diff --git a/changelog.d/2677-quoted-failure.fixed.md b/changelog.d/2677-quoted-failure.fixed.md new file mode 100644 index 000000000..5428cb2b2 --- /dev/null +++ b/changelog.d/2677-quoted-failure.fixed.md @@ -0,0 +1 @@ +A workflow step no longer fails because its own answer quotes `could not answer your message` while auditing that sentence: the host marks its real failure notice structurally, and an old host without the marker is read only from the notice shape at the start of the reply ([#2677](https://github.com/ThinkFlowLab/FlyBot/issues/2677)). diff --git a/core/src/host/turns.js b/core/src/host/turns.js index 24f0966a2..ca1945622 100644 --- a/core/src/host/turns.js +++ b/core/src/host/turns.js @@ -711,6 +711,8 @@ export function install(rt) { hop: (source.hop ?? 0) + 1, replyTo: source.exchangeId, origin, + // #2677: the same marker the remote notice uses, so a reader never parses the sentence. + failure: true, }).catch(error => rt.warn('ds-bot: failure notice delivery failed: %s', error)) } else if (reply !== '' || reason?.kind === 'aborted') { // A turn stopped before it could answer still owes the sender a line, diff --git a/core/test/failure-notice-marker.test.js b/core/test/failure-notice-marker.test.js new file mode 100644 index 000000000..227bb5e41 --- /dev/null +++ b/core/test/failure-notice-marker.test.js @@ -0,0 +1,56 @@ +// #2677: a workflow step used to be failed whenever its own answer quoted the failure +// sentence. The host's notice now says structurally that the turn failed, so a reader does +// not have to parse the text it may be auditing. +import assert from 'node:assert/strict' +import { test } from 'node:test' +import { install as installTurns } from '../src/host/turns.js' +import { SOURCE_KIND } from '../src/host/constants.js' + +function turnsHost() { + let listener + const delivered = [] + const rt = { + ctx: { on: (name, fn) => { if (name === 'session/event') listener = fn } }, + state: { exchangeTurns: {}, questions: {} }, + bots: { a: { id: 'a', name: 'Chief' }, b: { id: 'b', name: 'Writer' } }, + isBotSession: () => true, + roomOf: () => undefined, + groupOwners: new Map(), + chatOwners: new Map(), + botOf: id => rt.bots[id], + selfOf: id => id, + save: async () => {}, + deliver: async (args) => { delivered.push(args) }, + noteFailure() {}, + warn() {}, + switchJobs: () => [], + } + installTurns(rt) + let seq = 0 + const emit = (id, type, data) => listener({ id }, { type, seq: ++seq, data }) + return { emit, delivered } +} + +const replySource = (extra = {}) => ({ kind: SOURCE_KIND, role: 'request', senderSessionId: 'a', senderName: 'Chief', exchangeId: 'e1', hop: 0, ...extra }) +const textBlock = content => [{ type: 'text', text: content }] +const assistantText = text => ({ message: { content: textBlock(text) } }) +const joined = host => { + host.emit('b', 'turn/start', { turn: 1 }) + host.emit('b', 'user/message', { content: textBlock('write the letter'), source: { kind: 'user' } }) + host.emit('b', 'assistant/message', assistantText('first half')) + host.emit('b', 'user/message', { content: textBlock('[Message from Chief]\nmore'), source: replySource() }) +} + +test('#2677: the failure notice carries the structural marker; a normal reply does not', () => { + const failed = turnsHost() + joined(failed) + failed.emit('b', 'turn/end', { turn: 1, reason: { kind: 'error', error: { code: 'SERVER', message: 'upstream busy', status: 503 } } }) + assert.equal(failed.delivered.length, 1) + assert.equal(failed.delivered[0].failure, true) + + const answered = turnsHost() + joined(answered) + answered.emit('b', 'turn/end', { turn: 1, reason: { kind: 'completed' } }) + assert.equal(answered.delivered.length, 1) + assert.equal(answered.delivered[0].failure, undefined) +}) diff --git a/docs/subsystems/core-host.md b/docs/subsystems/core-host.md index 620c6afcf..de2b1dd2f 100644 --- a/docs/subsystems/core-host.md +++ b/docs/subsystems/core-host.md @@ -440,7 +440,7 @@ Work a member hands a Bot with `message_bot` from its group Session runs in a hi ### Untrusted fences in every shared room - `authoredRoomLine` fences peer-controlled text by the line's origin, not the room kind (#98): a remote participant, a local remote-chat record (`state.bots[id].remote`), a botId that is no team Bot, or `line.remote`. A friend's Bot speaking in a non-face group is fenced the same way as in a face room. A complete wrap is left as-is so outbound fanout does not nest residue; a mention of the marker in the body still fences. `stripUntrusted` is one fixed-point pass over both `untrusted-message` and `untrusted-file`, and its pattern allows whitespace inside the tag (`< / untrusted-message>`), since a parser that ignores whitespace reads the same marker; `stripMarkers` does the same for one tag. `core/test/fence-property.test.js` fuzzes a corpus of reconstitution attempts to keep this a fixed point. Inbound-share lines from the other instance are `remote: true` so owner-typed text is fenced too. Face join/leave notices stay unfenced platform copy. Channel room previews use `unwrapUntrusted` so the sidebar never shows the wrap. A human line whose `who` is `User` is rewritten from the signed-in display name before it is stored or fenced. -- A late group model failure posts a short friendly line into the room (`This Bot could not answer right now.` / `This Bot has no model set up yet.`); `failureLine` stays on Bot-to-Bot deliver and in diagnostics, not in the human transcript. +- A late group model failure posts a short friendly line into the room (`This Bot could not answer right now.` / `This Bot has no model set up yet.`); `failureLine` stays on Bot-to-Bot deliver and in diagnostics, not in the human transcript. That Bot-to-Bot failure notice carries `failure: true` (#2677), the same field the remote notice uses, so a consumer reads the outcome from the message instead of parsing the sentence. ### Remote chats and mentions diff --git a/docs/subsystems/workflow.md b/docs/subsystems/workflow.md index 8969952ea..3cca99c3e 100644 --- a/docs/subsystems/workflow.md +++ b/docs/subsystems/workflow.md @@ -59,7 +59,7 @@ Everything in `plugins/workflow/`: the graph rules, the run store, the matcher, 1. Friends' Bots are not nodes. A `remote` Bot is refused by create, define and the matcher. 2. The canvas and the run share one graph: nodes, edges and positions. A drag commits `workflow/graph`, but only while **Edit** is on — a factory preset is copied first (#1378). A wire or a reconnect that would list a pair the graph already carries commits nothing new, because `normalizeGraph` refuses a graph that lists one pair twice and then saves none of it: unguarded, the plane would show no edge, no change and only a save error (#2163). Status highlighting reads that same graph. Replacing edges does not recompute who is already running. The Settings canvas paints that graph on the same 960px design plane, scaled to the width the settings page gives it, so the canvas is not a second scroll container inside the page's own scroll; its pointer deltas are divided by the scale, and its wires keep a stronger stroke than a plain divider so an edge can be followed at that size (#1483). While another dsh holds the Bot team the host is read-only (`roster.readOnly`): the page still reads through `workflow/state`, keeps its controls disabled, and offers no **Edit**, **Add link** or wire port, so it cannot take a change that core would refuse with `bot/read-only` and then drop (#2187). The same window draws the sidebar's workflow list without **New**, the row's **Rename**/**Delete** or **Hand a task to a Workflow**: the list is read-only too, and an `roster.readOnly` that is not a string, or holds only whitespace, counts as writable — trimmed the way `core/src/client/workflows.js:570` trims it (#2200). The Stage a row opens carries the same lock: it prints the host's sentence under `role="status"` and draws no **Hand a task to a Workflow**, no **Re-match**, no **Pause**/**Resume**/**Stop**, no **Edit**/drag palette and no **Reset**, while the row click, **Progress**, **Flow** and viewing a run still work; a writable roster draws every control exactly as before (#2213). Mounting the list in a read-only window still reads `projects` so the Project filter fills, but it sends no `workflow/ingest` for the open cards: ingest is a claim/match write the host refuses, so a reader opening the list writes nothing; a writable window still ingests every open card (#2222). The read-only sentence can also land after the list has mounted — a cold window starts on the empty roster, `ready: false` and no `readOnly` — so the ingest step asks the value the latest render saw rather than the one the mount saw, and it skips the board read and stops between two cards once the window has turned read-only. The Settings page draws no read-only sentence of its own: the host's Settings header already prints one for every page it shows (`core/src/client/settings.js` `bt-read-only`, `role="status"`, `hostText(roster.readOnly)`), and a second copy inside the page said the same thing twice (#2202). -3. **A live run advances only when its ephemeral Bot replies, following the graph's edges in order.** That reply advances the run and may appear on the workflow stage log; it is not delivered into a persistent sidebar Bot conversation (for example Chief). The `workflow-step` send middleware denies those replies after `stepResult` runs. Nodes start only when their forward predecessors finish and edges reach them. **Early termination (#1537, #1538, #1546, #1570, #1571):** When a node completes **or fails**, the run ends immediately with status `done`, sets `run.earlyEnd` to the condition name, and does not advance to successors, take a success-path `loop`, or burn `onFail` recovery attempts. The two markers that report the task itself end it only on a structured signal (#1546): `needs-human` and `already-claimed` must be a declared OUTPUT whose value **equals** the marker, or a whole reply line that is exactly `EARLY-END ` in any case. A step that quotes an issue title, negates the label or reports it absent keeps running. `tip-unchanged` (#1570) and `open-pr` (#1571) accept that same structured signal — in a declared OUTPUT the value counts as a whole word, so a brief that names `data-open-project` is not `open-pr` (#1866) — and also honor their prose fallbacks (`tip unchanged` / `unchanged tip`, `open PR already` / `already has an open PR` / `has an open PR` / `found an open PR`), so a Bot that forgets the structured marker cannot soft-lock Claim↔Verify on the same tip. See `checkEarlyEnd` in Key files. Factory prompt guidance lives in `EARLY_END_GUIDANCE` (`plugins/workflow/src/template.js`) for Settings graphs that wire Triage / Claim / Verify: Verify must escalate once on an unchanged tip, not re-run the full suite. There is no scripted or seeded run path: `workflow/state` exposes only `live === true` runs — the running ones, plus the last run when it is still `live` and ended `failed` — a failed sibling stays on the stage beside whatever is running now, with `run.failure` (`{ code, at, nodeName?, text? }`) naming what ended it, so the room does not have to open Runs to learn that a run failed (#1493, #1494). The failed sibling is left out when a running run carries the same task id: that is a retry, and the live run keeps the row (#1507). History entries carry the run `id` and the same `failure`, so a Runs row shows when the run ended and why, and a row for a run already on the stage is not listed twice. A run does not create a schedule. `scheduleLabel` is display text. When the host can create Bots, starting a live run (`assign`, `run_workflow`, `ingest`) spawns one ephemeral Bot per node from that node's prompt, role, model and tools, sends it the step from the Main Bot, and removes every instance when the run finishes, fails for good, or is reset. A `failOnce` retry keeps them. They are hidden and left out of the sidebar. Loading the plugin does not spawn Bots for the default workflows. Plugin unload (`dispose`) and the load-time `reap` both end a live run still marked `running`: in-flight nodes are failed, the run is archived as `failed`, and any ephemeral Bot ids are released. A **read-only** host — a second window on the same `DSH_HOME`, the reader `home-lock` describes — does neither: it never ran those runs, so it cannot tell which of them a dead host left behind, and it must not end them, release their step Bots or write its older copy over the writer's store; both calls return without touching the record and the writer's own `reap` decides (#2673). A run already written as `running` with no ephemeral map (an older unload) is healed the same way on the next load (#1414). Loading is also the one moment a run the host left behind is noticed when its ephemeral map is still present: `reap` releases those instances and records the run as `failed`. Nothing resumes it. A run that already ended keeps its own status (#1388). `workflow/reset` stopping one live run ends it the same way before its Bots are released: nodes still `running` become `failed`, so the strip cannot keep painting a run that is over (#1424). A graph edit that drops a node that is still running ends that run the way Stop does: the node is failed, the run is archived `failed`, its node Bots are released, and the queue drains (#1549); an id whose release failed stays named on that archived run, and the next load's `reap` retries it, instead of the edit forgetting it (the #1616 contract). Every wait is bounded (#1530). A dispatch that does not land is not marked handed: the chat names the code and the attempt, and it is retried after a short backoff at most twice, never twice inside one hand-off, so a 429 is not hammered. When it still does not land, the step and the run fail with the reason written into the chat. Each running live step also holds a deadline — its workflow's `stepTimeoutMinutes`, `DEFAULT_STEP_TIMEOUT_MINUTES` (45 minutes) unless the record sets one, an integer from 1 to 120 given at create, through `define_workflow`, or through `workflow/step-timeout` (armed even when the host cannot spawn ephemeral Bots; #1530, #1580); when it expires the node and the run fail as `failed`, the run's ephemeral Bots are released, the chat and `run.failure` say the step's deadline was reached (`step deadline reached (45 min)`, the number its lane set), and `warn` names it once; the line never reads as a reply that did not arrive, because a step still working is killed by the same timer (#1868). The deadline is cleared as soon as the step moves, and by Pause, Reset, delete, `reap`, `dispose` and a Stop that released every Bot; a run the owner paused holds none, and Resume arms it again from that moment. A timer never outlives the run that started it. A reply that matches no running step — its run ended, or the node already moved — is reported through `warn` once per Bot and run instead of being dropped silently. A reply that says `could not answer your message` or `model call failed` fails that node and does not start its successor. A 429 or rate-limit mention in any other reply does not. Stop and Reset (`workflow/reset`) follow the same order as `dispose`: they release the run's Bots before the record forgets their ids, and an id whose release failed stays named on the still-live run, so the next Stop, Reset, dispose or load-time `reap` retries it (#1476). A run left that way is marked `paused` and `releasePending` and shows as Paused: it hands off no further node, `resume` is refused with an error, and Stop again is the retry. It also keeps the deadline of the node it was on — the one paused state that holds one — and a graph save re-syncs timers without taking it away (a save that cleared it left the run holding its slot with nothing left to free it); the deadline releases again there, ends the run as `paused-expired` and frees the slot, re-arming itself while a Bot still will not go (#2453, P-560). A whole-workflow Reset that could not release every Bot freezes every run it did not reset, because each of them has lost the ids whose release did go. The markers go away when the run ends: Stop again, `dispose`, `reap`, or its own completion. +3. **A live run advances only when its ephemeral Bot replies, following the graph's edges in order.** That reply advances the run and may appear on the workflow stage log; it is not delivered into a persistent sidebar Bot conversation (for example Chief). The `workflow-step` send middleware denies those replies after `stepResult` runs. Nodes start only when their forward predecessors finish and edges reach them. **Early termination (#1537, #1538, #1546, #1570, #1571):** When a node completes **or fails**, the run ends immediately with status `done`, sets `run.earlyEnd` to the condition name, and does not advance to successors, take a success-path `loop`, or burn `onFail` recovery attempts. The two markers that report the task itself end it only on a structured signal (#1546): `needs-human` and `already-claimed` must be a declared OUTPUT whose value **equals** the marker, or a whole reply line that is exactly `EARLY-END ` in any case. A step that quotes an issue title, negates the label or reports it absent keeps running. `tip-unchanged` (#1570) and `open-pr` (#1571) accept that same structured signal — in a declared OUTPUT the value counts as a whole word, so a brief that names `data-open-project` is not `open-pr` (#1866) — and also honor their prose fallbacks (`tip unchanged` / `unchanged tip`, `open PR already` / `already has an open PR` / `has an open PR` / `found an open PR`), so a Bot that forgets the structured marker cannot soft-lock Claim↔Verify on the same tip. See `checkEarlyEnd` in Key files. Factory prompt guidance lives in `EARLY_END_GUIDANCE` (`plugins/workflow/src/template.js`) for Settings graphs that wire Triage / Claim / Verify: Verify must escalate once on an unchanged tip, not re-run the full suite. There is no scripted or seeded run path: `workflow/state` exposes only `live === true` runs — the running ones, plus the last run when it is still `live` and ended `failed` — a failed sibling stays on the stage beside whatever is running now, with `run.failure` (`{ code, at, nodeName?, text? }`) naming what ended it, so the room does not have to open Runs to learn that a run failed (#1493, #1494). The failed sibling is left out when a running run carries the same task id: that is a retry, and the live run keeps the row (#1507). History entries carry the run `id` and the same `failure`, so a Runs row shows when the run ended and why, and a row for a run already on the stage is not listed twice. A run does not create a schedule. `scheduleLabel` is display text. When the host can create Bots, starting a live run (`assign`, `run_workflow`, `ingest`) spawns one ephemeral Bot per node from that node's prompt, role, model and tools, sends it the step from the Main Bot, and removes every instance when the run finishes, fails for good, or is reset. A `failOnce` retry keeps them. They are hidden and left out of the sidebar. Loading the plugin does not spawn Bots for the default workflows. Plugin unload (`dispose`) and the load-time `reap` both end a live run still marked `running`: in-flight nodes are failed, the run is archived as `failed`, and any ephemeral Bot ids are released. A **read-only** host — a second window on the same `DSH_HOME`, the reader `home-lock` describes — does neither: it never ran those runs, so it cannot tell which of them a dead host left behind, and it must not end them, release their step Bots or write its older copy over the writer's store; both calls return without touching the record and the writer's own `reap` decides (#2673). A run already written as `running` with no ephemeral map (an older unload) is healed the same way on the next load (#1414). Loading is also the one moment a run the host left behind is noticed when its ephemeral map is still present: `reap` releases those instances and records the run as `failed`. Nothing resumes it. A run that already ended keeps its own status (#1388). `workflow/reset` stopping one live run ends it the same way before its Bots are released: nodes still `running` become `failed`, so the strip cannot keep painting a run that is over (#1424). A graph edit that drops a node that is still running ends that run the way Stop does: the node is failed, the run is archived `failed`, its node Bots are released, and the queue drains (#1549); an id whose release failed stays named on that archived run, and the next load's `reap` retries it, instead of the edit forgetting it (the #1616 contract). Every wait is bounded (#1530). A dispatch that does not land is not marked handed: the chat names the code and the attempt, and it is retried after a short backoff at most twice, never twice inside one hand-off, so a 429 is not hammered. When it still does not land, the step and the run fail with the reason written into the chat. Each running live step also holds a deadline — its workflow's `stepTimeoutMinutes`, `DEFAULT_STEP_TIMEOUT_MINUTES` (45 minutes) unless the record sets one, an integer from 1 to 120 given at create, through `define_workflow`, or through `workflow/step-timeout` (armed even when the host cannot spawn ephemeral Bots; #1530, #1580); when it expires the node and the run fail as `failed`, the run's ephemeral Bots are released, the chat and `run.failure` say the step's deadline was reached (`step deadline reached (45 min)`, the number its lane set), and `warn` names it once; the line never reads as a reply that did not arrive, because a step still working is killed by the same timer (#1868). The deadline is cleared as soon as the step moves, and by Pause, Reset, delete, `reap`, `dispose` and a Stop that released every Bot; a run the owner paused holds none, and Resume arms it again from that moment. A timer never outlives the run that started it. A reply that matches no running step — its run ended, or the node already moved — is reported through `warn` once per Bot and run instead of being dropped silently. A reply fails that node on a structured signal only: the host's failure notice carries `failure: true`, which the middleware hands `stepResult` as `ok: false` (a caller-set `ok: false` counts too), so an answer that quotes `could not answer your message` or `model call failed` while auditing it advances the step (#2677). An old host without the marker falls back to the notice shape at the start of the text only (` could not answer your message: …`), never a free substring, and a 429 or rate-limit mention in any other reply does not fail a step. Stop and Reset (`workflow/reset`) follow the same order as `dispose`: they release the run's Bots before the record forgets their ids, and an id whose release failed stays named on the still-live run, so the next Stop, Reset, dispose or load-time `reap` retries it (#1476). A run left that way is marked `paused` and `releasePending` and shows as Paused: it hands off no further node, `resume` is refused with an error, and Stop again is the retry. It also keeps the deadline of the node it was on — the one paused state that holds one — and a graph save re-syncs timers without taking it away (a save that cleared it left the run holding its slot with nothing left to free it); the deadline releases again there, ends the run as `paused-expired` and frees the slot, re-arming itself while a Bot still will not go (#2453, P-560). A whole-workflow Reset that could not release every Bot freezes every run it did not reset, because each of them has lost the ids whose release did go. The markers go away when the run ends: Stop again, `dispose`, `reap`, or its own completion. 4. Each workflow keeps a FIFO queue and up to `maxConcurrent` live runs (default 2, an integer from 1 to 8 on the record, via `workflow/concurrency`, or in Settings → Workflow graphs). A task that does not fit is enqueued. When a run finishes, every workflow with a free slot is drained, oldest enqueue first. A new assign on a workflow that already has a queue waits behind it. A task does not replace a running run or release that run's Bots. Migrating an old store keeps only runs that were still `running` with ephemeral Bot ids; other legacy scripted rows are dropped from active state (finished tasks may remain in lane history). History records a finished or failed live run. Lowering the cap does not stop a run that has already started. Assigning `{ botId }` claims the card for that local Bot and does not enqueue. A remote Bot is refused. `activeCount` is `runningOf(lane)` — live runs with `live === true` and `status === 'running'`. The stage aggregate chip and the Runs tab (运行记录) must agree: every such run appears as its own running row and progress card with its own `status` (never a hardcoded Running), the chip count does not exceed that list, and a failed sibling appears with its time and reason. On the Runs list, the clock stays in a non-shrinking column with a gap before the status chip, and a Failed row paints the Failed chip once — a generic Failed with no host detail does not also draw `wf-hist-reason` (#1618). 5. A node declares up to eight inputs and eight outputs. A required input is an output of the same name on an ancestor, or `task`, `title`, or `brief` on a root. A start with no provider is refused, and Bots spawned for that attempt are released. On a Bot-backed run a successor stays queued until its required inputs are present. If every forward predecessor is done, failed or skipped and an input is still missing, that node fails and the run stops. The dispatch text carries `Input : `. It also lists the OUTPUT lines the step owes — its declared ports plus every output its outgoing edges read (`edgeOutputs` in `src/graph.js`), spelled `OUTPUT : ` when an edge compares that output against a value — so a node that declares no outputs is still told which line its own gate reads (#2669). The Bot reports `OUTPUT : `; a single declared output may be the reply body. A declared output a gate reads is required (§3): when a `when`/`else` or `until` edge reads it and the step's reply does not produce it, the step fails with `failure.text` `Missing OUTPUT ` instead of settling, so the gate never reads silence as "not this branch" and the run cannot close `done` on it; a delivered value that does not equal `eq` still completes the step and skips the untaken branch (#1773). One exception keeps the work a step already did (#2583, #2658): when `state` is the only gate-read output missing and the step's own reply shows its result — a pull request link, `pull/`, a line saying it opened PR #n, a `Claimed-by:` / `Claimed #n` line, or its own declared OUTPUT names the delivery (`pr`/`pull` with a pull number such as `2657 (verified)`, or `merged`/`merge-commit` carrying a 40-character commit sha; any other output name counts for nothing) — the step stands as `done`, the run takes the node's first declared `state` branch (on the fast line `state: ready → claim`, so the pull request reaches gate/enqueue instead of the `else` route the step never chose), and the run record carries `outputMissing: true` and `producedPull: ` when a pull request was named; every other missing verdict still fails, and a step whose reply shows no such result still fails the same way. `stepArtifacts` reads that evidence and `firstBranchValue` names the branch, both in `src/graph.js`; the engine has no GitHub client, so the reply is the only evidence it can read. Nodes start in edge order. 6. Each node is a real ephemeral Bot for that run, then it is removed. A permanent sidebar Bot (`botId`) is optional and is not required for the step to run. **File and shell rights are the same as a normal sidebar Bot's:** every node Bot gets the `danger-full-access` preset, and `EPHEMERAL_DENIED_TOOLS` still refuses the team-admin tools at the guard and approval layers (#1466, 29 names). **`node.tools` selects connectors only:** `spawn` passes the node's cleaned list to `fb.bot.create`, the host records it as `stepTools`, and `connector-access.js` allows `mcp__github__*` for `github` and `mcp__composio__*` for `composio` — a connector the node did not name is refused, and a node that names none reaches none. The list is also stored in `fields.tools` for the UI. A hidden run Bot cannot show an approval card, so the approval seam answers the same way the guard does. The host stores this on the run Bot as `stepTools` and `permission`. diff --git a/patches/ds-bot/0001-core-fixes.patch b/patches/ds-bot/0001-core-fixes.patch index 175f7e225..af05b5bf2 100644 --- a/patches/ds-bot/0001-core-fixes.patch +++ b/patches/ds-bot/0001-core-fixes.patch @@ -8505,11 +8505,13 @@ diff --git a/src/host/turns.js b/src/host/turns.js if (!expectsReply(source) || answersUser(source)) return const name = rt.botOf(rt.selfOf(id))?.name ?? 'The other Bot' // A sender that wrote from a group gets the reply in its Session there, while -@@ -434,6 +434,7 @@ export function install(rt) { +@@ -434,6 +434,9 @@ export function install(rt) { role: 'reply', hop: (source.hop ?? 0) + 1, replyTo: source.exchangeId, + origin, ++ // #2677: the same marker the remote notice uses, so a reader never parses the sentence. ++ failure: true, }).catch(error => rt.warn('ds-bot: failure notice delivery failed: %s', error)) } else if (reply !== '' || reason?.kind === 'aborted') { // A turn stopped before it could answer still owes the sender a line, diff --git a/patches/ds-bot/files/test/failure-notice-marker.test.js b/patches/ds-bot/files/test/failure-notice-marker.test.js new file mode 100644 index 000000000..227bb5e41 --- /dev/null +++ b/patches/ds-bot/files/test/failure-notice-marker.test.js @@ -0,0 +1,56 @@ +// #2677: a workflow step used to be failed whenever its own answer quoted the failure +// sentence. The host's notice now says structurally that the turn failed, so a reader does +// not have to parse the text it may be auditing. +import assert from 'node:assert/strict' +import { test } from 'node:test' +import { install as installTurns } from '../src/host/turns.js' +import { SOURCE_KIND } from '../src/host/constants.js' + +function turnsHost() { + let listener + const delivered = [] + const rt = { + ctx: { on: (name, fn) => { if (name === 'session/event') listener = fn } }, + state: { exchangeTurns: {}, questions: {} }, + bots: { a: { id: 'a', name: 'Chief' }, b: { id: 'b', name: 'Writer' } }, + isBotSession: () => true, + roomOf: () => undefined, + groupOwners: new Map(), + chatOwners: new Map(), + botOf: id => rt.bots[id], + selfOf: id => id, + save: async () => {}, + deliver: async (args) => { delivered.push(args) }, + noteFailure() {}, + warn() {}, + switchJobs: () => [], + } + installTurns(rt) + let seq = 0 + const emit = (id, type, data) => listener({ id }, { type, seq: ++seq, data }) + return { emit, delivered } +} + +const replySource = (extra = {}) => ({ kind: SOURCE_KIND, role: 'request', senderSessionId: 'a', senderName: 'Chief', exchangeId: 'e1', hop: 0, ...extra }) +const textBlock = content => [{ type: 'text', text: content }] +const assistantText = text => ({ message: { content: textBlock(text) } }) +const joined = host => { + host.emit('b', 'turn/start', { turn: 1 }) + host.emit('b', 'user/message', { content: textBlock('write the letter'), source: { kind: 'user' } }) + host.emit('b', 'assistant/message', assistantText('first half')) + host.emit('b', 'user/message', { content: textBlock('[Message from Chief]\nmore'), source: replySource() }) +} + +test('#2677: the failure notice carries the structural marker; a normal reply does not', () => { + const failed = turnsHost() + joined(failed) + failed.emit('b', 'turn/end', { turn: 1, reason: { kind: 'error', error: { code: 'SERVER', message: 'upstream busy', status: 503 } } }) + assert.equal(failed.delivered.length, 1) + assert.equal(failed.delivered[0].failure, true) + + const answered = turnsHost() + joined(answered) + answered.emit('b', 'turn/end', { turn: 1, reason: { kind: 'completed' } }) + assert.equal(answered.delivered.length, 1) + assert.equal(answered.delivered[0].failure, undefined) +}) diff --git a/plugins/workflow/index.js b/plugins/workflow/index.js index 77d76d719..2552a9a4c 100644 --- a/plugins/workflow/index.js +++ b/plugins/workflow/index.js @@ -15,6 +15,25 @@ function runnerName(base, taken) { return `${stem} ${randomBytes(3).toString('hex')}`.slice(0, 40) } +// #2677: a step's answer is its own text, so quoting the host's failure sentence in an +// audit or a pasted log must not fail the node. The host marks a real notice on the +// message; an old host without that marker is read only in the notice shape at the start +// of the text (` could not answer your message: …`). +const LEGACY_FAILURE_NOTICE = [ + /^[^\n:`"'<>]{1,80} could not answer your message:/i, + /^could not answer your message$/i, + /^[^\n:`"'<>]{0,80}model call failed\.?$/i, +] +const legacyFailureNotice = text => LEGACY_FAILURE_NOTICE.some(pattern => pattern.test(text)) +// `rt.deliver` hands the middleware its own args under `deliverArgs`, where the notice +// carries the marker; a direct `bot.send` message carries it at the top level. +function failureMarker(msg) { + const marked = msg.failure ?? msg.deliverArgs?.failure + if (marked !== undefined) return marked === true + if (msg.ok !== undefined) return msg.ok === false + return undefined +} + export const name = 'flybot-workflow' export const inject = ['flybot'] @@ -168,7 +187,8 @@ export function apply(ctx, config = {}) { const owned = service.ownsBot(from) if (owned !== true) return undefined const text = String(msg.text ?? '') - const failed = /could not answer your message|model call failed/i.test(text) + const marked = failureMarker(msg) + const failed = marked === undefined ? legacyFailureNotice(text) : marked === true // The reply advances the run only. It must not land in a persistent Bot chat (e.g. Chief). try { await Promise.resolve(service.stepResult({ botId: from, text, ok: !failed })) diff --git a/plugins/workflow/test/step-body.test.js b/plugins/workflow/test/step-body.test.js index 5c42340d2..4e77625d4 100644 --- a/plugins/workflow/test/step-body.test.js +++ b/plugins/workflow/test/step-body.test.js @@ -116,3 +116,66 @@ test('#2669: the run walks the branch once the node prints the OUTPUT it was tol assert.equal(run.nodeStatus.escalate, 'skipped', 'the else route is not taken') assert.equal(run.status, 'running') }) + +// #2677: a step's answer is its own text, so an audit that quotes the host's failure +// sentence, a pasted log, or a retold failure must not fail the node. The host marks the +// real notice on the message; only an old host falls back to the sentence, and only when +// it is the shape at the start of the text. +function replyMiddleware() { + let handle + const seen = [] + const fb = { + bot: { + list: () => [], + send: Object.assign(async () => ({ action: 'allow' }), { + use(spec) { + handle = spec.handle + return () => {} + }, + priority: { ROUTING: 10 }, + }), + }, + readOnly: () => false, + api: { handle: () => () => {} }, + tools: { register: () => () => {} }, + warn: () => {}, + } + const started = apply({ flybot: { scope: () => fb }, effect(fn) { fn() } }, { persist: false }) + started.service.ownsBot = () => true + started.service.stepResult = async body => { seen.push(body); return {} } + return { handle, seen } +} + +async function stepOk(message) { + const { handle, seen } = replyMiddleware() + const verdict = await handle({ from: 'bot-step', role: 'reply', text: '', ...message }) + assert.equal(verdict?.action, 'deny') + assert.equal(seen.length, 1) + return seen[0].ok +} + +const QUOTED_FAILURE = '候选 3：turns.js:709 的正文是 ` could not answer your message: its model call failed (…)`' + +test('#2677: an answer that quotes the failure sentence still advances the step', async () => { + assert.equal(await stepOk({ text: QUOTED_FAILURE }), true, 'a quoted failure inside the answer is not the step failing') + assert.equal(await stepOk({ text: '查完了，3 条候选。' }), true, 'an ordinary answer advances the step') +}) + +test('#2677: the host marks a real failure notice, and the marker decides', async () => { + assert.equal(await stepOk({ text: 'Scout abc could not answer your message: its model call failed (RATE_LIMIT 429).', failure: true }), false) + assert.equal(await stepOk({ text: 'anything at all', ok: false }), false) + // `rt.deliver` hands the middleware its own args under `deliverArgs`, which is where + // the host's notice carries the marker on a live run. + assert.equal(await stepOk({ text: 'Scout abc could not answer your message: its model call failed (RATE_LIMIT 429).', deliverArgs: { failure: true } }), false) + // A marker that says the answer is fine wins over text that looks like the notice. + assert.equal(await stepOk({ text: 'Scout abc could not answer your message: its model call failed (RATE_LIMIT 429).', failure: false }), true) +}) + +test('#2677: an old host without the marker is read only at the start of the text', async () => { + assert.equal(await stepOk({ text: 'Scout abc could not answer your message:' }), false, 'the notice shape at the start still fails') + assert.equal(await stepOk({ text: 'Scout abc could not answer your message: its model call failed (RATE_LIMIT 429).' }), false) + // These two are what the middleware caught before #2677; the anchored fallback keeps them. + assert.equal(await stepOk({ text: 'could not answer your message' }), false) + assert.equal(await stepOk({ text: 'The model call failed.' }), false) + assert.equal(await stepOk({ text: 'The log said "` could not answer your message: its model call failed`".' }), true, 'a quoted notice after the start is not the step failing') +})