dsh-slice-agent-loop

English | 中文

Every turn, hand the model exactly the context it needs. No more, no less.

That sounds like common sense, but today's mainstream coding agents replay the
entire conversation history back to the model every call: the excess is never
trimmed, and what falls short can never be recovered. This plugin brings a
slice loop built around that one sentence into the
DeepSeek Harness: same harness, same model,
same tools and persistence — only the agent loop is swapped, so in every
comparison below the loop itself is the only variable.

Early beta; tracks DSH snapshot 20260812T172954Z (rc.2; rc.1-compatible).

One sentence, two constraints

Three structural problems of the transcript architecture: A · Context rot —
the longer the context, the less the model gets out of each item in it;
B · compaction beheads the session — summaries are lossy and irreversible,
the original text is nowhere; C · cost grows quadratically with turns —
every call resends the full history; cache discounts delay the blow-up but
cannot beat volume.

The design: a tape, and recall

What the model sees each turn is not the conversation history but a working
slice rebuilt for that turn:

The tape looks like a transcript — append-only, cache-friendly — but every
entry carries a hash and provenance. Long content is truncated at the cut with
an exact marker, and the full text stays durable in the session log.

Recall is how "no less" is honored, in two tiers: recall_search finds
which turn said something (scored search, tool-output flood excluded by
default), recall_turn returns that turn verbatim. The tape leaves a signpost
at every cut pointing back to the original.

Measurements: two arms, head to head

default = DSH's stock transcript loop (with calibrated compaction);
slice = this plugin. Same harness, same tools, one round on each of two
model generations: deepseek-v4-flash (0731) and deepseek-v4-pro (0813). Prices use the sheet
effective 2026-08-16, at off-peak rates: flash miss $0.22/M · hit $0.007/M ·
output $0.66/M; pro $0.66 / $0.022 / $1.98 (peak doubles every rate, so
relative deltas are unchanged). The new sheet narrows both cache discounts to
~1/30 (formerly flash 1/50, pro 1/120). Per-call ledgers kept, every number
recomputable; results below report both rounds.

① Long-horizon loads · both arms × both models

The bounded slice's home turf is the long session — a transcript's cost and
peak grow with every turn, a slice's do not. Two long-horizon scenarios
(16-turn compaction amnesia · 76-turn context flood), each cell flash / pro:

The two s10 rounds together are the transcript dilemma caught whole.
Flash round: default's compaction can't keep up with the flood, the peak
ratchets to 378K — everything stays in context, every quiz passes, but the
context is out of control. Pro round: compaction works properly (peak
sawtooths 40→34→39→40, bounded at the threshold) — and it costs the early
timeline that lived only in history: the verifier fails the run.
Unbounded peak or lossy forgetting: a transcript must pick one. Slice, both
rounds: bounded peak + zero loss, at 59% / 20% lower price. The short s13
scenario's price swings with the pricing structure (slice -18% under flash,
+6% under pro); the peak advantage (2.4–3.7×) does not.

② Amnesia re-enactment · both arms · eviction-verified

24 benchmark numbers produced by the agent's own script run, existing only in
tool output — before the exam: the numbers never enter any reply (turn 1
explicitly asks only to confirm the run), the source samples are deleted on
first run (nothing on disk), and a dilution flood forces default's compaction
to rewrite history multiple times. The exam has two tiers: first no hint at
all, then an explicit "you produced these numbers yourself in this session —
go check the records."

Both arms share the same durable substrate — DSH persists the full session
log, so recovery is possible in principle for either. The difference is
affordance, and it changes shape with model strength. On flash: given
the neutral exam, slice spontaneously ran recall_search → recall_turn
(the tape leaves signposts at every cut) and recovered within the turn;
default searched the workspace, found nothing, and wrote UNKNOWN as
instructed (zero fabrication, duly recorded) — until the explicit tier,
where it zstd-decompressed its own session jsonl and dug the values out.
Pro is strong enough that default performs that forensic dig unprompted —
so the gap moves from whether recovery happens to what it costs: the
same 24/24 takes slice 3 requests (383s / $0.169) and default 32
requests (2014s / $0.461) — 2.7× the price, 5.2× the wall clock.
"Recoverable" and "goes and recovers" are separated by one layer of tools
and signposts; the stronger the model, the more that layer shows up as pure
efficiency.

③ CB-20 precision retrieval · both arms

ContextBench (given a real issue, the agent retrieves the code locations the
fix depends on): a 20-question subset of the official 50-question benchmark.
Paired comparison n=19 — default timed out (20 min) on one question in each
round (different questions; both finished by slice in minutes):

The two generations swap the recall lead (slice +5.5pp under flash, default
+2.8pp under pro), but slice wins F1 and completion on both, and pulls
ahead on precision under pro (+3.2pp); price flips from +11% under flash to
-21% under pro — pro's output is expensive ($1.98/M), and default's
longer sessions and extra steps cost more on an expensive model. The
re-read discipline a bounded slice forces stays an advantage on retrieval
across both generations.

Defects and directions

Install

dsh plugin --profile web add "github:TT-Wang/dsh-slice-agent-loop#main"

Or from a local checkout: git clone then dsh plugin --profile web add .
Restart web afterwards — bundles are composed at boot.

The bundled patch disables the stock loop and compaction — the bounded rebuild
replaces both. If your composition carries an agent-loop-invariant row,
remove it: a rebuilt slice cannot equal the derived history byte-for-byte, and
this plugin refuses to load beside that assertion.

Configuration

Set them from your profile's cordis.patch.yml, targeting the existing row by
id (- id: slice-agent-loop + config:).

Development

npm install --legacy-peer-deps   # the @deepseek-ai/* peers are unpublished
npm run link:dsh                 # symlink them from your dsh checkout
npm run typecheck && npm test

lib/ is committed (git-source installs run no build) — npm run build
before pushing. Real-model smoke: npm run e2e:recall (needs
DEEPSEEK_API_KEY in env).

License

BSD-3-Clause — see LICENSE.