[sroussey]

Summary

Implements text-retract stream events to allow producers (like AgentTask) to take back text they streamed before abandoning an attempt. When a model call fails partway and is retried, the failed attempt's text is retracted so consumers see only the successful retry's output, as if the failed attempt never ran.

Key Changes

• New StreamTextRetract event type (packages/task-graph/src/task/StreamTypes.ts):

  • Removes the last length characters (UTF-16 code units) from a port's accumulated text
  • Clamped at the start; later deltas append after what remains
  • retractionLength() is the one rule for how much a retraction removes (a fraction rounds down; anything not positive, NaN included, removes nothing), and retractText() is the fold built on it

• StreamProcessor retraction handling (packages/task-graph/src/task/StreamProcessor.ts):

  • Truncates accumulated text on retraction
  • Fails cache sinks that already teed retracted bytes (cannot be taken back), but only for a retraction that actually removes text
  • Detaches failed ports so later deltas go to accumulator only
  • Refuses to settle sunk ports without accumulator unless finish payload states the value

• AgentTask retry recovery (packages/ai/src/task/AgentRoundAttempt.ts, where main's refactor moved the round loop):

  • Forwards text and reasoning deltas live as they stream
  • On attempt failure, emits text-retract for exactly what that attempt sent
  • Counts what each port sent through retractionLength, so retraction lengths stay exact
  • Handles separator between reasoning rounds (retracted when all reasoning is taken back)
  • Abandons the in-flight model call if the stream stops being read before the attempt settles

• Host round runners can retract (packages/ai/src/task/AgentRoundRunner.ts):

  • AgentRoundContext.onTextRetract(port, length) lets an AGENT_ROUND_RUNNER withdraw text its own call streamed; the turn caps it at what that attempt forwarded

• StreamPump edge forwarding (packages/task-graph/src/task-graph/StreamPump.ts):

  • Retractions travel with their port's deltas on dataflow edges
  • Port-filtered edges carry retractions only for their own port

• UI and CLI updates:

  • streamedText.ts: Added classifyTextRetract() (rounds the length like retractionLength) and retraction folding
  • AgentTaskTurnControls.test.ts: New tests for retraction on failed/timed-out attempts
  • AgentTaskReasoningAndImages.test.ts: Tests for multi-port retraction
  • AgentRoundAttempt.test.ts: the retry test now expects the partial to be forwarded, then retracted; a new test stops reading mid-stream and expects the call to be aborted; another passes a host runner's retraction through, capped
  • CLI chat, eval and web UI: Display retraction events and fold them into accumulated text

• Comprehensive test coverage (StreamTextRetract.test.ts):

  • Processor truncation and cache sink failure, including a retraction that removes nothing
  • Port-filtered edge forwarding
  • Accumulator and event folding behavior, including fractional and NaN lengths

Implementation Details

• Retractions are counted per port to handle multi-port streams (text + reasoning)
• A port retracted to empty settles like one that never streamed: absent unless the producer states it in finish (AgentTask always does)
• Separators between reasoning rounds are included in retraction when all reasoning is taken back
• Every consumer folds with retractText(acc, length), which removes retractionLength(length) characters

🤖 Generated with Claude Code

https://claude.ai/code/session_01Hti4mgm8Y8uukHyEnzhJLo

[mergestorm-vortex]

Review progress ██████████ 28/28 files

Approve — reviewed files look good at 6141d32.

✅ All clear — nothing to fix.

---

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Comment — found 1 issue(s) at a6601a3.

Actionable comment posted: 1

🤖 Prompt for AI agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

Findings to address:
1. In `@packages/ai/src/task/AgentRoundAttempt.ts` (line 226, Important):
   `Math.max(0, NaN)` and the resulting `Math.min(...)` are `NaN`. Unlike the shared `retractText()` fold (which treats `NaN` as a no-op), this makes `textSent` become `NaN`; no retraction is yielded here, and the failure cleanup's `textSent > 0` check is false. If a child stream emits a `text-retract` with `length: NaN` after sending text and then fails, the failed attempt's text remains in the outer stream alongside the retry. Preserve the fold's no-op semantics or reject non-finite lengths before updating the counter.

---

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Approve — reviewed files look good at 82daa71.

✅ All clear — nothing to fix.

---

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Comment — found 1 issue(s) at 5578b44.

Actionable comment posted: 1

🤖 Prompt for AI agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

Findings to address:
1. In `@packages/ai/src/task/AgentRoundAttempt.ts` (line 256, Important):
   `yield event`
   If a consumer calls `return()`/breaks from `AgentTask.executeStream()` while this yield is suspended, `attemptRound` exits without aborting `roundAbort`/the owned task, clearing `timer`, or unsubscribing `off`. The in-flight model request can therefore keep running (and continue billing) after the stream has been abandoned; with no `timeoutMs`, the task and listener can remain indefinitely. Add cleanup for the generator-return path, including aborting the active attempt.

---

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Comment — found 1 issue(s) at 61d93e6.

Actionable comment posted: 1

🤖 Prompt for AI agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

Findings to address:
1. In `@packages/ai/src/task/AgentRoundAttempt.ts` (line 232, Important):
   `Math.min(event.length, textSent)` subtracts the raw fractional length, but `retractText` folds it through `slice`, which removes an integer number of code units. For example, with earlier-round text `P` and this attempt's `abc`, a child retraction of `1.5` removes `bc` from the outer stream while leaving `textSent === 1.5`; if the attempt then fails, the cleanup retraction of `1.5` also removes the earlier `P`. Normalize the effective length to the same amount the fold removes (or reject non-integer lengths) before updating these counters; apply the same treatment to the reasoning branch.

---

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Comment — found 1 issue(s) at d0d3597.

Actionable comment posted: 1

🤖 Prompt for AI agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

Findings to address:
1. In `@packages/task-graph/src/task/StreamTypes.ts` (line 37, Important):
   `acc.slice(0, Math.max(0, acc.length - length))` does not implement the newly documented round-down rule for fractional `length`. For example, folding `"Pabc"` with `length: 1.5` using this expression yields `"Pa"`, while `retractText`/`retractionLength` yields `"Pab"`. A consumer following this public example will diverge from `StreamProcessor` and the agent retry fold; document the canonical helper or normalize `length` in the expression.

---

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Request changes — found 1 issue(s) at 58ce47e.

Actionable comment posted: 1

🤖 Prompt for AI agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

Findings to address:
1. In `@packages/task-graph/src/task/StreamTypes.ts` (line 57, Blocker):
   `return length > 0 ? Math.floor(length) : 0;`
   This makes a `length` such as `0.5` a no-op, but `StreamProcessor` still tests the raw `event.length > 0` before detaching a cache router. With `noAccumulation` enabled, an append stream that emits text, then a `text-retract` of `0.5`, then `finish` without that port is therefore detached despite retracting nothing; the finish handler then throws because the detached port has no accumulator or stated value. Use the normalized retraction length for that sink decision too.

---

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Request changes — found 2 issue(s) at 0de83cd.

Actionable comments posted: 2

🤖 Prompt for AI agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

Findings to address:
1. In `@packages/task-graph/src/task/StreamProcessor.ts` (line 310, Blocker):
   `retractText` can leave this accumulator entry as `""` when a retraction removes the entire port. The finish merge later skips every empty accumulated value (`if (text.length > 0)`), so a stream like `text-delta("partial"), text-retract(99), finish({})` completes with no `text` property instead of `text: ""`; downstream consumers of this declared output port then receive an absent value despite accumulation being enabled.

Additional findings outside changed lines:
1. In `@packages/task-graph/src/task/StreamTypes.ts` (line 57, Blocker):
   A sub-unit retraction can fail a cached no-accumulation run
   `return length > 0 ? Math.floor(length) : 0;`
   This makes a `length` such as `0.5` a no-op, but `StreamProcessor` still tests the raw `event.length > 0` before detaching a cache router. With `noAccumulation` enabled, an append stream that emits text, then a `text-retract` of `0.5`, then `finish` without that port is therefore detached despite retracting nothing; the finish handler then throws because the detached port has no accumulator or stated value. Use the normalized retraction length for that sink decision too.

---

Still open from earlier reviews (1) — the lines they point at have not changed:

• 🔴 Blocker · A sub-unit retraction can fail a cached no-accumulation run · packages/task-graph/src/task/StreamTypes.ts line 57 · from 58ce47e
return length > 0 ? Math.floor(length) : 0;
  This makes a length such as 0.5 a no-op, but StreamProcessor still tests the raw event.length > 0 before detaching a cache router. With noAccumulation enabled, an append stream that emits text, then a text-retract of 0.5, then finish without that port is therefore detached despite retracting nothing; the finish handler then throws because the detached port has no accumulator or stated value. Use the normalized retraction length for that sink decision too.

---

[mergestorm-vortex]

🔴 Blocker · Blocker

Preserve an accumulated empty string after full retraction

retractText can leave this accumulator entry as "" when a retraction removes the entire port. The finish merge later skips every empty accumulated value (if (text.length > 0)), so a stream like text-delta("partial"), text-retract(99), finish({}) completes with no text property instead of text: ""; downstream consumers of this declared output port then receive an absent value despite accumulation being enabled.

🤖 Prompt for AI Agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

In `@packages/task-graph/src/task/StreamProcessor.ts` (line 310, Blocker):
`retractText` can leave this accumulator entry as `""` when a retraction removes the entire port. The finish merge later skips every empty accumulated value (`if (text.length > 0)`), so a stream like `text-delta("partial"), text-retract(99), finish({})` completes with no `text` property instead of `text: ""`; downstream consumers of this declared output port then receive an absent value despite accumulation being enabled.

[sroussey]

Not changing this here. As far as I can tell it's the intended behaviour, not a regression.

• The finish merge's if (text.length > 0) predates this PR (df50135). Empty accumulated text never overrides the finish payload.
• So a port retracted to "" ends up exactly like a port that never streamed. The port is absent unless the producer states it in finish. That's the contract text-retract documents: the folded port reads as if the withdrawn text had never been sent.
• AgentTask, the only producer that retracts, always states text and reasoning in its finish payload, so its output always carries them.
• Writing "" for a fully retracted port would make it differ from a never-streamed one. Doing that for every empty accumulator would also change ports that only ever streamed empty deltas, which is outside this PR.

Leaving this thread open for the author to decide.

---

Generated by Claude Code

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Request changes — found 2 issue(s) outside changed lines at b4716c0.

Actionable comments posted: 2

🤖 Prompt for AI agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

Findings to address:
1. In `@packages/task-graph/src/task/StreamProcessor.ts` (line 310, Blocker):
   Preserve an accumulated empty string after full retraction
   `retractText` can leave this accumulator entry as `""` when a retraction removes the entire port. The finish merge later skips every empty accumulated value (`if (text.length > 0)`), so a stream like `text-delta("partial"), text-retract(99), finish({})` completes with no `text` property instead of `text: ""`; downstream consumers of this declared output port then receive an absent value despite accumulation being enabled.

2. In `@packages/task-graph/src/task/StreamTypes.ts` (line 57, Blocker):
   A sub-unit retraction can fail a cached no-accumulation run
   `return length > 0 ? Math.floor(length) : 0;`
   This makes a `length` such as `0.5` a no-op, but `StreamProcessor` still tests the raw `event.length > 0` before detaching a cache router. With `noAccumulation` enabled, an append stream that emits text, then a `text-retract` of `0.5`, then `finish` without that port is therefore detached despite retracting nothing; the finish handler then throws because the detached port has no accumulator or stated value. Use the normalized retraction length for that sink decision too.

---

Still open from earlier reviews (2) — the lines they point at have not changed:

• 🔴 Blocker · Preserve an accumulated empty string after full retraction · packages/task-graph/src/task/StreamProcessor.ts line 310 · from 0de83cd
retractText can leave this accumulator entry as "" when a retraction removes the entire port. The finish merge later skips every empty accumulated value (if (text.length > 0)), so a stream like text-delta("partial"), text-retract(99), finish({}) completes with no text property instead of text: ""; downstream consumers of this declared output port then receive an absent value despite accumulation being enabled.
• 🔴 Blocker · A sub-unit retraction can fail a cached no-accumulation run · packages/task-graph/src/task/StreamTypes.ts line 57 · from 58ce47e
return length > 0 ? Math.floor(length) : 0;
  This makes a length such as 0.5 a no-op, but StreamProcessor still tests the raw event.length > 0 before detaching a cache router. With noAccumulation enabled, an append stream that emits text, then a text-retract of 0.5, then finish without that port is therefore detached despite retracting nothing; the finish handler then throws because the detached port has no accumulator or stated value. Use the normalized retraction length for that sink decision too.

---

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Request changes — found 2 issue(s) outside changed lines at 241edfa.

Actionable comments posted: 2

🤖 Prompt for AI agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

Findings to address:
1. In `@packages/task-graph/src/task/StreamProcessor.ts` (line 310, Blocker):
   Preserve an accumulated empty string after full retraction
   `retractText` can leave this accumulator entry as `""` when a retraction removes the entire port. The finish merge later skips every empty accumulated value (`if (text.length > 0)`), so a stream like `text-delta("partial"), text-retract(99), finish({})` completes with no `text` property instead of `text: ""`; downstream consumers of this declared output port then receive an absent value despite accumulation being enabled.

2. In `@packages/task-graph/src/task/StreamTypes.ts` (line 57, Blocker):
   A sub-unit retraction can fail a cached no-accumulation run
   `return length > 0 ? Math.floor(length) : 0;`
   This makes a `length` such as `0.5` a no-op, but `StreamProcessor` still tests the raw `event.length > 0` before detaching a cache router. With `noAccumulation` enabled, an append stream that emits text, then a `text-retract` of `0.5`, then `finish` without that port is therefore detached despite retracting nothing; the finish handler then throws because the detached port has no accumulator or stated value. Use the normalized retraction length for that sink decision too.

---

Still open from earlier reviews (2) — the lines they point at have not changed:

• 🔴 Blocker · Preserve an accumulated empty string after full retraction · packages/task-graph/src/task/StreamProcessor.ts line 310 · from 0de83cd
retractText can leave this accumulator entry as "" when a retraction removes the entire port. The finish merge later skips every empty accumulated value (if (text.length > 0)), so a stream like text-delta("partial"), text-retract(99), finish({}) completes with no text property instead of text: ""; downstream consumers of this declared output port then receive an absent value despite accumulation being enabled.
• 🔴 Blocker · A sub-unit retraction can fail a cached no-accumulation run · packages/task-graph/src/task/StreamTypes.ts line 57 · from 58ce47e
return length > 0 ? Math.floor(length) : 0;
  This makes a length such as 0.5 a no-op, but StreamProcessor still tests the raw event.length > 0 before detaching a cache router. With noAccumulation enabled, an append stream that emits text, then a text-retract of 0.5, then finish without that port is therefore detached despite retracting nothing; the finish handler then throws because the detached port has no accumulator or stated value. Use the normalized retraction length for that sink decision too.

---

[mergestorm-vortex]

Review progress ██████████ 29/29 files

Request changes — found 2 issue(s) outside changed lines at 23d3f3d.

Actionable comments posted: 2

🤖 Prompt for AI agents

Pull request workglow-dev/libs#1039: https://github.com/workglow-dev/libs/pull/1039 (head branch `claude/wonderful-brahmagupta-4xsmg4`)

Verify each finding against current code. Fix only still-valid concrete bugs, skip the
rest with a brief reason, keep changes minimal, and validate. Skip Decision required,
policy forks, and "consider X" alternatives. Do not add new features, refactors, or
architecture beyond the fix; prefer the smallest diff.

Findings to address:
1. In `@packages/task-graph/src/task/StreamProcessor.ts` (line 310, Blocker):
   Preserve an accumulated empty string after full retraction
   `retractText` can leave this accumulator entry as `""` when a retraction removes the entire port. The finish merge later skips every empty accumulated value (`if (text.length > 0)`), so a stream like `text-delta("partial"), text-retract(99), finish({})` completes with no `text` property instead of `text: ""`; downstream consumers of this declared output port then receive an absent value despite accumulation being enabled.

2. In `@packages/task-graph/src/task/StreamTypes.ts` (line 57, Blocker):
   A sub-unit retraction can fail a cached no-accumulation run
   `return length > 0 ? Math.floor(length) : 0;`
   This makes a `length` such as `0.5` a no-op, but `StreamProcessor` still tests the raw `event.length > 0` before detaching a cache router. With `noAccumulation` enabled, an append stream that emits text, then a `text-retract` of `0.5`, then `finish` without that port is therefore detached despite retracting nothing; the finish handler then throws because the detached port has no accumulator or stated value. Use the normalized retraction length for that sink decision too.

---

Still open from earlier reviews (2) — the lines they point at have not changed:

• 🔴 Blocker · Preserve an accumulated empty string after full retraction · packages/task-graph/src/task/StreamProcessor.ts line 310 · from 0de83cd
retractText can leave this accumulator entry as "" when a retraction removes the entire port. The finish merge later skips every empty accumulated value (if (text.length > 0)), so a stream like text-delta("partial"), text-retract(99), finish({}) completes with no text property instead of text: ""; downstream consumers of this declared output port then receive an absent value despite accumulation being enabled.
• 🔴 Blocker · A sub-unit retraction can fail a cached no-accumulation run · packages/task-graph/src/task/StreamTypes.ts line 57 · from 58ce47e
return length > 0 ? Math.floor(length) : 0;
  This makes a length such as 0.5 a no-op, but StreamProcessor still tests the raw event.length > 0 before detaching a cache router. With noAccumulation enabled, an append stream that emits text, then a text-retract of 0.5, then finish without that port is therefore detached despite retracting nothing; the finish handler then throws because the detached port has no accumulator or stated value. Use the normalized retraction length for that sink decision too.

---