Compaction config in cordis.patch.yml is silently ignored on web profiles — preset-mounted dsh-compaction-basic always runs with defaults,

[idkjustgoogleit]

Compaction config in cordis.patch.yml is silently ignored on web profiles — preset-mounted dsh-compaction-basic always runs with defaults, and the 65,536-token headroom default triggers compaction at ~32% of context on smaller models

DSH version: 0.2.0-rc.2 (dsh web, profile web)

Q: What happened?

Compaction fired at ~32% of context (~38.2K of 120K) while my cordis.patch.yml explicitly configured compaction-basic with thresholdRatio: 0.8. Tuning the patch (headroom, ratios, per-model policies) had zero effect — and the behavior appeared to change "randomly" between sessions, because different presets/models resolved different effective thresholds.

Q: Why?

Two stacked issues:

1. The profile-patch entry targets a disabled row. dsh-base defines a top-level row id: compaction-basic, but the web-app layer disables it (- id: compaction-basic / disabled: true in dsh-web-app/cordis.patch.yml). A user patch override on that id merges silently into a row that never mounts. No warning is emitted.

2. The actually-running plugin comes from the agent preset (preset-minimal, preset-standard, preset-ptc, preset-cordis inserts in dsh-web-app/presets/*.patch.yml), and every preset ships compaction-basic with no config — so the plugin's built-in defaults apply. With defaults, the trigger is:

   min(W × 0.8, W − O − 65536)
   

   On a 120K-context local model with 16K reserved output that is min(96000, 38464) = 38,464 tokens ≈ 32% of the window.

Evidence:

• Session logs show compaction events at step boundaries (agent/pre-step pressure path) with metered totals of 31,239 and 14,433 tokens — far below the configured 96,000 threshold. The second fired immediately after the first because the usage-anchored baseline stays above 38,464 right after a compaction.
• The GUI context meter matched the 38,464 threshold at the moment of compaction.
• Editing the two default lines in dsh-compaction-basic/lib/index.js (headroomTokens ?? 65536 → ?? 0) made the 80% trigger work immediately for every preset — confirming the patched config never reached the mounted instance.

Q: What is the expected behavior?

At minimum, one of:

• A warning when a profile-patch override targets a disabled (or non-mounted) row — e.g. patch: entry compaction-basic targets a disabled row, skipping — instead of silently merging into it. The loader already warns on entry not found; a disabled target deserves the same.
• A documented, supported way to configure compaction globally for all presets from the profile patch. Today the only patch lever is overriding each preset row's entire config.plugins (whole-config replacement, no per-plugin merge), which freezes the preset definition and is impractical across four built-ins plus user presets.
• Reconsider the 65,536-token default headroom. For models with windows ≤ ~130K (common for local/self-hosted engines), it silently dominates the ratio and moves the trigger far below 80% — or, with a larger reserved output (W − O − 65536 ≤ 0), disables proactive compaction entirely with only a single warn-once log line. A fraction-of-window default (or falling back to 0 when 0.8 × W is smaller) would behave sanely across window sizes.

Q: How to reproduce

1. Run dsh web (profile web), select the minimal preset, and use a model declared with contextWindow: 120000 and maxTokens: 16000.

2. Add to cordis.patch.yml:

   - id: compaction-basic
     config:
       thresholdRatio: 0.8
       retainRatio: 0.16
       headroomTokens: 0
       maxTokens: 8192

3. Restart the profile and run a long session.

4. Observed: compaction fires at ~38.2K tokens (32%), ignoring the patched config entirely. min(W × 0.8, W − O − 65536) predicts the observed trigger exactly.

Q: Workaround

Edit the plugin defaults in @deepseek-ai/dsh-compaction-basic/lib/index.js:

// before
const headroomTokens = config.headroomTokens ?? 65536;
const maxTokens = config.maxTokens ?? headroomTokens;

// after
const headroomTokens = config.headroomTokens ?? 0;
const maxTokens = config.maxTokens ?? 8192;

One file covers every preset because profiles symlink to the shared install.

Caveat: package updates/reinstalls overwrite this file and restore the old defaults — the two-line diff must be reapplied.

Written by GLM 5.3 Max

[PerryLink]

The override lands on a row that never mounts.

dsh-base defines compaction-basic (packages/bundle/base/cordis.patch.yml:355-356) and the web-app layer disables that row (packages/bundle/web-app/cordis.patch.yml:557-558). The plugin that actually runs is mounted from each preset's config.plugins: presets/minimal.patch.yml:30-31, and the same two lines in standard, ptc and cordis, with no config on the row.

A profile patch keyed by id: compaction-basic is applied by the include patch applier, which matches by id over the composed tree and does not look at disabled; it warns only for a missing id or a name mismatch. So your override is accepted into a row that is never mounted, and nothing warns. That is why editing headroomTokens there did nothing.

The defaults are what you computed. thresholdRatio 0.8 and retainRatio 0.16 (packages/compaction/compaction-basic/src/config.ts:20-23), headroomTokens 65,536 (:75), pressureBudgetTokens = (contextWindow - reservedCompletion) - headroom (:181), threshold = floor(min(contextWindow * ratio, pressureBudget)) (:191-194). For 120K with a 16K reserved completion: 120000 - 16000 - 65536 = 38464, exactly 32%.

What actually changes the running instance today: the preset row, not the base row. The preset-* entry's config.plugins is data, not tree entries, so you cannot patch compaction-basic inside it by id; a profile patch on id: preset-minimal replaces its whole config, which is what the Web preset editor writes (the comment in minimal.patch.yml:1-3 says exactly that). The alternative is to insert your own @deepseek-ai/dsh-agent-preset row with the compaction-basic config you want and select it — the registry README calls that out as the supported way to add a preset or override a shipped one (packages/preset/agent-preset-registry/README.md:48).

On the three asks: a warning when a patch targets a disabled row would have to live in the include applier, which currently has no disabled concept; the "configure globally for all presets" lever does not exist because the plugin is per-preset by design; and the 65,536 default is a flat number, so the fix for small windows is to set headroomTokens on the preset that mounts it.

[ByteRax]

I checked the dsh-v0.2.0-rc.2 source and found one detail worth checking before choosing a workaround:

The web patch disables the top-level compaction-basic row, while the standard preset declares its own instance inside a compaction group. That is consistent with a top-level override not affecting the preset-scoped instance.

However, the minimal preset at the rc.2 source commit does not appear to contain compaction-basic; the current source snapshot does. Could you check the exact installed bundle version, selected preset, and effective composed configuration? A local override or a different build could explain the difference.

Rather than editing the shared installed lib/index.js, the rc.2 web documentation describes editing preset child plugins through the Web editor, which saves to the profile patch. I haven’t tested whether that version’s UI exposes every nested compaction field. If editing a patch manually, please preserve the complete preset configuration and verify the effective composition afterward.

With a 120,000-token window and an effective 16,000-token output reservation, the reported 38,464 threshold follows from the default formula. If setting headroomTokens: 0, also set a positive summary maxTokens; however, choose headroom based on the model and workload rather than assuming zero is safe.

I haven’t reproduced this on your installation, so the version/preset discrepancy remains an observation rather than a confirmed explanation.

[PerryLink]

At dsh-v0.2.0-rc.2 you are right; the row appears later.

I checked the same file at four refs: 639ed015 (dsh-v0.2.0-rc.2) has no compaction-basic row in packages/bundle/web-app/presets/minimal.patch.yml; 687f282e86 (0.2.1-rc.2), 66491b05 (0.2.2-rc.1) and fcc0ad46 have it at :30-31. So the difference is the shipped bundle, not a local override. My earlier line numbers were master; on the 0.2.0-rc.2 tree minimal mounts nothing, which is why the workaround depends on the version.

Which preset ran is in the session, not the patch file: the session meta carries agentPreset (packages/api/session-controller/src/agent.ts:445-448), and a later change appends agent-preset/selected (:410). That is what separates minimal from standard, whose own instance sits inside a group.

For the effective configuration: dsh web --dump-config composes the profile layers without booting the host and without evaluating !!js (apps/cli/src/dump-config.ts:1-7), so the selected preset's config.plugins list is in the output. For the number that actually fired, take contextWindow from the session's request/context event and apply the defaults: threshold = floor(min(W * 0.8, messageBudget - 65536)) (packages/compaction/compaction-basic/src/config.ts:20、:75、:191-194). That reproduces the threshold without a running host.

If you send the request/context line and the agentPreset value, the effective threshold is decidable from those two alone.