Remove the provider stream throughput watchdog and the live tok/s working-line readout

[code-yeongyu]

Summary

#1743 (issue #1739) shipped two user-facing behaviours that are being withdrawn in full:

1. a sustained-rate watchdog that aborts an in-flight provider stream whose measured rate falls below retry.provider.minThroughputTokensPerSecond (default 8 tok/s over a 20s window), surfaced as Provider stream throughput degraded: <n> tok/s over <n>s (floor <n> tok/s);
2. a live - N.N tok/s segment on the interactive working line, added only so that verdict was observable while it was being measured.

Both must be removed completely - code, settings, docs, changelog entry, fork-tracking notes and tests - so the tree reads as if neither had ever been written.

Why

• The floor fires on healthy turns. A normal turn measured at 6.1 tok/s over the 20s window had its request aborted mid tool call; the transcript showed the tool call as error plus the raw watchdog string. Thinking-heavy models and gateways that batch several tokens per delta legitimately sustain rates under the shipped floor, so the guard turns a slow-but-correct answer into a hard failure and spends a fallback hop on it.
• It cannot be softened in place. The wording is simultaneously the classifier token (packages/ai/src/utils/retry.ts), so the message cannot be reworded, and the verdict aborts the request controller, so the partial answer for that turn is discarded rather than delivered slowly.
• The readout has no remaining purpose. With the watchdog gone, a per-delta rate on the working line is noise on every turn; end-of-turn rate is already reported by the builtin TPS extension.

Expected (ideal state)

No trace of either feature on main:

• packages/agent/src/stream-throughput-watchdog.ts deleted, together with AgentLoopConfig.streamThroughput, AgentOptions.streamThroughput and the packages/agent/src/index.ts exports.
• The agent-loop assistant reader back to the start-bound and idle-bound guards only.
• isProviderStreamThroughputDegradedError, the "provider stream throughput degraded" retryable pattern, the _handleRetryableError branch, the stream_throughput_degraded session event and its interactive notice box all gone.
• retry.provider.minThroughputTokensPerSecond / throughputWindowMs / throughputGraceMs removed from the settings type, from SettingsManager.getAgentStreamThroughputOptions(), from core/sdk.ts, from the test harness and from the docs/settings.md rows and example.
• Working line back to Working (<elapsed> - esc to interrupt); StreamRateMeter, estimateStreamedUnits and getWorkingTokensPerSecond() gone from interactive mode.
• Tests: packages/agent/test/agent-loop-throughput-watchdog.test.ts and packages/coding-agent/test/suite/retry-fallback-throughput-degraded.test.ts deleted; the settings-manager, working-status and interactive-mode-status (test(coding-agent): unbreak main's working-indicator status test #1748) additions reverted; the explanatory comment added to agent-loop-stream-start-timeout.test.ts restored to its previous text.
• The CHANGELOG bullet and the changes.md sections describing the feature are deleted, not annotated: the removal is a withdrawal, not a documented change.

Acceptance criteria

• git grep -E 'minThroughputTokensPerSecond|throughputWindowMs|throughputGraceMs|StreamRateMeter|StreamThroughputDegraded|stream-throughput-watchdog|streamThroughput|estimateStreamedUnits|stream_throughput_degraded|getWorkingTokensPerSecond' returns nothing.
• The interactive working line renders Working (1m 12s - esc to interrupt).
• Green: packages/agent agent-loop suites, packages/coding-agent settings-manager / interactive-mode-status / interactive-mode-working-status / test/suite retry-fallback suites, packages/ai retry suites.

Related

• Slow-but-alive provider streams are undetectable: a ~2 tok/s turn never times out, never retries, and never falls back #1739, fix(agent): fail over provider streams that trickle below a throughput floor (#1739) #1743 - the feature being removed
• test(coding-agent): unbreak main's working-indicator status test #1748 - test stub that exists only to feed the tok/s reader
• Stream-start timeout is shown to the user as a raw error instead of failing over #1740, fix(coding-agent): explain provider stream stalls instead of showing the watchdog string #1744 - stall surfacing must keep working; it is the silence class, not the rate class, and stays

[code-yeongyu]

Removal evidence (measured)

Tree state — reverts ced1dadeb1 + d04674187c (#1743) and e5ab39dba8 (#1748) at HEAD 73bd1577c2.

• 18 of the 25 touched files are byte-identical to the pre-fix(agent): fail over provider streams that trickle below a throughput floor (#1739) #1743 tree (62c03f1b37). The other 7 (packages/ai/src/utils/retry.ts, agent-session.ts, interactive-mode.ts, three changes.md, CHANGELOG.md) differ only by unrelated later work - fix(coding-agent): explain provider stream stalls instead of showing the watchdog string #1744 stall copy, the resident-store change and /rename - with 0 throughput-related diff lines against that base.
• git grep -E 'minThroughputTokensPerSecond|throughputWindowMs|throughputGraceMs|StreamRateMeter|formatWorkingRateSegment|StreamThroughputDegraded|stream-throughput-watchdog|streamThroughput|estimateStreamedUnits|isProviderStreamThroughputDegradedError|stream_throughput_degraded|getWorkingTokensPerSecond' on the branch: no matches.

Focused suites — run on a second machine (macOS arm64, bun 1.4.2, vitest 4.1.11) from a worktree checked out at 73bd1577c2 (HEAD asserted equal to the pushed branch before running):

package	files	result
packages/agent	agent-loop-stream-start-timeout, agent-loop-idle-timeout	6 passed (1 file)
packages/coding-agent	interactive-mode-status, interactive-mode-working-status, settings-manager	168 passed (3 files)
packages/ai	retry, provider-retry, retry-profile-classifiers	133 passed (3 files)
packages/coding-agent	all 21 test/suite/retry-fallback*.test.ts	242 passed (21 files)
packages/coding-agent	5 stall suites (stream-start-stall-*, retry-fallback-stall-shared-budget, ...)	18 passed (5 files)

Stall handling (#1740) is verified intact by the last row: it is the silence class and keeps its own bounds, copy and shared retry budget.

One file-level failure, classified environmental, not a regression: packages/agent/test/agent-loop.test.ts fails to load with Cannot find package '@earendil-works/chord/context' imported from packages/agent/src/harness/context.ts. The same file fails identically on a BASE worktree of main (a7dee82bb1) in the same environment - the linked node_modules has chord unbuilt. It is a setup artifact of running vitest against a symlinked worktree, not something this diff touches; CI's own Test jobs are the authority.

Changelog gate: the PR carries no-changelog. The entry for #1739 was deleted from the released [2026.9.16-2] section (same treatment #1756 gave the #1754 entry) and no "Removed" note replaces it, because the decision is that the feature reads as never having shipped rather than as an added-then-removed behaviour.

[code-yeongyu]

#1761 merged as 1197629472 at 2026-09-16T08:42:33Z, with every required check green (Check and test fan-in, Changelog gate, Review claim gate; CI, Changelog gate, Read summary binary parity, Publish Model Catalog and Review Claims all success at fedd1ab594).

The Windows Compiled extensions job failed once on the pre-rebase head with scripts/build-binaries.sh: line 195: zip: command not found (exit 127) and passed on the rebased head in 1m44s. That step shells out to zip/unzip, which are not guaranteed on the Windows runner image; if it recurs it deserves its own issue rather than a rerun.

One trace remained after the merge: the four changes.md trackers recorded the removal by naming the identifiers they describe as deleted, so a grep of the tree still found the feature there. #1762 rewords those entries in prose, keeping the file paths and all four canonical sections the changes.md policy requires. Closing this issue once that lands and the grep is clean on main.