feat(goal): goal/operate runs are chopped mid-pass by the interactive max_steps — add a goal-scoped step budget

[7jrxt42BxFZo4iAnN4CX]

Problem

A goal/operate run is documented as unbounded by default: [goal] max_continuations defaults to 0 = unlimited, and the only terminal stops are completion, blocked, or the opt-in continuation backstop (goal_loop.rs: "Goals are unlimited by default: completion, blocked status, or explicit user control ends the run", DEFAULT_MAX_GOAL_CONTINUATIONS = 0; docs/CONFIGURATION.md "Goal loop").

But a single goal pass is still one turn, and every turn is capped by the interactive [tui].max_model_steps (default 200 — the R1 runaway guard). The turn loop itself says the intra-turn goal-continuation hook is "bounded by the step budget" (turn_loop.rs ~4912). So when a model enters a long tool loop inside one pass and does not close the turn, that pass is cut at 200 model steps with Maximum model steps reached before completion (limit: N) and the turn ends Failed — even though the goal is meant to run until done.

Result: an autonomous goal run is simultaneously unbounded (no pass/step cap at the goal level by default) and arbitrarily chopped mid-pass at the interactive ceiling. The knob designed to catch an interactive runaway loop is what stops an intentional autonomous run.

Current behavior

• Goal run started (operate mode); the model works a long tool loop inside one pass without emitting a tool-less closing answer → the turn never closes → max_steps (200) fires mid-pass → TurnOutcomeStatus::Failed with Maximum model steps reached before completion (limit: 200). Work done up to that point stays in the transcript, but the run is over.
• Goal-level budgets already exist but do not gate: GoalBudget { token_budget, time_budget_seconds, max_continuations } (goal_loop.rs), and "token/time accounting stays visible as telemetry but does not gate continuation".
• [goal] max_continuations counts passes, not steps, and defaults off (0).
• The only way to give an autonomous run more room today is to raise [tui].max_model_steps globally — which loosens the interactive runaway guard for everyone, including non-goal turns that a human is watching.

Desired behavior (options — maintainer's choice)

Give goal runs a step budget of their own, independent of the interactive max_steps:

• Option A — [goal] max_steps: a model-step ceiling scoped to goal runs (total across the run, or per pass), separate from [tui].max_model_steps. Default could stay generous/unbounded at the goal layer (matching the goal "unlimited" contract) or inherit a large finite default — but it must not be the interactive 200.
• Option B — surface the governing budget: wherever the 80% soft-landing / exhaustion notice is shown, name which budget governs (interactive max_steps vs goal budget), so "200" is not ambiguous.
• Either way, exhaustion of a goal step budget should end the run with the existing honesty contract (soft landing at ~80%, one bounded final-report turn) rather than a bare Failed that looks like a crash.

The goal layer already owns a natural home for this: GoalBudget in goal_loop.rs gains a step dimension alongside token_budget / time_budget_seconds / max_continuations.

Motivation

Autonomous/operate runs (including overnight or long unattended work) are an intended use of the goal loop. Today a long single pass inside such a run dies at the interactive guard, which the operator did not expect to govern autonomous work — the goal documentation says "unlimited by default". Splitting the budgets keeps the interactive runaway protection intact (default 200 untouched for watched turns) while letting autonomous runs have their own honest, visible limit.

Acceptance criteria

• A goal run that would previously die at 200 model steps inside one pass now runs under its own goal step budget and ends with the soft-landing/final-report honesty contract, not a bare Failed.
• Interactive [tui].max_model_steps default stays 200 for non-goal turns (guard unchanged).
• The effective governing budget is visible/disclosed where the exhaustion notice is shown (Option B).
• docs/CONFIGURATION.md goal section and config.example.toml document the new key.

Environment

• codewhale 0.9.12 (dev), Linux, tmux. Reproduced in an operate-mode goal run.

[github-actions]

Thanks @7jrxt42BxFZo4iAnN4CX for the report.

This issue is staying open for maintainer triage. CodeWhale gets better because people bring us real edge cases from real machines, providers, regions, and workflows.

If you can add a reproduction, logs, version output, screenshots, or the provider/model involved, that makes it much easier for us to verify and harvest the fix. Maintainers may comment /lgtmi to mark recurring issue reporters as approved so this intake note is skipped next time.

[Hmbown]

Problem

A goal/operate run is documented as unbounded by default: [goal] max_continuations defaults to 0 = unlimited, and the only terminal stops are completion, blocked, or the opt-in continuation backstop (goal_loop.rs: "Goals are unlimited by default: completion, blocked status, or explicit user control ends the run", DEFAULT_MAX_GOAL_CONTINUATIONS = 0; docs/CONFIGURATION.md "Goal loop").

But a single goal pass is still one turn, and every turn is capped by the interactive [tui].max_model_steps (default 200 — the R1 runaway guard). The turn loop itself says the intra-turn goal-continuation hook is "bounded by the step budget" (turn_loop.rs ~4912). So when a model enters a long tool loop inside one pass and does not close the turn, that pass is cut at 200 model steps with Maximum model steps reached before completion (limit: N) and the turn ends Failed — even though the goal is meant to run until done.

Result: an autonomous goal run is simultaneously unbounded (no pass/step cap at the goal level by default) and arbitrarily chopped mid-pass at the interactive ceiling. The knob designed to catch an interactive runaway loop is what stops an intentional autonomous run.

Current behavior

• Goal run started (operate mode); the model works a long tool loop inside one pass without emitting a tool-less closing answer → the turn never closes → max_steps (200) fires mid-pass → TurnOutcomeStatus::Failed with Maximum model steps reached before completion (limit: 200). Work done up to that point stays in the transcript, but the run is over.
• Goal-level budgets already exist but do not gate: GoalBudget { token_budget, time_budget_seconds, max_continuations } (goal_loop.rs), and "token/time accounting stays visible as telemetry but does not gate continuation".
• [goal] max_continuations counts passes, not steps, and defaults off (0).
• The only way to give an autonomous run more room today is to raise [tui].max_model_steps globally — which loosens the interactive runaway guard for everyone, including non-goal turns that a human is watching.

Desired behavior (options — maintainer's choice)

Give goal runs a step budget of their own, independent of the interactive max_steps:

• Option A — [goal] max_steps: a model-step ceiling scoped to goal runs (total across the run, or per pass), separate from [tui].max_model_steps. Default could stay generous/unbounded at the goal layer (matching the goal "unlimited" contract) or inherit a large finite default — but it must not be the interactive 200.
• Option B — surface the governing budget: wherever the 80% soft-landing / exhaustion notice is shown, name which budget governs (interactive max_steps vs goal budget), so "200" is not ambiguous.
• Either way, exhaustion of a goal step budget should end the run with the existing honesty contract (soft landing at ~80%, one bounded final-report turn) rather than a bare Failed that looks like a crash.

The goal layer already owns a natural home for this: GoalBudget in goal_loop.rs gains a step dimension alongside token_budget / time_budget_seconds / max_continuations.

Motivation

Autonomous/operate runs (including overnight or long unattended work) are an intended use of the goal loop. Today a long single pass inside such a run dies at the interactive guard, which the operator did not expect to govern autonomous work — the goal documentation says "unlimited by default". Splitting the budgets keeps the interactive runaway protection intact (default 200 untouched for watched turns) while letting autonomous runs have their own honest, visible limit.

Acceptance criteria

• A goal run that would previously die at 200 model steps inside one pass now runs under its own goal step budget and ends with the soft-landing/final-report honesty contract, not a bare Failed.
• Interactive [tui].max_model_steps default stays 200 for non-goal turns (guard unchanged).
• The effective governing budget is visible/disclosed where the exhaustion notice is shown (Option B).
• docs/CONFIGURATION.md goal section and config.example.toml document the new key.

Environment

• codewhale 0.9.12 (dev), Linux, tmux. Reproduced in an operate-mode goal run.

Thank you so much for filing this one, too!! This is so helpful

[7jrxt42BxFZo4iAnN4CX]

Implemented on main in f5755a883 ("Report throughput and cost truthfully and make goals durable"). Verified: that commit is an ancestor of main (compare/main...f5755a883 → behind_by=363, ahead_by=0).

Goal turns no longer share the ordinary interactive step ceiling. A goal-scoped budget now exists:

• GoalConfig::max_steps (crates/tui/src/config.rs:2090), resolved by Config::goal_max_steps() (crates/tui/src/config.rs:7459).
• Default DEFAULT_GOAL_MAX_STEPS = 1_000 (crates/tui/src/goal_loop.rs:40); 0 or absent resolves to the default, never to unlimited. Range clamps to 1..=100_000.
• Applied per engine turn at crates/tui/src/core/engine.rs:4977 — goal_turn && let Some(goal_max_steps) = self.config.goal_max_steps — and wired from the live config at crates/tui/src/runtime_threads.rs:9997 and crates/tui/src/tui/ui/frame.rs:902.
• Active goal turns now use goal.max_steps instead of tui.max_model_steps (documented at docs/CONFIGURATION.md:2292, Goal loop section at :2353, key at :2374).

Exhaustion behavior is also fixed: at ~80% of the budget the model gets a soft-landing notice, at exhaustion one bounded final report, the turn classifies as budget-exhausted, and the unfinished goal pauses with BudgetLimit instead of re-arming another turn.

[goal]
max_steps = 1000   # default; range 1..=100_000; bounds one turn, not continuation passes

Test coverage: crates/tui/src/core/engine/tests.rs:7758 exercises the goal budget; the commit reports the goal-budget selection green.

Verified in the source tree at 8d66b9a (0.9.13 dev). Closing as fixed.