SFT of hamishivi/Qwen3.5-9B for self / dynamic context management in long-horizon agents. The model decides when and how to compact its own context (via an edit_context tool) while solving long-context agentic tasks.

This is the mk2 iteration. It differs from the previous 2-domain model in two ways:

Three domains instead of two — terminal/coding (CLI-Gym, TMax) plus long-horizon web research (OpenResearcher), so compaction is learned on research logs as well as terminal transcripts.
A thinking-capped teacher. Rollouts were distilled from bottlecapai/ThinkingCap-Qwen3.6-27B rather than Qwen/Qwen3.6-27B. The mk1 evaluation found that generation verbosity, not compaction, dominated inference cost — generation is the one term neither prefix caching nor compaction can discount — so mk2 targets the length of the distilled THOUGHT block. Measured on the built datasets, this cut teacher THOUGHT tokens per turn by −25% (OpenResearcher), −37% (TMax) and −11% (CLI-Gym) versus mk1.

Training data

Segmented self-compaction trajectories generated with a Codex/Claude-Code-style harness and split at compaction boundaries into independent SFT sequences. The model is trained on one pre-balanced mixture, osieosie/drctx-mix-mk2-nc20-bs20-os2 (17,595 rows), built from the three per-domain archives:

The mixture is balanced by compaction form (nc20-bs20-os2 = 20% non-compaction, 20% of compactions from the forced backstop, scarce self-initiated forms oversampled 2×) so the model sees non-compaction turns and rarer edit types, not just the dominant summarization behavior:

form	rows	share
`self:summarization`	10,181	57.9%
non-compaction	3,519	20.0%
`backstop:full_compaction`	2,815	16.0%
`self:offload`	668	3.8%
`self:full_compaction`	320	1.8%
`self:deletion`	92	0.5%

Each per-domain archive ships normal / medium / high compaction-effort splits (generated at n1000 × k4 per domain per effort).

Recipe

Base model: hamishivi/Qwen3.5-9B (gated-delta hybrid)
Objective: assistant-only loss with prompt-length-gated masking (carried-over context is masked; only newly generated tokens after each compaction boundary are trained)
Max sequence length: 32,768 (matches the 32K generation budget)
Epochs: 2 · global batch: 128 · LR: 2e-5 (linear, warmup 0.03) · bf16
Parallelism: DeepSpeed ZeRO-3 (no offload) + Ulysses sequence parallelism
Trainer: vendored open-instruct fork

Evaluation

Same unified (compaction) harness for every policy: strict 32K cap, max_steps=64, max_tokens=16384, temp=0.7, top_p=0.95, k=1, single seed. base = hamishivi/Qwen3.5-9B; mk1 = same recipe with the un-capped Qwen3.6-27B teacher.

OpenResearcher (rl split, n=250, LLM-graded F1) — mk2 restores base-level answer quality while keeping the compaction benefit:

metric	base	mk1	mk2
mean F1	0.449	0.429	0.448
empty-answer rate	28.8%	40.4%	32.8%
fully correct	0.380	0.375	0.388
generation tokens / rollout	5,587	10,293	6,029
KV-cache-aware FLOPs / rollout	8.51e15	1.65e15	1.53e15 (−82% vs base)

Terminal-Bench 2 (89 tasks) — task success is flat within noise (SE ≈ 0.039 at n≈88, k=1), and compute remains above base:

metric	base	mk1	mk2
resolved rate	0.182	0.161	0.148
generation tokens / rollout	31,597	58,485	58,257
KV-cache-aware FLOPs / rollout	3.96e15	4.52e15	4.42e15 (+11% vs base)

Compaction behavior. edit_context anchor apply-success is 97% (OpenResearcher) / 92% (TB2), and 73% / 64% of edits span multiple turns, so the model compacts across whole sub-threads rather than only the last observation. mk2 reaches mk1's context control with roughly half the edits (OR 1,032 → 528; TB2 428 → 128) at equal-or-tighter peak context — and since every edit breaks the prefix cache, fewer-but-equally-effective edits is the cheaper behavior.

Known limitations

Form collapse toward summarization. Despite the mixture oversampling them, offload and deletion are almost never chosen at inference (≈0% on OpenResearcher; 2 offload events on TB2, neither read back). If you need those behaviors, prompt for them explicitly.
Terminal-domain cost is unimproved. The terser teacher shortened thinking on both domains, but on TB2 the savings were reallocated into larger bash payloads (heredocs writing whole source files), leaving total generation flat. Terminal output is ~40% code payload, which a thinking-length lever does not reach.
Trained and evaluated at a 32K context budget; behavior at much larger budgets is untested.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "osieosie/qwen3.5-9b-drctx-mix-mk2-nc20-bs20-os2-sft-e2"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="bfloat16", device_map="auto")

Serving with vLLM: the checkpoint ships open-instruct's flattened qwen3_5_text config (Qwen3_5ForCausalLM), which older vLLM builds do not recognize as a top-level architecture. If vllm serve rejects it, serve the base model's wrapper config.json (model_type: qwen3_5) over these weights with --language-model-only; the weight keys match the wrapper's language tower.

The model expects the harness system prompt that documents the edit_context tool and stamps [[B<n>]] block markers on context messages — it anchors its edits on those markers, which is why anchor apply-success is >90%.