Rescue a tight tool-call loop sideways instead of killing the turn

[devin-ai-integration]

Summary

A tight tool-call loop currently ends the turn: handleToolCallLoopBreak writes a synthetic end_turn and expires the pin. Prod session loom-398a68c6… (2026-08-25 19:31:12) shows why that buys nothing — the pin was cleared, the client resent 37s later, and the HMM re-picked the same arm (gpt-5.6-luna, score 0.88). Net effect: one dead turn on a session that was not actually stuck. Three fixes, all in internal/proxy:

1. Move the session sideways before stopping it. The tight detector now calls handleToolCallLoopSideways first, which reuses the struggle machinery (StruggleEscalationRoster.EscalationTarget: cheapest cluster above the pin's own, else a sideways arm within it) and re-pins with a new translate.ReasonLoopSideways before routing, so this same turn dispatches on a different model instead of dying:

rescue := s.handleToolCallLoopSideways(ctx, sig, count, installationID, sessionKey, loopRole)
if !rescue.Moved {
    loopingModel, loopingProvider := loopAttribution(rescue.LoopingModel, rescue.LoopingProvider, feats.Model, providers.ProviderAnthropic)
    return s.handleToolCallLoopBreak(ctx, ..., loopingModel, loopingProvider, rescue.UserForced, feats.Tokens)
}

The synthetic stop survives as the fallback for every case where a move isn't valid, tracked by an explicit action taxonomy: no_pin, user_forced, already_moved (the pin is already a loop/struggle rescue — looping again there is a task problem, not a misroute), no_target, disabled. The pin write uses context.Background() like the other durable auto-pins; a failed write degrades to the stop. ReasonLoopSideways joins the immutable-sticky set in turnloop.go so the rescued arm can't be re-scored back onto the looping model.

handleToolCallLoopBreak gains a preserveForcedPin argument (same shape as handleNoProgressBreak): the fallback path is now reachable with a /force-model pin in place, and automatic eviction must not clear it.

2. Attribute the loop to the model that was serving it. The detector runs before routing, so feats.Model is the client's inbound baseline — the prod session logged looping_model=claude-fable-5 / anthropic while every one of its 36 turns was served by gpt-5.6-luna / openai. loopAttribution prefers the pin and falls back to the requested model only when there is no pin to read.

3. Exempt poll tools from the identical-args detector. shell_output (and get_output / BashOutput) drain a buffer that changes underneath a byte-identical request — repetition is the tool's normal usage, and it is what tripped the detector in prod. Poll signatures are skipped when counting; they still occupy window positions, so a genuine repeat interleaved with polls still trips.

Not addressed here: struggle escalation needs turns AND wall-clock (30 / 10m), so a fast session like this one (36 turns in 4.5 min) never arms it — that gate is worth revisiting separately.

Link to Devin session: https://app.devin.ai/sessions/e32d7e9712e94512b8ecb03f29c14a37
Requested by: @steventohme

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

PR author is not in the allowed authors list.

[claude]

Claude finished @devin-ai-integration[bot]'s task —— View job

---

Posted an advisory comment-length review with 4 suggestionable nits — won't block merge.

[workweave-bot]

Advisory only — comment-length nits. Won't block merge.

[workweave-bot]

Advisory only — comment-length nits. Won't block merge.

[claude]

Claude finished @devin-ai-integration[bot]'s task —— View job

---

Claude finished @devin-ai-integration[bot]'s task —— View job

---

Posted an advisory comment-length review with 2 suggestionable nits — won't block merge.

[workweave-bot]

Advisory only — comment-length nits. Won't block merge.

[claude]

Claude finished @devin-ai-integration[bot]'s task —— View job

---

• Run git diff and analyze added comment blocks
• Check for 3+ line verbose comment blocks
• Post suggestion review if needed, or stay silent

View job

[cursor]

Unservable rescue pin skips stop

High Severity

The sideways writer pins Providers[0] after an availableModels check only, with no forcedModelBinding / request-enabled-provider validation. Moved is still true, so the synthetic stop is skipped. runTurnLoop then drops the pin on provider ineligibility and the scorer can re-select the looping model, after which writeNewPin overwrites the rescue.

Additional Locations (1)

• internal/proxy/service.go#L2662-L2667

 

^{Triggered by learned rule: Force-model and user-facing model commands must resolve via catalog, not assume default provider}

^{Reviewed by Cursor Bugbot for commit ef3a5f9. Configure here.}

[devin-ai-integration]

Runtime verification — loop-sideways rescue

Ran the router locally in docker compose from this branch with a mocked Fireworks upstream and a mocked HMM roster sidecar (balanced: [z-ai/glm-5, qwen/qwen3-coder]), a seeded session pin, and hand-crafted tool-call histories over /v1/messages. Every case was also replayed against a docker image built from the merge-base (75710ced) for a real before/after.

Main regression — 5× identical shell_output is no longer stopped

before (main)	after (this PR)
body: "→ Tool-call loop detected: shell_output was called 5 times..." log: Tool-call loop detected; breaking turn tool_name=shell_output upstream never called	X-Router-Model: z-ai/glm-5 body: MOCK_UPSTREAM_OK... no loop log lines; upstream request recorded

Sideways re-pin on a real loop (5× identical Bash, pinned session)

X-Router-Decision: loop_sideways
X-Router-Model:    qwen/qwen3-coder
marker:            "→ qwen/qwen3-coder · switched model to break a tool-call loop"
body does NOT contain "Tool-call loop detected"

router.loop_sideways looping_model=z-ai/glm-5 looping_provider=fireworks action=moved
  escalation_target=qwen/qwen3-coder escalation_cluster=balanced policy_group=balanced
  loop_tool=Bash repeat_count=5 window_size=10

mock upstream: + accounts/fireworks/models/qwen3-coder-480b-a35b-instruct
session_pins:  default_mid | qwen/qwen3-coder | fireworks | balanced | loop_sideways | live

The same turn dispatched on the new arm, and the next turn stays pinned (pin_reason=loop_sideways, planner_outcome=skipped) — sticky immutability works.

Fallback paths

• Replay after rescue → action=already_moved, synthetic stop, looping_model=qwen/qwen3-coder, no upstream call.
• Fresh session with no pin → action=no_pin, synthetic stop.
• /force-model session → action=user_forced, synthetic stop, and the user-forced pin is still live afterwards (no eviction).
• Roster unreachable/invalid → action=no_target, synthetic stop (graceful).

Attribution fix

Same seeded pin (z-ai/glm-5), same payload:

main:     Tool-call loop detected; breaking turn looping_model=claude-sonnet-4-5 looping_provider=anthropic
this PR:  router.loop_sideways               looping_model=z-ai/glm-5        looping_provider=fireworks

Adversarial: poll exemption doesn't mask real loops

5× shell_output interleaved with 5× identical Bash (window exactly 10) still trips on Bash and moves sideways to qwen/qwen3-coder.

Not exercised at runtime: ROUTER_LOOP_ESCALATION_ENABLED=false, pin-write-failure fallback, and the contents of the loop_escalation_events row.

One prod-relevant nuance found while testing: a genuinely degenerate reply evicts the pin (isDegenerateResponse), so a loop turn on that session afterwards falls back to no_pin rather than moving sideways.

Devin session

[claude]

Claude finished @devin-ai-integration[bot]'s task —— View job

---

• Run git diff and analyze added comment blocks
• Check for 3+ line verbose comment blocks
• Post suggestion review if needed, or stay silent

Posted an advisory comment-length review with 1 suggestionable nit — won't block merge.

[workweave-bot]

Advisory only — comment-length nit. Won't block merge.

[claude]

Claude finished @devin-ai-integration[bot]'s task —— View job

---

• Run git diff and analyze added comment blocks
• Check for 3+ line verbose comment blocks
• Post suggestion review if needed, or stay silent

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ad27221. Configure here.}

[cursor]

Pro suffix over-excludes distinct models

Medium Severity

sameEngine treats any catalog ID that differs only by a -pro suffix as the same engine, so a tight-loop rescue excludes gpt-5.4-pro and gpt-5.5-pro when the pin is gpt-5.4/gpt-5.5. Those rows are separate models (own prices, no UpstreamID alias), unlike gpt-5.6-luna-pro which maps to the same upstream weights. A looping session can skip a genuine different arm and fall through to a weaker sideways pick or a stop.

Additional Locations (1)

• internal/proxy/loop_detection.go#L374-L385

 

^{Reviewed by Cursor Bugbot for commit ad27221. Configure here.}