agent-harness-settings-parity
Skill3 starsChanged 53 days agoMITView on GitHub
Audit and apply one policy — model, thinking effort, compaction threshold, approval mode — across every installed coding harness, each with its own config file, knob name, and unit. Use when a setting must hold everywhere or before claiming it is already set. Not for a single-harness edit.
In this repository59 other files it gives its agents.
- AGENTS.mdAGENTS.md
- agent-operating-contract-evolutionSkill
- ai-coding-agent-delegationSkill
- arch-improvement-planSkill
- contract-bound-deliverablesSkill
- cross-agent-repair-routingSkill
- define-goalSkill
- delegated-work-acceptanceSkill
- design-an-interfaceSkill
- design-mdSkill
- discord-rich-messageSkill
- dotnet-repo-workflowSkill
- evidence-backed-analytical-writingSkill
- evidence-boundary-disciplineSkill
- find-nearbySkill
- gddSkill
- generated-runtime-repairsSkill
- ghc-searchSkill
- git-auto-sync-pipelinesSkill
- github-code-reviewSkill
- github-repo-managementSkill
- grounded-citationsSkill
- harness-engineeringSkill
- hermes-context-file-loadingSkill
- hermes-modelsSkill
- hermes-updatesSkill
- hoa-introspect-distillSkill
- hoa-introspectSkill
- hoa-skill-reviewSkill
- hoa-skill-workSkill
- html-artifactSkill
- jupyter-live-kernelSkill
- kanban-workflowsSkill
- m365-copilot-cdpSkill
- measurement-campaign-evidenceSkill
- microsoft-officeSkill
- notes-to-prdSkill
- ocr-and-documentsSkill
- official-document-verificationSkill
- omp-headless-orchestrationSkill
- ops-claim-verificationSkill
- package-release-verificationSkill
- plan-to-issuesSkill
- project-phase-closeoutSkill
- report-claim-auditSkill
- research-stream-consolidationSkill
- review-revision-candidatesSkill
- rss-feedsSkill
- skill-library-curationSkill
- skill-porting-and-publishingSkill
- spec-convergenceSkill
- structural-validity-gatesSkill
- subagent-driven-developmentSkill
- systematic-debuggingSkill
- teams-cdpSkill
- tmux-agent-controlSkill
- whisperSkill
- writing-plansSkill
- writingSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 28
Before claiming a setting is already in place anywhere — the audit half is the point.
- Very specific: line 35
Find each harness's own knob from its own docs or CLI: Never map one harness's key name onto another: the same concept is thresholdPercent …
- Distinctive: line 58
Every one of these harnesses reads config at process start: a running TUI keeps its old value.
- Distinctive: line 86
When you must touch a package file, say so in the report and keep the backup so the edit can be replayed after an upgrade.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Agent harness settings parity
An operator runs several coding harnesses on the same machine (for example omp/oh-my-pi, pi, codex, opencode, dsh) and expects a policy — model, thinking effort, compaction threshold — to hold on ALL of them. Each harness has its own config file, its own knob name, and its own unit for the same concept. Treat this as a repeatable audit-and-apply task, never as one edit.
When to use
- "给所有 harness 配上 X" / "check whether every harness still has X" / "apply this on all machines".
- Any policy that must hold across several harnesses at once: model, thinking effort, compaction threshold, approval mode, telemetry.
- Before claiming a setting is already in place anywhere — the audit half is the point.
Procedure
- Inventory what is installed before touching anything.
command -vfor each CLI, then locate the real binary (version-manager installs and npm globals are usually NOT on PATH). Record each config root. - Find each harness's own knob from its own docs or CLI. Never map one harness's key name
onto another: the same concept is
thresholdPercent(omp),buffer(opencode),reserveTokens(pi),model_post_turn_compact_threshold_percent(codex),thresholdRatio(dsh),compression.threshold(the gateway harness). - Read the current value and its default first. "The operator already set this" is a hypothesis; most often nothing was ever written on this host. Report what was actually there.
- Write via the harness's own config command when one exists, otherwise edit the file with a backup into the current task dir. Then read the value back — a successful write is not a verified setting.
- Verify with a parse/dump command, not by re-reading your own edit:
codex doctor("config.toml parse ok"),dsh --profile <p> --dump-config,omp config get <key>, or the gateway harness'sconfig get <key>. If a harness has no readback, say so instead of claiming success. - Prove a value that only shows up at runtime from the session transcript. Per-model settings
(
thinkingLevel, resolved compaction threshold) are not visible in the config file alone and the TUI banner shows only the model id. A fresh omp session writes its resolution into its session JSONL as{"type":"thinking_level_change","thinkingLevel":"max"}— read the newest file and quote that line as the proof. - Report a per-harness table: key, before → after, and the exact verification output. Grade anything you could not verify as unverified in the same table.
Restart policy
- Every one of these harnesses reads config at process start: a running TUI keeps its old value.
- Never restart a harness that is mid-turn on real work. Apply the config, let the turn finish,
then relaunch with the resume flag (
omp --resume,dsh --resume <session>). Losing an hour of in-flight work to prove a setting is a bad trade. - Prefer a fresh session over a resume when the goal is to make a config change live. A resumed session carries the value it resolved at its first start; only a new process re-reads config. When the next piece of work starts anyway, start it in a new session and let the setting load for free (the delivery pipeline's post-merge reset does exactly this).
- A gateway restart kills the live conversation turn and its background watchers. Config reaches new sessions anyway, so restarting is opt-in, stated explicitly with its cost.
- Harnesses that are not running need no restart; say so instead of listing them as pending.
Pitfalls
- A "reserve/buffer" knob is not a threshold.
buffer,reserveTokens,autoCompactBufferTokensare subtracted from the window: threshold = window − reserve, so 50% means reserve = half the window, and the same absolute reserve means something different on every model window. Prefer a percentage form when the schema accepts one. - Patch/write tools may refuse a gateway config file by design. Use the harness's own
config set <dotted.key> <value>; it edits and validates the file for you. - Some settings live in nested realms, not the top-level config. dsh is the hard case: the host
compaction-basicrow isdisabled: truebecause each agent preset owns its own isolate realm, so a host-layer patch changes nothing for TUI/web sessions. Patch the preset copy the session actually uses (${DSH_HOME:-$HOME/.dsh}/.agent-presets/<preset>/agent.cordis.yml). - Know which files an upgrade will overwrite. Preset/config files inside an npm or
version-manager package directory get reset on update; user-layer copies
(
${DSH_HOME:-$HOME/.dsh}/.agent-presets/,${OMP_HOME:-$HOME/.omp}/agent/config.yml,${PI_HOME:-$HOME/.pi}/agent/settings.json,${CODEX_HOME:-$HOME/.codex}/config.toml) survive. When you must touch a package file, say so in the report and keep the backup so the edit can be replayed after an upgrade. strings <binary>is a last resort. It returns megabytes of unrelated keys and eats context. The vendors publish real docs — read those first and usestringsonly to confirm a key name you already suspect.- A managed preset may be rewritten by its owner package. dsh TUI presets carry
.dsh-tui-managed.json; an edit there is legitimate but can be replaced on a preset update, so note it as such rather than treating it as permanent.
References
references/context-compaction.md— per-harness compaction knobs, defaults, and the exact verification command for each, plus the dsh realm map.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

