[DISCUSS] Turning the simplification audit into issues

[Astro-Han]

PR #3545 removed more than 13,000 lines by collapsing several compaction paths into one authority. We used the same test on the rest of the repository: if a type, state, bridge, or facade disappeared, what real behavior would we lose?

The audit found five larger candidates that look ready for implementation:

• fold RuntimeRunner and the Invocation/Flow shell into RuntimeKernel;
• remove the old Session/AgentRun inspection documents while keeping Session and Turn traces;
• ship one Eval runtime in the npm CLI package instead of two copies;
• reduce the Windows sandbox launcher to the selected --broker-local path;
• narrow maka-cu to the macOS host executor that Maka actually uses.

It also found smaller cases such as write-only projection tables, producerless states, duplicate DTO fields, and private extension points with no consumer. Some remove only a few lines, but they still remove a maintenance contract. Product questions and candidates with unresolved migration or compatibility work are not in this draft.

I do not want to open thirty issues without giving the community a chance to challenge the audit first. The attached draft contains the proposed issue boundaries and the evidence we used.

Full working draft

The proposed rule for each issue is simple: one end-to-end deletion, one PR, and a strict reduction in maintained concepts. Moving the same idea behind a new facade or compatibility bridge does not count.

Please point out any current consumer, compatibility obligation, overlapping work, or behavior that the draft missed. Unless the discussion changes the direction, we will revise the list and open the remaining issues unassigned after 2026-08-26 13:39 UTC.

Codex agents helped trace consumers and prepare the working draft. A human contributor still needs to verify each issue and own any resulting change.

[M4n5ter]

+1 from me

[me2seeks]

+1

[Astro-Han]

Status update before the stated date

Three candidates from the draft have already landed during the discussion window, so they are dropped from the list:

• Collapsing the Invocation/Flow shell into RuntimeKernel — refactor(runtime): collapse runner shell into kernel #3718
• Typed request() as the sole direct Runtime Host operation API — refactor(runtime-host): use typed request as sole operation API #3784
• Removing MAKA_DISABLE_DEFERRED_TOOLS — no longer present on main

Two candidates change shape:

• Narrowing maka-cu is withdrawn from this batch. It targets an external repository and overlaps with feat(windows): harden and productionize the maka-cu Computer Use executor #3785, which hardens the same executor. That direction needs its own discussion before anything is deleted.
• Inlining the External Session importer overlaps with refactor(storage): consolidate external session source parsing #3625, which consolidates external session source parsing over the same path. The issue will say that if refactor(storage): consolidate external session source parsing #3625 lands first, the deletion is re-scoped against its result rather than duplicated.

Five candidates — the retired live permission event protocol, the expired Session-status wire normalization, the optional Daily Review bridge methods, the orphaned Computer Use Core outcome/frame contract, and the unwired Computer Use approval projector — are described in the draft conceptually rather than by their real symbols, and a re-check against current main could not confirm the code they name. Those issues will ask the contributor to confirm the symbols still exist before opening a PR, and to close the issue if the deletion has already landed.

That leaves 28 issues, which we will open unassigned as planned. Each keeps the same rule: one end-to-end deletion, one PR, and strictly fewer maintained concepts.

Claude re-checked every candidate against current main for this update. A human contributor still needs to verify each issue and own any resulting change.

[Astro-Han]

Correction to the previous update

Re-checking the highest-leverage candidates against main found that the draft is wrong in two places, so this batch is three issues rather than 28:

• refactor(runtime-host): remove the uncalled execution.inspect.resolve operation #3932 — remove the uncalled execution.inspect.resolve operation
• refactor(cli): ship one Eval runtime in the npm CLI package #3933 — ship one Eval runtime in the npm CLI package
• refactor(storage): retire the unread Task Ledger and Plan projection tables #3934 — retire the unread Task Ledger and Plan projection tables

Two corrections to the draft:

• The legacy inspection candidate is real but much smaller than stated. Desktop calls only execution.inspect.query with kind: 'turn_trace'; execution.inspect.resolve has no non-test caller. The two files involved are 935 lines together and query stays, so the draft's 2.2k–2.4k estimate does not hold.
• The Windows sandbox candidate is withdrawn. windows-sandbox.ts already has only the --broker-local path. The rejected launch paths live in experiments/windows-sandbox/launcher/ (Rust) and its PowerShell smokes, and touch the W0 evidence lane — that is a separate discussion, not a TypeScript deletion.

The remaining smaller candidates are not being opened as issues for now. The draft stays attached above for anyone who wants to pick one up; opening thirty low-line issues would bury the tracker without making the work more likely to happen.

Claude re-checked these candidates against current main. A human contributor still needs to verify each issue and own any resulting change.

[liuxiaocs7]

+1 for me

Hi, @Astro-Han, I can help work on these related issues.

Thanks!

[Astro-Han]

Closing this discussion

Three issues came out of the audit and are open, unassigned, and ready to claim:

• refactor(runtime-host): remove the uncalled execution.inspect.resolve operation #3932 — remove the uncalled execution.inspect.resolve operation
• refactor(cli): ship one Eval runtime in the npm CLI package #3933 — ship one Eval runtime in the npm CLI package
• refactor(storage): retire the unread Task Ledger and Plan projection tables #3934 — retire the unread Task Ledger and Plan projection tables

If you want one, comment on it and go ahead — no assignment ceremony needed. Each carries its own consumer evidence, and each still needs a human contributor to confirm that evidence before the PR.

The maka-cu narrowing and the Windows sandbox launcher were withdrawn from this batch; both need their own discussion. The remaining smaller candidates stay in the working draft attached to the top post — pick one up and open an issue for it if you want to take it on.

Thanks to @M4n5ter and @me2seeks for the early reads. Note that the list shrank from about thirty candidates to three after a re-check against main, so it is not the same list you saw.