feat: tool truncation + auto-compaction to bash harness

[mikasenghaas]

Summary

• add an optional CompactionConfig to the bash harness: compact into a handoff summary at summarize_at_tokens, or when 16k tokens remain below the model context window
• the program discovers the window itself from the provider's /models card (via models.list()), so the interception server gains a stateless GET /v1/models relay (one route serves every dialect)
• truncate a tool result over 20KB to its head and tail with a warning naming the original size
• compact reactively on an attributed 400/413 overflow error, from the current state; a rejected checkpoint falls back to the last state that passed a threshold check (which holds a full reserve of room by definition), and an empty or tool-calling reply is resampled
• after three failed checkpoint attempts the program ends the run cleanly — still a trainable sample
• an overflow with no history beyond the task propagates: nothing to reclaim

The RLM harness gets the same policy in #2459, stacked on this PR so this one can merge before nano-rlm does. Together with #2453 (merged) and #2459 this supersedes #2448.

Breaking

• None on its own: compaction is a new optional field, unset by default.

Verification

• uv run pytest -q tests/v1 — passed; live E2E tests skipped without PRIME_API_KEY

• Tool truncation checked by exact match: a synthetic 55KB output produces the byte-identical expected head/tail + warning header; short outputs pass through untouched.
  Terminal-Bench 2 e2e: 8 tasks, local vLLM poolside/Laguna-XS-2.1 at 32k (glm45 reasoning + glm47 tool parsers), compaction = {} so the threshold auto-discovers to 32768 − 16384 = 16384. Trace analysis:

• 8/8 rollouts end ok=true with scoring run (1 solved); no harness errors, no truncation stops.

• The threshold triggers exactly where designed: 7/8 episodes compacted with peak contexts of 17.0k-23.4k tokens; the one episode that stayed at 15.4k never compacted.

• 20KB tool truncation visible in 3 episodes' traces (Warning: truncated output with original size).

• Laguna puts its entire checkpoint reply in the reasoning channel, so under the summaries-are-content-only rule every compaction on this model exercises the failure path end to end: three resampled checkpoint attempts, then the program ends the run cleanly - the rollout stays a trainable sample. Summary carry-over itself was demonstrated on models that answer in the content channel (Qwen3-0.6B at 4k/8k, deepseek-v4-flash on tb2): every non-final branch ends with the checkpoint prompt followed by the summary, and the next branch opens with [system, framed summary].

• A final combined verification run on a content-channel model is pending.

🤖 Generated with Claude Code

---

Note

Medium Risk
Opt-in compaction rewrites conversation history and adds a new upstream relay route; mistakes could affect long agent rollouts or model discovery, but default behavior is unchanged when compaction is unset.

Overview
Adds optional context compaction to the bash harness agent loop so long rollouts can keep going without hitting the model window. BashHarnessConfig gains CompactionConfig (optional summarize_at_tokens); enabling compaction forwards --compaction / --summarize-at-tokens into the uv program.

When compaction is on, the program uses a Compactor to trigger on token usage (default threshold: model context window minus 16k, discovered via models.list()), on finish_reason == length, or on provider 400/413 overflow errors. It replaces history with a handoff summary (checkpoint prompt, tool_choice=none, up to three retries with fallback to the last “good” snapshot), then continues; unrecoverable compaction ends the loop cleanly instead of crashing.

Tool results over 20KB are middle-truncated (head/tail + warning) before they enter the transcript, including interception rewrites.

The interception server adds GET /v1/models as a dialect-aware relay to upstream so the bash program can read context-window fields from model cards when no explicit threshold is set.

^{Reviewed by Cursor Bugbot for commit cad252f. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add tool truncation and auto-compaction to bash harness

• Adds CompactionConfig to BashHarnessConfig so callers can enable compaction and set a summarize_at_tokens threshold; the harness passes --compaction and --summarize-at-tokens CLI flags to the program
• Introduces a Compactor class in the bash program that summarizes conversation history when token thresholds are hit or context overflow is detected, rebuilding to a compacted system+user state; raises CompactionFailed after 3 failed checkpoint attempts to terminate cleanly
• Truncates tool output middle-out to TOOL_OUTPUT_MAX_BYTES (20000) before inserting into the conversation, preserving head and tail with a warning
• When compaction is enabled without an explicit threshold, auto-discovers one from the upstream model's context window via a new proxied /v1/models endpoint in InterceptionServer
• Behavioral Change: chat() now returns the raw completion object instead of a pre-extracted message, and accepts tool_choice; tool outputs exceeding 20000 bytes are truncated before entering the conversation

^{Macroscope summarized cad252f.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This PR adds a substantial context-management workflow to the bash harness, including truncation of existing tool outputs, model discovery, summarization retries, and a new authenticated relay route. Its runtime behavior and unresolved overflow-recovery concern warrant human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[cursor]

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

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

^{Reviewed by Cursor Bugbot for commit f07f60a. Configure here.}

[cursor]

Overflow recovery requires a threshold

Medium Severity

Reactive compaction on a 400/413 overflow is skipped whenever threshold is None. An empty compaction config that cannot discover a window therefore never recovers from overflow, even though that recovery is described as independent of the proactive threshold.

 

^{Reviewed by Cursor Bugbot for commit f07f60a. Configure here.}

[mikasenghaas]

By design: compaction requires a known context window (discovered or explicit); without one the overflow propagates. Documented scoping - training always knows the window.