dsh-glm-mode

dsh-glm-mode

The "GLM Mode" agent preset for DeepSeek Harness, tuned end-to-end for zhipuai/glm-5.3-flash — PTC programmatic tool calling, GLM-specific auto-compaction, plus a reasoning loop guard and a tool spin guard out of the box.

中文 · Releases · Changelog

Install

dsh plugin add dsh-glm-mode

(The npm package name is unique; GitHub source also works: dsh plugin add hoyyang/dsh-glm-mode.)

Zero config required (every switch defaults to on). In a new session pick the "GLM Mode" agent preset and the zhipuai/glm-5.3-flash model.

Why it helps

• PTC tool calling: the full tool catalog is presented as one Code Mode tool — the model composes multi-step TypeScript programs instead of per-call JSON round trips (A/B measured −15% input tokens at equal task).
• Reasoning loop guard: glm-5.3-flash degenerates on long decode at high reasoning effort (same defect as vLLM #56868) — observed live: 131,072 tokens of pure reasoning, one sentence repeated ~2,500 times, 29.5 minutes burned into the maxTokens ceiling. The guard watches the stream for sentence-level repetition and truncates early, reusing the existing "output token limit" flow — reply "continue" to resume.
• Tool spin guard: in PTC mode the model can hallucinate a "top-level write tool" and emit placeholder calls (one live incident: 37 in a row, 23 consecutive steps of no-ops). The guard detects the stall and injects one corrective message through the official channel to put the model back on track — corrective, never truncating a healthy stream.
• GLM-specific auto-compaction: the 1M context auto-compacts at 75% threshold / 22% retention, swapping history for an in-place summary; sessions that skip the preset still get an overflow-level fallback.
• Native multimodal: the preset declares text+image input — attach screenshots directly.

30-second start

1. dsh plugin add dsh-glm-mode
2. Restart DSH (or hot-reload)
3. New session → agent preset "GLM Mode"
4. Model → zhipuai/glm-5.3-flash
5. Just work — both guards run silently in the background

Advanced

All knobs (in cordis.patch.yml; defaults are the recommended values):

Guard audit ledger: ~/.dsh/dsh-glm-mode/guard-events.jsonl (append-only JSONL; the kind field separates reasoning-loop from tool-spin).

How it works

• Presentation: agent-presentation mode: ptc folds the whole catalog into a single run_code tool; the model writes TypeScript and the runtime bridges tools.* calls.
• Reasoning guard: an llm/stream waterfall wraps the chunk stream (zhipuai glm-* routes only), feeding reasoning deltas into a digit-normalized sentence repeater; on trigger it synthesizes finish{max-tokens} and hard-cancels the upstream request.
• Spin guard: llm/stream inspects each step's tool calls (placeholder = run_code with no real tools.* reference and short code); past the threshold, an agent/pre-step decision rewrite appends one corrective user message — a tail append that never touches the request prefix cache.
• Compaction: the preset ships its own isolated compaction owner (official "nearest owner wins" semantics) that coexists deterministically with host engines such as dsh-kiro.

Reliability & acceptance

• Synthetic self-test 52/52 (scripts/selftest-loop-guard.mjs: truncation grammar compliance, route matrix, arming thresholds, fail-open, inject/reset/zero-awareness matrix, stall window, ledger safety)
• Cold-start static checks (boot-check [A]–[E]) all green; uninstall/reinstall idempotency verified (real uninstall → reinstall → preset dir restored)
• Headless fresh-process live smoke test on the real model passed; guards stay silent on healthy streams
• Both guards fail open: internal errors release the original stream with a loud log — they never take the pipeline down

FAQ

Does it only affect GLM models? Yes. Both guards and the compaction policy gate on the route (zhipuai glm-*); every other model is untouched.

Will it break normal answers? The reasoning guard arms only after 2,000 reasoning chars and ignores sentences under 48 chars; a spin-guard false positive costs one corrective message and never truncates anything.

How is this different from the built-in Kiro mode? Kiro mode is tuned for DeepSeek-family models. This plugin gives glm-5.3-flash its own preset (PTC + 1M compaction at 0.75/0.22) and the guard family. Both can coexist — compaction ownership yields deterministically per official semantics.

Where do the wire settings (thinking format / tool_stream) live? In your settings.yaml provider route (the officially recommended z.ai format), not inside this plugin — install and go.

Build locally

pnpm install
npm run build      # tsc compile + npm pack
npm run selftest   # 52 synthetic self-test checks

License

MIT — see LICENSE.