. Only a scroll toward older content un-pins now.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Keep a chat selected while another tab starts it
ed673e6 
The selected chat is shared by every open tab. A tab whose chat list did
not yet include a chat just started elsewhere cleared the selection at
once, and the clear reached the originating tab: its running chat
vanished behind an empty "New chat". A missing selection is now cleared
only when that chat is not running and stays unknown for three seconds.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Mark a stopped ACP run in the transcript
50f3ebb 
A stopped run ended with no marker, so a red tool row was the only sign
anything happened; after a reload fx's replay produced a notice instead.
A cancelled prompt now adds the same "This run was stopped before it
finished." notice live, so both views agree.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Refetch a chat after its history reloads instead of patching it
93538a8 
A cold load (after the idle release or an environment change) replays the
chat with new turn ids: user turns get replay ids and assistant turns
restart their counters, so broadcasting them patched an open client's
transcript with duplicated and overwritten turns. Replayed turns now stay
server-side, and one `session_reloaded` frame makes clients refetch the
transcript; events that arrive during the refetch are applied on top.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Document fx 0.0.11 behavior and label its vision tool
a710539 
The fx notes describe the restored tool results, backgrounded commands,
operational notices, queued messages, first-message titles and the effort
probe, and their remaining limits now reflect what is fixed. fx's vision
tool (used for models without native vision) gets a label and its focus as
the brief.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Keep an idle fx chat's live transcript across its release
6ca7b82 
Idle chats free their fx process after ten minutes. Reopening one reloaded
it from fx's replay, which has no thoughts and only clipped results, so a
chat the user had just read lost them. fx chats now keep the transcript moi
saw live (bounded, dropped when the chat is forgotten): it is served without
starting a process, and the next load still attaches through session/load
but keeps the retained transcript unless the history changed elsewhere (a
different number of user turns). New runs continue after its turn ids.
Replayed durations are applied before replay ends, so no replayed turn is
broadcast, and clients drop stale previews when a chat reloads.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Note that idle fx chats keep their live transcript
499db73 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Cut redundant fx broadcasts and list spawns, fix replay model labels
9b311d8 
- Reuse one warm fx process for session/list instead of spawning per call.
- Skip the title broadcast during a prompt; the turn end already sends one.
- Drop no-op tool updates and the duplicate final-turn meta emit.
- Label only the latest replayed run with the loaded model, never the
  model picked for the send that resumed the chat.
- Derive totalTokens when fx omits it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Show fx tool results as readable text instead of envelopes
e3d52e5 
Subagent reports and vision summaries render as markdown; fetched pages,
skills, saved command output and file searches drop their model-facing
wrappers. Shell status sentences and fx's diff line counts move to the
row's summary line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
claude added 14 commits September 25, 2026 12:18
@claude
List fx gateway models by name under vendor headings
67d9d54 
fx names each model by its raw id, so a few hundred 'vendor/model' rows
sat in one section and the composer truncated them to the same prefix.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Capitalize fx vendor headings the catalog does not name
71aa4eb 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Document the fx 0.0.11 verification and session changes
60aa458 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Make the fx replay probe compare content and await process exit
38aab04 
The verdict now requires replayed tool names, inputs and message text to
match the live transcript, real-gateway runs drop inherited fx gateway
overrides, and the cold load waits for the killed process to exit.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Unwrap fx skills and pages cut before their closing tag
afa4548 
fx saves at most 4,096 bytes of a tool result, so a long SKILL.md or page
reached the chat without its closing tag and kept its envelope.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Note fx's 4,096-byte cap on saved tool results
f179df6 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Keep full streamed fx output when history saved only part of it
1c17257 
fx saves at most 4,096 bytes of a result, so a long command's saved
envelope is cut mid-JSON. The post-run history read replaced the complete
streamed output with that raw fragment. A cut envelope now yields its exit
status and, only for a row with no output of its own, the decoded prefix.

Also from review:
- The enrichment hook sees the current rows and sets failure text
  explicitly, so a cold-loaded failed command shows its saved output.
- Late output from a backgrounded command no longer splits the reply
  the model is writing.
- A stop while a cold chat reads its history cancels the send.
- A reconnect that fails or is stopped keeps the retained transcript.
- A new ACP chat reports running while its agent starts, so other tabs
  keep a shared selection they do not list yet.
- fx reads cut before their closing tag are unwrapped and marked.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Note how fx's saved-result cap affects restored output
b084119 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Keep live fx shell output and place replayed notices after their turn
1b74358 
From the re-verification runs:
- A shell row keeps the output it streamed live; fx's saved copy (at most
  4,096 bytes, stdout and stderr interleaved mid-line) fills only a row
  with none of its own. Restored rows get their duration, a command that
  printed nothing reads as no output, and a backgrounded command keeps its
  early output after a cold load.
- Replayed notices name the turn they follow: replayed turns carry no
  time, so a stopped run's notice used to land at the end of the chat.
- An environment change keeps idle fx chats' live transcripts.
- The warm chat-list process stays for ten minutes, like an idle chat.
- Vendor headings for spacexai and inclusionai; '1 match' for one result.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Note that fx can repeat text when replaying an interrupted turn
40a0bff 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Keep a live wait row's outcome when fx history is restored
fad138d 
The saved result of fx's shell interact call is the output the command
printed while waited on, which the command's own row already shows live.
It now fills the wait row only after a cold load.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Place a replayed notice after the run that contains its turn
21bb898 
The chat merges a run's assistant turns under the first turn's id, so a
notice naming a later turn in the run found no match and fell to the end
of the chat. Anchors now resolve through the grouping.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Keep Hermes env and stop hooks to Hermes chats
390b869 
Every harness hears an env change, and Hermes forgot and killed every ACP
session in the workspace before fx's own hook ran. An fx chat therefore
lost the transcript it was meant to keep across the restart, and its long
shell output shrank to fx's saved 4,096 bytes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@claude
Record the fx 0.0.11 re-run and final spot check
116af4c 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019wu29L9PZkqMDtA224wbUy
@molefrog
molefrog force-pushed the claude/sweet-faraday-fy7od4 branch from e50d57d to 116af4c Compare September 25, 2026 12:26
@molefrog molefrog changed the title Add fx 0.0.9 chats over ACP with reasoning effort Add fx chats over ACP with reasoning effort and full tool results Sep 25, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@chatgpt-codex-connector chatgpt-codex-connector[bot] chatgpt-codex-connector[bot] left review comments

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@molefrog @claude