Stage 01// ingestturn 01 · step 1 of 6
Out of the whole repo: two files, one memory, three tools. Everything else stayed outside the window.
▸ the packing list, verbatim
task fix the flaky retry test
files src/retry.py 1.2k
tests/test_retry.py 0.8k
memory "flaked 3x this week,
always on slow CI"
tools run_tests · read_file · edit_file// the skill here was leaving the rest of the repo out.
Stage 02// assemblestep 2 of 6
This is the entire world the model will see this turn. Rules first, the ask last: recency is weight.
▸ the assembled window, block by block
[system] 8.0k rules + tool schemas
[history] 0.0k first turn, empty
[knowledge] 8.9k both files + the memory
[ask] 0.1k "fix the flaky retry test"
─────────────────────────────
window 17.0k of 64k// prompt = f(context). Can't diff it, can't debug it.
Stage 03// model callstep 3 of 6
The reply is just tokens: a thought, and a request. Note what it did not do: guess at a fix.
▸ the reply, raw blocks, nothing run yet
text "The failure is timing-shaped:
test_backoff asserts elapsed
< 0.3s. Reproducing it before
touching anything."
tool_call run_tests(
path="tests/test_retry.py")// a tool call is a sentence until something executes it.
Stage 04// parse & routestep 4 of 6
The parser doesn't understand the plan. It only sorts blocks: words to the transcript, calls to the sandbox.
▸ the routing table for those two blocks
block 1 text → say it
block 2 tool_call → run_tests
routed 1 say · 1 call
malformed 0 // errors go back around// on a bad turn, "malformed" is the row that saves the crash.
Stage 05// execute + observe
The agent didn't theorize. It ran the suite inside the wall and watched: 0.312 seconds against a 0.3 limit.
▸ the sandbox terminal, exactly as observed
$ pytest tests/test_retry.py -q
....F
FAILED test_backoff_under_300ms
AssertionError: 0.312 >= 0.3
1 failed, 4 passed in 2.41s// an observation beats a theory. Every turn.
Stage 06// compact + verify
Four kilobytes of test output became one sentence. The checklist moved by exactly one line.
▸ the state diff: kept vs dropped
+ "repro: jitter can push backoff
past the 0.3s assert"
- 4.1k of raw pytest output
todos ✓ reproduce the failure
· locate the cause // next// compaction is lossy on purpose. What survives is a design decision.
The gate// continue or exitend of turn 01
One turn produced one verified fact. The gate sends it around again, carrying everything it learned.
▸ the gate, evaluated on real state
state.done() ?
todos 1 of 5 complete
suite still red
→ False · around again// exit on verified todos, not on the model feeling done.
Turns 02–05// same ring, compounding
The next four take one lap each, because every lap starts with the last one's lesson already in the window.
// the bad patch on turn 04 was caught by the same gate you just watched say no.
Return// gate open5 turns · ctx 42k
A one-line fix, a green suite, a written summary. Out through the gate, the only way there is.
▸ the fix and the green run
- assert elapsed < 0.3
+ assert elapsed < 0.3 + JITTER_CAP
5 passed in 2.38s ✓