Inside Coding Agents
v0.2 preview
EXPERIMENT · CONTROLLED · PASSED

Reference Harness context compaction probe

Can the Reference Harness replace older turns with a deterministic checkpoint while preserving the verified conclusion?

reference-context-compaction-v12 repetitions5 metrics
Evidence boundary

This is a controlled experiment. It supports only the listed subjects, fixture, configuration, and outputs; it does not generalize to native product behavior or cross-agent ranking.

STATUSpassed
SUCCESSFUL2/2
FIXTUREd369431d683d
PROMPTd48225211fab
SUBJECTS + CONTROLS

Which variables are fixed?

Both repetitions will drop three messages, emit a 232-character checkpoint, preserve the declared conclusion, and share one normalized Trace fingerprint.

SDK

Reference Harness

reference-harness-2026-08-10-controlled

Model
scripted-model-v1
Provider
local-fixture
Comparability
not-comparable
Inspect snapshot →
CONTROLS

Run manifest

network
false
temperature
0
writes
false
sandbox
lesson-owned-temporary-fixture
RUNS + METRICS

Are repeated runs consistent?

python3 labs/runner.py reference-context-compaction-v1 --check
RUN 01passed

reference-context-compaction-v1-run-001

  • event-order-validpass
  • observations-matchpass
  • final-answer-matchpass
  • fixture-unchangedpass
  • deterministic-replaypass
d3e3171adaa9a3de960fc262798c5f56a65f7a634a63c44ccd7f38ea1fc45e59Replay 10 events →
RUN 02passed

reference-context-compaction-v1-run-002

  • event-order-validpass
  • observations-matchpass
  • final-answer-matchpass
  • fixture-unchangedpass
  • deterministic-replaypass
d3e3171adaa9a3de960fc262798c5f56a65f7a634a63c44ccd7f38ea1fc45e59Replay 10 events →
SUCCESS CRITERIA
  1. Both repetitions record three dropped messages and a 232-character compacted checkpoint.
  2. Both repetitions preserve the expected verified conclusion and share one Trace fingerprint.
LIMITATIONS
  • The checkpoint text is produced by deterministic lesson code, not a generative summarizer.
  • The probe checks declared facts, not semantic recall across long real-world sessions.
  • The Reference Harness result is not Native vendor evidence.
GENERATED REPORT

Reference Harness context compaction probe

  • Status: PASSED
  • Mode: controlled
  • Experiment: reference-context-compaction-v1

Question

Can the Reference Harness replace older turns with a deterministic checkpoint while preserving the verified conclusion?

Reproduction

python3 labs/runner.py reference-context-compaction-v1 --check

The command rebuilds the expected artifacts in memory and exits non-zero if a committed trace, result, fixture digest, prompt digest, or report differs.

Result

2 of 2 repetitions passed all declared metrics.

  • reference-context-compaction-v1-run-001: passed, 10 events, fingerprint d3e3171adaa9a3de960fc262798c5f56a65f7a634a63c44ccd7f38ea1fc45e59
  • reference-context-compaction-v1-run-002: passed, 10 events, fingerprint d3e3171adaa9a3de960fc262798c5f56a65f7a634a63c44ccd7f38ea1fc45e59

Metrics

  • event-order-valid: pass
  • observations-match: pass
  • final-answer-match: pass
  • fixture-unchanged: pass
  • deterministic-replay: pass

Input digests

  • Fixture: d369431d683da5c2cad36bb6dc6f76d429c2efbdc29171d5c8d9d4d7c3c575aa
  • Prompt: d48225211fab69c822df8f219e4e6c33c5c724a097f36375c43727194566e6e2

Evidence boundary

  • The checkpoint text is produced by deterministic lesson code, not a generative summarizer.
  • The probe checks declared facts, not semantic recall across long real-world sessions.
  • The Reference Harness result is not Native vendor evidence.
Back to Lab →