does optchat stay small after fifty tool calls?

the new optmem / optchat work belongs here too. this follow-up measures the size of a derived view against an accumulated transcript, while keeping answer quality and actual model requests as separate questions.

what we actually measured

the source report extends a file-review protocol to 50 public python files. each step adds a user line, a file echo capped at 8,000 characters, and a short reply. the comparison is the accumulated text versus a derived optchat view.

the local compactor was qwen3-0.6b-q8, with a 2,048-token context window. it saw a 1,600-character excerpt, not the full tool echo. that difference matters when asking what information survived.

the full trajectory

raw transcript

270,271 b

at turn 50

optchat view

16,621 b

same recorded turn

view / raw

6.15%

byte ratio, not token savings

placeholders

at the end of this run

solid: accumulated transcript. dashed: optchat view. bytes, not tokens; one recorded trajectory, not a distribution or a quality score.

the view is smaller, but it also grows. this run does not demonstrate constant context. the source report says higher tree levels were not needed within its configured budget; the observed reduction came from per-message summaries.

show all 50 recorded turns

context-size trajectory; source-reported bytes
turn	raw bytes	view bytes	view / raw
1	2,656	516	19.43%
2	5,134	724	14.10%
3	13,275	1,019	7.68%
4	16,976	1,672	9.85%
5	25,089	2,189	8.72%
6	33,208	2,671	8.04%
7	39,571	2,851	7.20%
8	46,547	3,154	6.78%
9	54,664	3,540	6.48%
10	61,301	3,818	6.23%
11	64,462	3,973	6.16%
12	67,221	4,174	6.21%
13	72,848	4,821	6.62%
14	80,985	5,150	6.36%
15	85,239	5,799	6.80%
16	93,386	5,995	6.42%
17	94,821	6,160	6.50%
18	96,017	6,559	6.83%
19	96,976	6,763	6.97%
20	105,099	7,065	6.72%
21	111,973	7,268	6.49%
22	120,117	7,601	6.33%
23	121,251	8,078	6.66%
24	129,374	8,290	6.41%
25	132,943	8,480	6.38%
26	141,090	8,696	6.16%
27	145,545	8,885	6.10%
28	153,682	9,238	6.01%
29	156,325	9,436	6.04%
30	164,454	9,643	5.86%
31	165,639	10,182	6.15%
32	172,666	10,326	5.98%
33	180,290	10,512	5.83%
34	184,598	10,691	5.79%
35	190,177	10,874	5.72%
36	196,106	11,086	5.65%
37	201,776	11,249	5.57%
38	204,191	11,672	5.72%
39	206,744	11,863	5.74%
40	210,017	12,046	5.74%
41	218,138	12,586	5.77%
42	224,559	12,803	5.70%
43	229,925	13,316	5.79%
44	237,491	13,721	5.78%
45	241,353	14,274	5.91%
46	249,482	14,860	5.96%
47	253,316	15,262	6.02%
48	254,622	15,483	6.08%
49	262,136	15,938	6.08%
50	270,271	16,621	6.15%

small is not the same as correct

there was no held-out answer-quality verifier. a small view could omit exactly the detail a later question needs. zero remaining placeholders says the nodes were built, not that their summaries were faithful or sufficient.

the source report records one failed compaction request and a fallback line. the manifest also says the optchat implementation was uncommitted relative to its recorded source pin. publishing that limitation is not the same as repairing the missing provenance.

the older observationpack / sol-pi comparison is background, not another arm in this trajectory. the earlier cross-harness memory studyremains separate and unscored. neither is promoted by making this article visible.

what still needs a test

freeze the exact implementation, measure the actual model-input boundary, and ask delayed questions against held-out evidence. keep correctness, stale-context errors, latency, bytes, and tokens separate. newer fresh-turn and hermes adapters need their own exact-head evidence; this old run cannot certify them retroactively.

what we actually measured

the full trajectory

small is not the same as correct

what still needs a test

source and provenance