docs(handover): the autonomous run is done — nine tasks, four questions - #44
Open
codeitlikemiley wants to merge 1 commit into
Open
docs(handover): the autonomous run is done — nine tasks, four questions#44codeitlikemiley wants to merge 1 commit into
codeitlikemiley wants to merge 1 commit into
Conversation
§2 moves to `4e1ce9c` and 109/101/8, recounted programmatically. M14.9 arrived from another session mid-run. §3 records the one thing blocking the rest — M14.8's decision, which tasks 2 and 3 have already mostly made — and names a pattern the run found three times: a protocol or manifest field that reads like a control and controls nothing. `NetPolicy::allow` granted every host when asked for one. `plugin.toml`'s `net:` was rendered at the consent prompt and reached no sandbox. `CallMeta.task` is promised by docs/12 and read by nobody. Each was individually latent; together they say the policy layer has been written ahead of the enforcement layer more than once, and nothing catches that class automatically. A lint that fails when a policy field has no reader would have caught all three. §6 is rewritten around what is actually left: four questions, then the dogfood, then the hardware-blocked list. The training *gates* are executable now (#40), so a model can be judged the day one exists — which is the point of shipping gates and not scores. Also recorded: the two known flakes with what triggers them, and what this machine does to the suite when another project builds alongside it — gates ran 60 minutes to 3.5 hours against a usual 15, and a parent process at 0% CPU is not evidence of a stall. `RUN-REPORT.md` is the full log: every task, every bug found outside the list, every test deliberately broken to prove it fails, the mistakes I made, the deviations I took and why, and what remains unverified. #39 is superseded by it. Claude-Session: https://claude.ai/code/session_017kFpYDqvz6sKGSkM4YKaRf
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes the run brief in
GOAL.prompt.md. All nine tasks shipped or resolved.109 milestones, 101 shipped, 8 open, recounted programmatically.Four questions waiting for you
All four are at the top of
RUN-REPORT.md, each phrased to be answerable in a word:CallMeta.task— honour it, or remove the field and docs/12's promise?The finding I'd most want you to see
Three fields, found one per task, only visible as a class in hindsight — each reads like a control and controls nothing:
NetPolicy::allowplugin.tomlnet:SandboxPolicyCallMeta.taskEach was individually latent. Together they say the policy layer has been written ahead of the enforcement layer more than once. A lint that fails when a policy field has no reader would have caught all three — that's the durable fix, and it isn't written.
And one thing that was reported and isn't true
A scoping agent claimed
nscurl -bgescapes the macOS sandbox viamach-lookup, with 559 bytes exfiltrated. Tested against the real generated profile: rc=139, no file. The 559 bytes matches my unsandboxed control exactly — its probe wasn't sandboxed. No vulnerability. Recorded so nobody "fixes" it later.Also in the report
Every test I deliberately broke to prove it fails (seven, all restored green), the mistakes I made during the run — a
git reset --softthat folded the report into a milestone commit, a CI red on Linux-only dead code, a test of mine that raced a deliberate design — the two deviations I took from the brief and why, and what remains unverified.Plus, practically: this machine stretched gates from ~15 minutes to 60–210 because another project built alongside and UTM/QEMU held ~127% of a core for seven hours. Nothing was wrong with the code, and a
cargo test --workspaceparent at 0% CPU is not evidence of a stall.Supersedes #39.
https://claude.ai/code/session_017kFpYDqvz6sKGSkM4YKaRf