Open Items

Open Items

This page lists the measures, decisions, and claims about dotclaude that are still open after 0.28.0. Each open item has a reason and an owner. The owner is the user or the next session. Labels: measured means a run on this machine, and reported means a claim from a user.

Summary

• 14 items are open.
• 11 items were closed or made moot by 0.27.0 and 0.28.0.
• 11 claims from users were checked and not acted on.

Open Items

Item	Reason it is open	Owner	What would close it
The context_management hook in a live session	The lab tests pass, but on this machine the built-in guard sec-default loads, because the machine has managed settings, and it skips the hook (measured). Both fixes need sudo (Request build).	The user	Remove the managed settings files, or add dotclaude to managed prependPlugins from a local marketplace folder. Then check the system prompt in a sandbox session.
The mods API can change between releases	The hooks module uses many events of the mods API. The API is documented, but it is new, and a change can break hooks/register.ts with no compatibility layer.	The next session	Run just lab on each new Claude Code release before you move the pin.
The 300K window	The window is a judgment from reports (rot "around 300-400K" on a 1M window, and "200K is the sweet spot"). It is not from a measure on this machine (reported).	The user	A week of sessions, then a look at where the quality drops and where sessions stop.
Usage report after a week on 0.27	No week of 0.27 data exists yet.	The next session	Run just usage --days 7 and compare it with Usage evidence.
A capability eval suite	0.28.0 ported a suite of 16 cases to claude plugin eval (Evals of 0.28.0). It is not the blind, frozen suite that this item asks for.	The user	A blind, frozen suite written before any run, for long sessions, corrections over several turns, and large repositories.
implementer on Opus 5.5 against Sonnet 5.5	The runs on this machine had different tasks (see Model fit). The ProjectArchitect Bench A/B is outside evidence on a different harness with a private hard tier.	The user	Run both models on the same tasks. Anthropic recalibrated the effort levels of Sonnet 5.5, so also re-check the medium setting of the Sonnet agents.
Per-turn effort changes in one long-lived process	No data: each effort value in the transcripts of 2026-10-04 and 2026-10-05 is medium.	The next session	Measure the changes in one long-lived process.
A Bash read of a directory does not load its CLAUDE.md or path-scoped .claude/rules files	Open in Claude Code (#90450). Claude Code 2.1.288 loads rules on Write and Edit.	The user (a fix in Claude Code)	A fix in Claude Code.
The model and effort of the agents	The user set them in the agents/*.md frontmatter on 2026-10-09 from a router search of the removed eval harness and the effort guide of Anthropic. The search, web, bulk, and audit cases did not separate the pairs. The pairs for debug and audit, and each medium that is one step above a tested low, are inferred (Agents).	The user	Harder cases for those roles, run with claude plugin eval (Evals).
The source of the xhigh effort cap	A hook-set max arrives as xhigh. The cause is probably an account limit for each model (unverified).	The next session	A probe on a second account.
A fixed prompt variant after a model switch	The variant stays as chosen at session start, and the user can switch the model with /model. Whether this costs adherence is not measured (System prompt). The eval format has no arm for a variant.	The user	A hand comparison of the variants for each model.
The compact prompt	Claude Code has no route to override it. The settings snippet disables compaction instead (Prompt variants).	None	A route in a later Claude Code version.
The cause of the 30K start in some windows	Research found it only when the settings snippet did not apply (inferred).	The user	A check in the terminal, the desktop app, and the IDE with /context.
The 0.28.0 evals	The output style size, the <text_format> block, and the evidence line have no arm in the ported suite. The reason is that the format has only the plugin arm and the no-plugin arm (Evals of 0.28.0).	The user	A suite and its figures in Evals.

Closed Items

Item	Closed by
The compaction handoff and its retention baseline	0.27.0. Compaction is off, and the user writes each note.
A week with compaction at 150k	0.27.0. The settings turn compaction off.
What the user does after a handoff note	0.27.0. No hook tells the user to run /clear.
The line-break hook (sembr)	0.28.0 adds a new formatter and hooks (Semantic line breaks).
The behavior of the tool.check ask in auto mode on 2.1.292	Measured on 2026-10-07: the classifier allowed the ask of the module, so 0.27.0 added a classic hook for the rm ask. 0.28.0 replaces that hook with ask permission rules and the sandbox.
The stop at the 300K window	Measured on 2026-10-07: the turn ended with blocking_limit, and no compaction ran (Design).
The OpenSpec resume after /clear	Measured on 2026-10-07: /opsx:apply in a fresh session continued at the first unchecked task (OpenSpec history).
Full model id, haiku alias, and the effort clamp	A full id works in agent.spawn. The haiku alias is Haiku 5.5 on the Anthropic API and Haiku 4.5 on Bedrock, Vertex, and Foundry (docs only, no probe there). maxEffortLevel: max causes no clamp, and a hook-set max arrives as xhigh (Decisions).
The session ledger race	0.20.0
The debug escalation of the router	0.28.0 removed the router (design D35). The agent files hold the model and effort.
The shape of a usage-limit hit	The transcripts of 2026-10-04. It is an assistant entry with error set to rate_limit, isApiErrorMessage set to true, apiErrorStatus 429, and the model <synthetic>. The usage report counts this shape.

Claims Not Acted On

Claim	Finding
Two Max 5x plans against one Max 20x	One user reports that Max 20x gives about 1.7 times the weekly usage of Max 5x. Anthropic does not state a weekly ratio. 0.27.0 has no plan detection.
Auto mode's classifier sends the whole transcript on every Bash call	Not checked in the binary.
A Haiku ping keeps the cache warm	Caches are per model, so a Haiku request cannot keep an Opus or Fable cache warm.
A cold cache costs 10 times more on Opus and 40 times on Fable	The official cache prices do not produce these ratios.
A Sonnet session started 8 Fable agents on extra usage	One report. The 0.27 profile leaves Fable out of availableModels.
Opus 5.5 got worse after its release	Public trackers do not agree. One shows 103.8% of the score at release. The Reddit threads of 2026-10-03 report the same. They are in external/2026-10-03/, a local source of the maintainer.
A tool-call batching rule saves about 20% of tokens on Opus 5.5	One user measured it, and a stronger wording made reviews worse. Claude Code already tells Claude to make independent calls together.
Non-Claude model serving (parts 04, 05, 23, and 29 of the external reports)	Quantization, routing, and serving of the models of other vendors are outside the reach of a plugin.
Refusals by a safety classifier	A plugin cannot change the classifier of the model or of the engine.
DensePack	A tool that a user built to save usage. No measurement shows that it helps, and it compresses text with loss.
Domain checklists	A checklist per field, such as security or accessibility. No evidence shows that a checklist raises the pass rate, and the reports are not a spec.

Related Pages

• Usage Evidence
• Evals
• Eval History
• Archive of 0.28
• Archive