Skip to content

Record the installed NeoTech stack on ds-bot 0.3.2 - #16

Open
neotechnet wants to merge 9 commits into
upstream/v0.3.2from
stack/installed-0.3.2
Open

neotechnet wants to merge 9 commits into
upstream/v0.3.2from
stack/installed-0.3.2

Conversation

@neotechnet

@neotechnet neotechnet commented Nov 10, 2026 •

Copy link
Copy Markdown
Owner

Records the NeoTech local stack as installed on ds-bot 0.3.2. Not for upstream, not meant to merge; same role as #4 for 0.2.7.

What is in it

Five commits over upstream v0.3.2:

  • c76ba068 Carry the local stack onto 0.3.2 (part 1 of the 0.3.1 stack)
  • 07165163 Carry the local stack onto 0.3.2 (part 2 of the 0.3.1 stack)
  • 2133df4c Pass the system prompt to Claude Code in a file
  • 7fa7c623 Carry the codebots Claude Code settings onto 0.3.2
  • 5a9571a7 Fit the local stack to 0.3.2's new waits, adopted turns and run-mode tests

Installed into the live desktop profile as patches/ds-bot@0.3.2.patch (406,429 bytes, sha256 d48ea978b6065342ba1a3cf4c2f398b1ae1b99bfb7d933bc1b875c6ed7062ad1) on 2026-11-09 19:31 local.

Proven live (2026-11-10)

  • gate-loaded at 2026-11-10T11:13:42Z after the operator restart; state.lock PID 9748 alive.
  • 47 bots, 7 rooms, state revision 20223 to 20231.
  • No new crash log after the start. updateCheck: false and checkRatio: 0.35 intact.
  • 17 hook/result events after the start, all exitCode: 0.
  • Operator confirmed a Bot opens from the sidebar and replies.

Base upstream/v0.3.2 is the upstream tag, pushed as a branch so the diff shows only the local stack.

🤖 Generated with Claude Code

neotechnet and others added 5 commits November 9, 2026 19:16
Replays 2da086d (the 0.2.7 stack carried onto 0.3.1) on top of v0.3.2.

Conflicts and how they were resolved, keeping 0.3.2's structure:
- sidebar.js: 0.3.2 added collapsible sections with drag and drop and a
  section-aware glide key. The Main Bot tile grid now sits above the
  ungrouped list and the sections, and every Main Bot is kept out of the
  section lists, as before. Both ui selectors (renameSection and
  accountSeated) are kept.
- handoffs.js: 0.3.2's held() returns { id, work } pairs and the waiting
  view gained a waits list for the give-up menu. The waiting-live-only
  filter now reads the pair, and waits lists only the live waits.
- delivery.js: the local handedBack flag joins 0.3.2's new deliver()
  parameters (fromName, abandoned, recordExchange).
- bot-messages and group-work tests: assert both 0.3.2's waits entries
  and the local waiting-live-only shape.
- client.js rebuilt from src/client.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replays b3307ed (the true installed stack: operator tasks, credit guard,
brainstorm, the ds-bot:sessions event, card gate and parts additions) on
top of v0.3.2.

Conflicts and how they were resolved, keeping 0.3.2's structure:
- index.js: 0.3.2's module list (which adds groupLog) with operatorTasks
  after im and creditGuard last, as installed.
- client index.js: 0.3.2 moved the shell sheets into SHELL_SHEETS so the
  settings dialog can borrow them; the operator tasks sheet joins that
  table after the schedules sheet instead of its own install line.
- sources.js: 0.3.2's EMPTY_ROSTER (sections) plus operatorTasks.
- handoffs.js and turns.js: 0.3.2's runtime exports (abandonWork,
  abandonedWork, adoptLiveWork) plus the local workOrigin and
  operatorTurnOf.
- view-reply test: the local stack's dropped closedCards line stays
  dropped; 0.3.2's dock-record assertion and its new no-call-row test
  are kept.
- client.js rebuilt from src/client.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cherry-picked from a7e4357 (fix/claude-code-enametoolong), the source form
of the 2026-11-09 system-prompt-file hunk in the installed 0.3.1 patch. It
applies cleanly to v0.3.2, whose claude.js still passes the whole prompt as
--append-system-prompt, so 0.3.2's Windows spawn fix (spawnAgent for npm
and pnpm shims) does not cover ENAMETOOLONG.

A Bot's assembled system prompt was pushed onto the command line as one
argument. Windows allows 32,767 characters for a whole command line, so a
Bot with a large instruction set never started: the process failed at launch
with an opaque spawn ENAMETOOLONG before any model call.

The prompt now goes to a per-session file under the OS temp folder and only
its path is an argument (--append-system-prompt-file). The file is removed
when the process exits or is killed, and a startup sweep deletes anything
older than a day that a hard kill left behind. A guard rejects any command
line over 30,000 characters so an oversized argument is named.

The installed 0.3.1 patch also carried a startup-deadlock hunk (count the
child's spawn as started) and an adapter hunk that saved the late Claude
session id. v0.3.2 ships both upstream (start resolves on 'spawn', and the
adapter waits on whenSessionId()), so neither is carried.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The installed 0.3.1 patch's claude.js hunk (2026-11-09) carried three
local settings beside the system-prompt file, ported here unchanged in
effect onto 0.3.2's engine:
- the DSH-built prompt's "powered by the <model> model" line is rewritten
  to name the Claude Code program and its model;
- --effort high unless the engine sets its own;
- --strict-mcp-config with only the DS Bot tool server and the work-history
  search (claude-memory-bridge), never the user's business MCP servers.

0.3.2's --permission-prompt-tool stdio and --allowedTools mcp__dsbot stay
as upstream has them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tests

Three places where a local feature met code that is new in 0.3.2:

- waiting-live-only and 0.3.2's subagent waits: a subagent run is a work
  child with neither a Bot nor a group, so the local filter, which asks
  whether the child's Bot is still on it, dropped it and the give-up menu
  lost its entry. A child with no Bot now always counts, as a group wait
  does. The wait shape for a request a Bot sent on its own stays without a
  waits list, since there is no handed-on work to give up there.
- hand-back pointer and 0.3.2's adopted-turn requests: a request that
  arrives during a recovered turn is read in that Session, so the local
  pointer replaces its repeat. The 0.3.2 test now expects the pointer and
  the handedBack flag.
- run-mode tests: they called the last pre-step listener, assuming the part
  switch; the local step failure reporter is registered after it so that it
  wraps it. The tests now find the switch by name, as the other part tests
  do. Runtime order is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 10, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

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

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-10T11:24:56.622824Z 5a9571a 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.

@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: 5a9571a752

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 ds-bot/src/host/parts.js
Comment on lines +1334 to +1336
service[seam.method] = async function wrappedSeam(...args) {
try {
return await original.apply(this, args)

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

P1 Badge Replace DSH service monkey-patches with supported hooks

Installing the plugin now directly replaces the shared DSH tools.resolve and systemPrompt.assemble methods. This violates the explicit “Build on DSH; never patch it” rule in AGENTS.md. These replacements also have no disposer: unloading or reloading DS Bot leaves wrappers referencing the old runtime, and each installation adds another wrapper because wrappedSeams is local to that installation. Use documented diagnostic hooks instead; if DSH exposes no suitable hook, report that limitation rather than replacing its service methods.

AGENTS.md reference: AGENTS.md:L22-L22

Useful? React with 👍 / 👎.

Comment on lines +190 to +193
const response = await fetchFn(url, { headers: { 'x-api-key': key, 'anthropic-version': '2023-06-01' } })
if (response.status === 401 || response.status === 403) throw Object.assign(new Error(`Admin API ${response.status}`), { kind: 'bad-key', status: response.status })
if (!response.ok) throw Object.assign(new Error(`Admin API ${response.status}`), { kind: 'error', status: response.status })
const body = await response.json()

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

P2 Badge Bound Admin API requests so the blind cutoff can execute

When the Admin API stalls while returning headers or streaming its response body, this request has no abort signal or deadline. check() waits for read() before evaluating blindFor, and subsequent checks and low-credit error handling are serialized behind that same pending operation through busy. Consequently, a stalled meter can prevent both the configured blind cutoff and an immediate low-credit switch from running while Bots continue using the credit route. Apply a deadline covering both the fetch and body consumption, and treat expiration as a failed reading.

Useful? React with 👍 / 👎.

Comment on lines +265 to +269
const botId = rt.selfOf?.(sessionId)
if (botId === undefined && !fresh) return undefined
const done = rt.state?.planDefaultOff ?? (rt.state === undefined ? undefined : (rt.state.planDefaultOff = {}))
if (done === undefined || done[sessionId] !== undefined) return undefined
const bot = rt.botOf?.(botId)

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

P2 Badge Resolve the owning Bot before applying fresh-session defaults

When a Bot with planMode: 'on' or an entry in plan-mode-keep.json rolls into a fresh part, parts.js calls rt.createSession() before assigning that part to the Bot. The new fresh: true call therefore reaches this code with no botId, bypasses both opt-out checks, disables plan mode, and records the Session as processed. Once ownership is assigned, later calls return at the done check and cannot honor the Bot's setting. Pass the intended Bot into fresh-session initialization or defer the decision until ownership is registered.

Useful? React with 👍 / 👎.

The status line other Bots read named a hand-off as still open whenever the work kept an open id, even after that reply had come back. Nothing cleared the id when the answer arrived by another route, so an idle Bot looked loaded on work that had ended days before, and every job watch chased it.

A work now stops counting as waiting when the reply for an open id is already on it, and when it has sat past a day with no result at all. A fresh work with no reply still shows as waiting.
The ceiling first applied to every open id, so a work holding a subagent run for more than a day dropped that wait and took the whole entry out of the wait view, which is the view a restart relies on. A subagent run has no Bot to hand back, so the arrived and ceiling rules now apply only to Bot hand-offs.
A work kept the open id until the Bot that asked got round to reading the answer, so a row read open with its result already attached and every watch saw a phantom handoff. Writing the reply now clears the open id in the same save that attaches the result.

A reply that never reaches the asker opens the wait again until the notice carrying it is read, which is the case the follow-through tests already pinned.

An ask sent outside any work now marks its sender waiting only while it is recent, so an old unanswered ask cannot label an idle Bot for days.
Stop showing a work as waiting once its reply has arrived
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant