heartmove/dsh-session-bridge

Targets DSH 0.1.7-0 and later: both the peer ranges and package.json's engines.dsh declare a floor only (>=0.1.7-0), never a ceiling. Since DSH 0.2.1-rc.1 engines.dsh decides alone (when present, peerDependencies are ignored), so the two state the same floor; verified against 0.2.1-rc.1. See compatibility changes and AGENTS.md (the version policy is a hard rule, guarded by pnpm test).

A DSH plugin that lets the current agent drive other real DSH sessions directly from a prompt — create sessions, send messages to any session, wait for and read replies, resume offline sessions, and find sessions across workspaces. On top of that it can monitor and schedule a main task (watch its progress, nudge or correct its direction, and stop it), and archive sessions the same way the DSH sidebar's Archive action does.

中文文档见 README.zh.md.

What it does

• Create real DSH sessions. session_bridge_create makes a new main session (top-level UI session) in the current workspace, or in another workspace when you pass workspaceId / cwd. It can send one first prompt and optionally block until the first reply. Provider / model / reasoning effort are inherited from the calling session by default. An async create (no waitForReply) returns a sinceSeq anchor for a later precise wait.
• Send messages to any session. session_bridge_send appends a turn (mode=queue) or injects steering into the running step (mode=steer), and can optionally wait for the next reply. An async send also returns a sinceSeq anchor.
• Wait for a reply or a segment. session_bridge_wait blocks until new assistant output appears after sinceSeq (default: the latest event seq at call time): waitFor=reply (default) returns as soon as a new text reply is readable; waitFor=segment returns as soon as any new completed output step appears (an assistant/message — text, reasoning, or tool-call turn), without waiting for the whole turn, so you can observe output paragraph by paragraph as it is produced. With requireTurnEnd it additionally waits for the turn to settle. Timeout / abort return the partial result rather than throwing. An already-landed reply is never lost: when no new output arrives within the budget, the latest PRE-EXISTING reply/segment is returned with stale: true (no more (no text)). To retrieve exactly "the reply to what I sent", pass the sinceSeq returned by session_bridge_send / session_bridge_create (works regardless of caller latency); sinceSeq: -1 counts existing events too, i.e. the anchor for a brand-new session.
• Read any session. session_bridge_read folds a session's event log into readable rows — live or offline (from persistence) — with sinceSeq paging, role filtering, and a limit (default 20, max 100).
• Read output paragraph by paragraph. session_bridge_segments returns the session's completed output segments — every finished assistant step (one assistant/message: its text, reasoning, and requested tool calls) as its own row, paged forward via sinceSeq, returning the next cursor. It works live or offline and does not wait for the whole turn, so you can follow a long agentic run step by step (chain-of-thought included when the model streams it).
• Resume offline sessions. session_bridge_resume brings a persisted session back online (idempotent); it can also override provider / model.
• Find sessions. session_bridge_find matches by title, id, workspace, or directory across all workspaces, returning live/running state, title, and working directory. Bridge-registered titles act as aliases.
• Monitor and schedule a main task. session_bridge_status reads a session's real-time progress (running/idle, open turn, time since the last event for stall detection, pending work, latest reply); only a running session is flagged [STALLED] (a quiet idle session is not stuck — same rule as the watchdog). session_bridge_cancel stops a running session. session_bridge_monitor_start runs a background watchdog loop that polls the task, nudges it when it stalls, corrects it when it drifts, terminates it after it stays stuck, and wraps up when it finishes. When the chain-of-thought loops (e.g. 写。好。Go。 over and over) it interrupts the session and transfers the task to a new session (with the subagent list carried over) so the work continues.
• Archive / unarchive sessions. session_bridge_archive adds a session to the DSH workspace archive set (hidden from every grouping surface, history and workspace position preserved); an active target is refused unless stopActivity: true is passed, which archives first and then stops its running work (turn, subagents, jobs, schedules) through the official path. session_bridge_unarchive drops it from the archive set so it reappears at its recorded position. session_bridge_archived lists the archive set, optionally resolving titles; a session whose title cannot be resolved is listed without one instead of failing the whole call. The archive set grows without bound, so archive/unarchive results report the affected id plus the archive size and summarize the newest ids, and session_bridge_archived windows the listing with an optional limit (default 50, newest first) while total always reports the real size.

Monitoring worker

session_bridge_monitor_start installs a timer-driven loop. Every poll it observes → judges → schedules → logs the target session:

Observation	Action
A doneKeywords string appears in the reply and the session is idle	Wrap up and stop the watchdog (log DONE)
Idle with no pending work	Wrap up (settled) — no pointless nudging / cancelling
running and no event for more than stalledMs	Record a stall → steer a nudge (with useLlm, judge offtrack/stuck first)
Stall repeats ≥ maxStuckCycles	cancel the session
Making progress	Reset the stall counter (steady)
running and the chain-of-thought loops (on by default)	Create a successor session carrying the task → interrupt the old one (transfer, log LOOP-TRANSFER)
Loop, but the transfer budget is exhausted / loopTransfer: false	Interrupt only (log LOOP-CANCEL)
A subagent's chain-of-thought loops	Cancel that subagent + steer the main session to re-delegate (no transfer)

The watchdog only treats running sessions as stalled, so a finished or idle task is wrapped up rather than nudged forever (the [STALLED] marker in session_bridge_status follows the same rule and is shown for running sessions only). Logs go to ~/.dsh/super-injector/dsh-session-bridge-monitor.log (overridable).

Control it with session_bridge_monitor_start / _stop / _list.

Chain-of-thought monitoring & rules

The bridge can watch another session's chain-of-thought (reasoning) in real time — not just its final reply — and act on it:

• Observe it live. session_bridge_status returns three chain-of-thought fields on a running session: lastReasoning (the most recent finalized reasoning block), liveReasoning (the in-flight reasoning streamed for the current handled turn, from assistant/chunk reasoning-delta events), and reasoningTail (a compact, char-bounded merged preview). Use the reasoning param (none | last | live | tail) to pick which fields come back; tail is the default and costs the least.

• Read paragraph by paragraph. session_bridge_segments returns each completed output step (one assistant/message — text, reasoning, or tool-call turn) as its own segment, paged forward via sinceSeq, without waiting for the whole turn.

• Or read it per message. session_bridge_read with includeReasoning returns the finalized reasoning of each assistant message.

• Enforce rules. session_bridge_monitor_start accepts coRules — an array of { match: contains|not-contains, field: reasoning|text|both, value: string, action: steer|cancel, message?: string } and cotMinHits. Each poll the watchdog matches the rule's condition against the live reasoning/text and, once it stays matched for cotMinHits consecutive polls (default 1), fires the action: steer injects a guiding user message, cancel terminates the session. Example — "stop the session the moment its reasoning no longer contains I'm" becomes:

  coRules: [{ "match": "not-contains", "field": "reasoning", "value": "I'm", "action": "cancel" }]
  

  Caveats: a not-contains rule on reasoning deliberately does not fire when the target session produces no reasoning at all (e.g. a non-reasoning model or reasoningEffort off), so you don't cancel sessions that simply don't stream a chain-of-thought. Repeated fires are throttled by a cooldown, and each evaluation/trigger is written to the monitor log.

Chain-of-thought loop detection & task transfer

Why the stall check cannot catch a loop: a looping session lands new events every cycle, so "ms since the last event" stays tiny and stalledMs never crosses the threshold — the stall branch is blind to loops by construction. A loop therefore has to be judged from the content itself.

Judging it (src/loop-guard.ts — pure functions, no host dependency): split the chain-of-thought into units on 。！？；!?; and newlines (punctuation stays inside the unit, so 写。 and 好。 are different units), then look for a periodic repeat of a small run inside a tail window (1200 chars). Measured: 写。/好。/Go。 ×3 → period=3 repeats=3 fires; 3 units × 2 rounds and 2 units × 2 rounds do not (the repeat floor is 3); long ordinary reasoning does not fire; a long prefix cannot hide a tail loop; punctuation-free repetition is caught by a character n-gram fallback (kind=char).

Parameter	Default	Meaning
loopGuard	true	false disables loop detection
loopField	reasoning	What to scan: reasoning (live chain-of-thought, liveReasoning first, else lastReasoning) / text / both
loopMinCycles	2	Minimum repeated cycles
loopMinRepeats	3	Minimum repeats of one unit; the effective floor is max(loopMinCycles, loopMinRepeats)
loopMinConfidence	0.6	Confidence threshold (weighted by unit length, covered rounds, period length)
loopMinHits	1	Consecutive polls with the same signature (field+period+unit) before acting
loopCooldownMs	max(intervalMs*2, 30000)	Per-signature cooldown, so a jittery signal cannot double-fire
loopMaxTransfers	3	Maximum transfers (also the transfer-depth cap)
loopTransfer	true	false = interrupt only, never transfer
loopAutoMonitor	true	Whether the successor session is watched too
loopDescendants	report	Subagent policy, see the table below
loopHandoffPrompt	built-in template	Custom handoff prompt; placeholders {{sessionId}} {{cwd}} {{reason}} {{task}} {{evidence}} {{descendants}}

On a hit the order is not negotiable: create the successor main session in the original working directory and deliver the handoff prompt first, and only cancel the old session once that succeeded. The reverse order (stop, then create) loses the task outright if creation fails; this implementation instead treats a failed creation as "keep the original watchdog, cancel nothing" and retries on a max(intervalMs, 10s) throttle. The handoff carries the original session id, the task instruction, the recent instructions / last reply / last chain-of-thought, and the subagent list — and tells the successor to session_bridge_read the old session log itself (no full transcripts inlined, to save tokens). The transfer chain inherits the same transfer record, so the budget does not reset with each successor.

When the budget is exhausted (loopMaxTransfers or the depth cap) or loopTransfer: false, the watchdog interrupts only (log LOOP-CANCEL). Log lines: LOOP-HIT / LOOP-TRANSFER / LOOP-TRANSFER-FAILED / LOOP-CANCEL / LOOP-SUBAGENT-CANCEL / LOOP-SUBAGENT-NOTIFY.

What happens to subagents

Scenario	Handling	Why
A. Main session loops, subagents still running normally	Not cancelled, listed in the handoff (report, the default)	Their work is not wasted; the successor is told not to redo it and can wait for results
B. Main session loops with loopDescendants: 'stop'	Cancelled one by one before the transfer, recorded under stopped	You asked for a clean handoff with no orphaned work
C. Main session loops with loopDescendants: 'ignore'	Neither cancelled nor mentioned	The subagents are irrelevant to the remaining task
D. A subagent loops (main session fine)	Cancel that subagent + steer the main session to re-delegate; no transfer	The broken thing is the subtask, not the task; the main session knows how to re-delegate
E. A subagent session itself is the monitored target and loops	Cancel only, never transfer	Its identity is a subagent; spawning a new top-level session would break the execution tree. The parent session is notified

"Who is a subagent" is decided with host capabilities first (agents.isOwnedBy / roots()), then falls back to the session header's origin === 'subagent' / parentSession, and finally treats the session as top-level — it can under-transfer, never mis-transfer. The transfer chain is bounded by loopMaxTransfers: a successor that loops again gets at most another transfer, then only interruptions, so there is no A→B→C relay.

Enabling and disabling (two ways)

Loop detection is on by default, and both the master switch and the global defaults live in the plugin config (the config of this plugin's entry in the profile patch), so the two paths below change the same value:

1. The /cot-loop slash command — change it from the interactive UI; it writes the config back and takes effect immediately (no restart, no plugin reload):

/cot-loop status                      # switch state, detection parameters, settings entry
/cot-loop off                         # disable chain-of-thought loop auto-interrupt (globally)
/cot-loop on                          # turn it back on
/cot-loop field text                  # scan the final reply only (default: reasoning)
/cot-loop descendants stop            # cancel still-running subagents before transferring
/cot-loop max-transfers 3             # transfer budget (0 = interrupt only)
/cot-loop help                        # every subcommand

Parameter changes apply from the next session_bridge_monitor_start; on / off apply to already running watchdogs at once (each tick reads the volatile reference live). When the host has no configEditor service (or the write fails) the command degrades to an in-process switch and says "this run only" — it never pretends the value was persisted.

2. The settings toggle — Settings → Plugins → dsh-session-bridge: the form is rendered automatically from the volatile fields of Config (dsh-settings only renders volatile fields as controls, so this half needs no client code at all), and loopGuard is the master switch. Saving it writes the profile patch and applies immediately as well.

A single watchdog can still opt out on its own: session_bridge_monitor_start({ sessionId: '...', loopGuard: false }).

Plugin config field	Default	Meaning
loopGuard	true	Master switch: false stops loop detection for every watchdog (this is what the settings page / /cot-loop toggles)
loopField	reasoning	What to scan: reasoning / text / both
loopMinCycles	2	Minimum repeated cycles
loopMinRepeats	3	Minimum repeats of one unit
loopMinConfidence	0.6	Confidence threshold
loopMinHits	1	Consecutive hits before acting
loopCooldownMs	60000	Per-signature cooldown in ms
loopMaxTransfers	3	Transfer budget (0 = interrupt only)
loopDescendants	report	Subagent policy, see the table below

Requirements

• Node.js ≥ 20
• pnpm
• DSH 0.1.7-0 or later (peerDependencies and engines.dsh both >=0.1.7-0; built and tested against 0.2.1-rc.1).

Build

# Install locked DSH release packages, then type-check and build
pnpm install --frozen-lockfile
pnpm build

# type-check src/ against the dsh that is actually installed (no checkout needed)
npm run check:compat

# regression tests: peer floor, wait/stall core, loop detection + handoff, watchdog loop, archived handler, V3→V4
npm test

# via the injector toolchain
dev_build_plugin dsh-session-bridge

Builds use the locked 0.2.1-rc.1 registry packages; no source checkout is required. check:compat type-checks src/ against the installed harness and then compares the artifact's inlined DSH version with it: inlining something older than the declared floor (0.1.7-0) fails, any other drift is a note (--strict restores the exact-match failure). The inlined cordis copy is tracked separately and compared with the installed cordis. CI uses the same lockfile.

Deploy

DSH web loads external plugins from the active profile. This package is a bundle: its package.json declares dsh.bundle.patch → cordis.patch.yml, whose insert row mounts the plugin. That declaration is what lets dsh plugin add install the package and activate it in one step.

Install from npm

The package is published to npmjs.com. Releases are cut on a v* git tag by the publish.yml GitHub Actions workflow; package.json and dsh.plugin.json versions are synced to that tag before publishing.

npx -p @deepseek-ai/dsh dsh plugin --profile web add dsh-session-bridge

A prerelease tag (v0.3.2-alpha.1) is published to its own dist-tag (alpha, beta, rc) instead of latest, so it cannot displace the stable release for other users. Opt in explicitly:

npx -p @deepseek-ai/dsh dsh plugin --profile web add dsh-session-bridge@alpha

pnpm installs the published tarball and runs its prepare script (tsdown) to ensure lib/ is present, then dsh activates the bundle.

For ~24 hours after a release, @latest resolves to the previous version. pnpm 11 enables supply-chain protection by default (minimumReleaseAge defaults to 1440 minutes): versions published less than 24h ago are excluded from resolution, and because minimumReleaseAgeStrict defaults to false, pnpm silently falls back to the newest old-enough version instead of erroring. pnpm view dsh-session-bridge dist-tags still reports the new latest, but the install picks the old one — whose peer range a newer DSH may already reject, which reads as "I asked for latest and it still says incompatible". Three ways out:

1. Exclude the package in the consuming profile's pnpm-workspace.yaml so @latest takes effect immediately:

   minimumReleaseAgeExclude:
     - dsh-session-bridge
   

2. Install a local tarball (no registry resolution): dsh plugin --profile <p> add <path>/dsh-session-bridge-x.y.z.tgz;
3. Wait 24 hours. Pinning an exact version does not bypass the policy (it fails hard with ERR_PNPM_MINIMUM_RELEASE_AGE_VIOLATION).

Install from GitHub

npx -p @deepseek-ai/dsh dsh plugin --profile web add github:heartmove/dsh-session-bridge

dsh plugin forwards to pnpm inside ~/.dsh/profiles/web/, then reconciles the bundle into the profile's dsh.profile.bundles layer list. A git install fetches sources, so pnpm runs the package's prepare script (tsdown) to build lib/ from src/ after checkout.

pnpm ≥ 10 refuses to run a git dependency's prepare script until it is allowlisted, so the first add fails with an "Ignored build scripts" hint. Copy the exact package key pnpm printed into the profile's pnpm-workspace.yaml (~/.dsh/profiles/web/pnpm-workspace.yaml):

allowBuilds:
  dsh-session-bridge: true

then re-run the add. That allowance means "run this package's code on my machine at install time" — only allow packages whose source you trust, and pin a commit (github:heartmove/dsh-session-bridge#<sha>) so a later push cannot silently change what runs.

Restart dsh web, then hard-refresh the page (Ctrl/Cmd+Shift+R).

Install from a local checkout

From the directory that contains this checkout:

npx -p @deepseek-ai/dsh dsh plugin --profile web add ./dsh-session-bridge

pnpm links the checkout and dsh activates the bundle the same way.

Manual link

To manage the profile by hand, link the package and list it as a bundle in ~/.dsh/profiles/web/package.json (the bundle's own cordis.patch.yml supplies the loader row, so no insert entry is needed):

{
  "dependencies": {
    "dsh-session-bridge": "link:D:\\path\\to\\dsh-session-bridge"
  },
  "dsh": {
    "profile": {
      "bundles": ["@deepseek-ai/dsh-base", "@deepseek-ai/dsh-web-app", "dsh-session-bridge"]
    }
  }
}

(On POSIX systems use link:/path/to/dsh-session-bridge.) Then run pnpm install in the profile directory and restart dsh web.

Inject directly (dev)

For fast iteration while developing the plugin, you can also load it directly through the injector toolchain (no bundle entry required):

dev_inject_plugin D:\code\dsh-session-bridge

Remove it with dev_uninject_plugin dsh-session-bridge (clears the injector registration and junction; not re-assembled on restart).

Usage

Tool	What it does
session_bridge_create	Create a main session (current or another workspace via workspaceId / cwd); optional first prompt + waitForReply; async creates return a sinceSeq anchor.
session_bridge_send	Send a message (mode=queue/steer); optional wait-for-reply; async sends return a sinceSeq anchor.
session_bridge_wait	Wait for new output after sinceSeq (default: latest seq at call time; -1 counts existing events): waitFor=reply (text) or waitFor=segment (any completed step, no full-turn wait); optional requireTurnEnd; falls back to the pre-existing reply with stale when nothing new arrives.
session_bridge_read	Read messages — live or offline; sinceSeq paging, role filter, limit.
session_bridge_segments	Read completed output segments (each finished assistant step) incrementally by paragraph — live or offline.
session_bridge_resume	Bring a persisted session back online (idempotent).
session_bridge_find	Find sessions by title / id / workspace / directory across workspaces.
session_bridge_status	Read a session's live progress (running, open turn, stall detection, pending work, latest reply) plus live/finalized chain-of-thought (reasoning param); [STALLED] is shown for running sessions only.
session_bridge_cancel	Stop a running session (abort active turn; clear queued/steering work unless keepInbox).
session_bridge_monitor_start	Start a background watchdog on a main session (poll, nudge, correct, cancel, wrap up); supports chain-of-thought coRules (e.g. reasoning not-contains "I'm" → cancel) and chain-of-thought loop detection (on by default: a hit interrupts the session and transfers the task to a new one, with a subagent policy via loopDescendants).
session_bridge_monitor_stop	Stop a watchdog (keep the session itself running).
session_bridge_monitor_list	List active watchdogs and their state.
session_bridge_archive	Archive a session (hidden from groupings; history and position preserved); requires stopActivity: true for a running one. Reports the id, the archive size and the newest ids.
session_bridge_unarchive	Unarchive a session (visible again at its recorded position; unknown ids are a no-op).
session_bridge_archived	List the archive set (newest first, optional limit, default 50), optionally resolving titles for the returned ids.

All tools return lossless JSON; wait-style tools never throw on timeout — they return a timedOut / aborted / stale flag.

Project layout

src/
  index.ts    host plugin entry (registers the tools; mounts the monitor)
  core.ts     shared host logic (create/send/wait/read/find, status snapshot, archive attach)
  tools.ts    tool registrations (bridge + status/cancel + monitor + archive)
  monitor.ts  the background watchdog loop (statusSnapshot + rules + loop detection/transfer + optional LLM judge)
  loop-guard.ts  chain-of-thought loop detection (pure: unit split + tail periodic repeat + n-gram fallback)
  handoff.ts  the handoff prompt (task/context extraction + subagent sections + template)
  session-create.ts  the single session-creation implementation (shared by the create tool and loop transfer)
  registry.ts bridge-side title/workspace registry (~/.dsh/session-bridge-registry.json)
scripts/
  build.sh    compatibility wrapper for npm run build
  test-bridge-core.mjs  wait/stall regression tests (npm test)
  test-loop-guard.mjs  loop-detection regression tests (写。/好。/Go。 shapes, tail window, fallback)
  test-handoff.mjs  handoff-prompt regression tests (placeholders + subagent sections)
  test-monitor-loop.mjs  end-to-end watchdog loop tests (transfer, the five subagent scenarios, budget, failure)
  test-tools-archived.mjs  archive/unarchive + archived-handler regression tests (npm test)
  check-dsh-compat.mjs  type-check src/ against installed DSH + verify bundle provenance (npm run check:compat)
  smoke-bundle.mjs  mount the built lib/index.js and assert every tool registers (npm run smoke)

Lifecycle and unloading

DSH ≥ 0.1.6 mounts and unloads plugins at runtime (Settings → Plugins, and the injector's hot reload). This plugin unloads cleanly: it registers no loader-level state, its tools are disposed with the plugin fiber, and the monitor's timer is released through ctx.effect on unload.

One consequence of that ownership model: sessions created by session_bridge_create are owned by the plugin fiber (the agent is created under the plugin's context), so unloading or reloading the plugin stops their live agents. The sessions themselves are persisted and show as offline; bring one back with session_bridge_resume. Monitor watchdogs also stop on unload.

License

MIT