Repository navigation
Open
Feature
17 / 3017 of 30 issues completed
Copy link
Labels
needs-humanHuman-in-the-loop required; autonomous sessions must not resolve items carrying this.Human-in-the-loop required; autonomous sessions must not resolve items carrying this.spec-containerSpec container: sub-issues are its work items (/work-items:decompose and /work-items:ship).Spec container: sub-issues are its work items (/work-items:decompose and /work-items:ship).
Description
TLDR
- Generic skills stop shipping model-read language, framework and tool choices; each skill reads the consumer's own records, then infers from the repository.
- Executed adapters, single-tool plugins and security floors stay; an adapter names what it cannot read instead of skipping it.
- Shipped defaults that pick a tool, command or library are retired; unattended lanes run only what the repository states or its build files imply.
- The enforcement ladder becomes the one representation of the talk's pyramid: capability-named rungs, a verification tier in third place, a proportion rule in words, and her order of work.
- The automation-gaps audit reports how a repository's enforcement is spread across the tiers; a review rule holds new enforcement-placing skills to the ladder.
Problem
- Gap: plugins ship per-language guidance and bundled defaults that give wrong or silent results in repositories nobody anticipated, and the ladder has no place for verification.
- Who / how many: every consumer repository; 1,106 ecosystem-specific items across 93 plugins, 584 read only by a model (
recommendations/design-ecosystem-specifics.mdsection 1, in the talk's slice). - Frequency: every run in an unlisted language; every pytest finding codify routes (
plugins/corrections/skills/codify/reference/finding-classes.md:93). - Trend: unknown.
- Cost: wrong tool runs (the bundled Python default runs pyright whatever the repository uses,
plugins/toolchain/reference/ecosystems/python.yaml:14); silent skips; upkeep is low and is not the reason. - Evidence: observed in the repository's files (rated: read directly), plus the workflow inventory (rated: relayed, spot-checked).
Intent
"I would be wary of shipping our skills with specific language implementations because that can all be looked up in the downstream repo ... it really should discover what the language is, what the framework is, what the ecosystem is, what the environment of the downstream system is ... Otherwise, we have to maintain and ship languages and frameworks, and then maintain that as they change."
- note: "I want to make sure that the plugin skills that we ship, whether they're new or existing, align with this pyramid ... which ones are preferred, from strongest to weakest, and which one you should have the most of versus the least of."
- note: "other skills can call other skills through the skill references, not the internals, because that breaks encapsulation."
- note: "As long as we're aligned with her talk, the facts/data/metrics and the latest /discovery:research from official, authoritative, and trusted sources."
- note: "It needs to cover her entire talk, deep dive, everything discussed -> integrated into new/existing plugins/skills ... And dogfooded into this plugins repo itself."
- note: "If there are any of these skills that would operate better, they look at the repo and generate another skill, a CLI, a tool, or a check ... It still remains user repo organization- and machine-agnostic, and it knows where things live ... so it reuses and maintains stuff that it generates and builds."
Goal
A consumer repository in any language gets correct, visible behavior from the generic skills without the marketplace shipping or maintaining that language's tool choices, and every skill that places enforcement follows one ladder that matches the talk's pyramid.
- Intent (why, in the user's words): "Otherwise, we have to maintain and ship languages and frameworks, and then maintain that as they change."
- Vision (where this heads): the whole talk integrated into plugins and skills and dogfooded here, tracked on the wayfind map for the talk.
- When goals conflict, prefer: "single responsibility ... low coupling and high cohesion ... things that change together live together ... semantically accurate names" (the owner's quality bar).
- Finish line: merged — checked by the owner via the widened portability gate passing on
mainand the ladder contract showing the verification tier. (Confirmed by the owner, 2026-11-12: "lets go w/ recommended".)
Work size
- Appetite: not stated; the owner said plan usage is not a constraint.
- Size: large; ten migration slices plus the ladder changes, one plugin per pull request for the guidance cuts.
- Uncertainty: medium; twelve questions are deferred to design and planning.
- Split signals: the slices are independent after the owner documents land; the first slice removes nothing.
Constraints
- No shipped language specifics, applied to existing content, for model-read guidance only; executed code stays.
- Security and safety floors stay; removing or narrowing one is the owner's decision each time.
/corrections:codifyonly picks the rung and proves the check;/implementation:implementbuilds the rule in the consumer's own tooling.- A skill reaches another only by
/plugin:skill, presence-gated with a stated fallback (docs/plugin-philosophy.md:101-116). - A cut to a skill ships only with eval evidence of no loss (
docs/review/skills-and-agents.md:7-9); a whole-page deletion uses a sensitivity control, every other cut keeps the detection control. - Unattended lanes: no install, no text copied from CI files, each command logged as assumed with its source; an unsure run counts as unattended.
- A lane's record pull request holds only commands that pass those limits; "a merged pull request counts as acceptance" is added for ecosystem records only.
- No new organization layer; organization conventions arrive through the repository's own synced standards.
- Mechanism-level decisions are settled at design and planning.
- Quality requirement: repeatability: the same repository run twice gets the same tool choice; target: a recorded choice is read before any inference.
Acceptance criteria
- AC1: No generic skill ships a model-read tool, library or command choice outside the named exceptions, and the portability gate, widened to agents and plugin-level reference and context files, passes.
- AC2: The enforcement ladder has no tool-named rung id, has a verification tier in third position, states the proportion in words with no numeric target, states her order of work (lint first, clean up, strengthen verification, then automate), and its five copies are in sync.
- AC3: Codify no longer routes a pytest
testing/audit/finding to a C#-specific tutorial, and its proposal names the fix-message requirement and hands the build to/implementation:implement. - AC4: Toolchain runs no bundled default command; a repository with no record and nothing inferable shows a visible unresolved skip; a source file outside every ecosystem is named, not skipped silently.
- AC5: An unattended run executes only stated or inferred commands, logs each as assumed with its source, installs nothing, writes no record, and reports a tie as unresolved.
- AC6: A lane's record pull request containing a command outside the unattended limits is refused by a deterministic check.
- AC7: Each whole-page deletion carries a sensitivity-control result; each partial cut carries a detection-control result.
- AC8: Each executed adapter names a file or value it cannot read, and an unlisted mutation tool value is accepted and reported as having no shipped adapter.
- AC9: The automation-gaps audit reports, read-only, a count per ladder tier and the rules that sit lower than they could.
- AC10: A review rule holds every new skill that recommends, places or builds enforcement or an instruction to: read the ladder, prefer the strongest tier that can decide the matter, and discover the consumer's tooling first.
- AC11: This repository's own ecosystem records exist, and its
/toolchain:checkcommand list on a fixed diff is identical before and after the change.
Captured assumptions
- Quality bar: production
- Variation intent: second variant = known (other languages and tools); existing structure: the ecosystem-commands contract, the four-step resolution order, executed adapters in six plugins; upstream: none adopted.
- The sensitivity control has no outside source; it rests on the repository's rule and judgment: revisit if a whole-page cut's control result is disputed.
- The design document's judge scores were read from the workflow result, not from the document: revisit if the design is re-run.
- Her team's numbers are unverifiable and are not targets: revisit if she publishes data.
- Decision records were not written during the interview; six decisions are candidates, not checked: revisit at design.
Out-of-scope
- The no-comments posture; the formal-verification proof of concept; the corrections counter; the live-app exploring skill and its fixture; the dotfiles rule changes (worktree rule, issue-filing wording, lessons check-first): each its own work item, tracked from the talk's wayfind map.
- The sixteen undecided menu items: decision items on that map.
- Moving the ecosystem record's home.
Deferred questions
- Q21: How does a skill that carries no ladder copy read the ladder?, defer until design; arbiter: USER-RESERVED
- Q22: Are the rule-to-enforcer inventory and the per-tier count one artifact?, defer until design; arbiter: USER-RESERVED
- Q23: Does the resolution procedure need its own contract, or does the existing ecosystem-commands contract hold it?, defer until design; arbiter: USER-RESERVED
- Q24: Does the verification tier ship in the same ladder major version as the rung rename?, defer until planning; arbiter: /planning:plan
- Q25: What stack is the live-app fixture and where does it live?, defer until the live-app skill's own work item; arbiter: USER-RESERVED
- Q26: Which models and how many runs per side does each compare use?, defer until planning; arbiter: /planning:plan
- Q27: What agreement level between two blind classifiers counts as a rule being split on?, defer until planning; arbiter: /planning:plan
- Q28: The old rung id's alias lifetime, and the order of the remaining cuts, defer until planning; arbiter: /planning:plan
- Q29: Final public names for the session's working names, defer until design; arbiter: USER-RESERVED
- Q40: Which skill writes the record a lane proposes?, defer until design; arbiter: USER-RESERVED
- Q41: Are toolchain's bundled files deleted, or kept as examples of the record's shape that never run?, defer until design; arbiter: USER-RESERVED
- Q42: Which plugins hold the synced resolution text?, defer until planning; arbiter: /planning:plan
Load-bearing assumptions
| Assumption | Source | Falsifying check | Response if false |
|---|---|---|---|
| The sensitivity control gives a usable answer for a whole-page cut | Brief | Phase 5 Sanity Check: the recorded control outcome is one of the two predicted outcomes | Major: stop Phase 14's whole-page cuts and bring the proof rule back to the owner |
| The design's judge scores are not re-read | Brief | none in this plan; no phase depends on a score | Minor: proceed and note |
| Her team's numbers are not targets | Brief | Phase 6 Sanity Check: the proportion rule's text holds no digit | Minor: proceed and note |
| Six candidate decisions were not recorded during the interview | Brief | Phase 3 Sanity Check: the one ADR file exists; Phase 6 Sanity Check: the 3.0.0 changelog entry states the tier's reason | Moderate: flag and continue |
| No shipped script reads the rung ids or the nine bundled toolchain files | plan (design B12, re-run of the search this session: only skill text, eval files and one test fixture name them) | Phase 6 and Phase 10 pre-flight consumer checks | Major: stop and replan the affected phase |
The lanes run on the opus and sonnet aliases only |
plan (docs/conventions/pr-pipeline.yaml:943, docs/conventions/pr-pipeline.yaml:960) |
each model-run phase's first work item re-reads models.profiles |
Moderate: flag and continue; add the new alias to that phase's compare |
docs/standards/README.md is at standards-contract 1.0.2 while the planning plugin's binding is 1.1.3 |
plan | none | Minor: best-effort grounding; re-run the standards setup |
Execution shape: per-item PRs
Metadata
Metadata
Assignees
Labels
needs-humanHuman-in-the-loop required; autonomous sessions must not resolve items carrying this.Human-in-the-loop required; autonomous sessions must not resolve items carrying this.spec-containerSpec container: sub-issues are its work items (/work-items:decompose and /work-items:ship).Spec container: sub-issues are its work items (/work-items:decompose and /work-items:ship).
Type
Fields
No fields configured for Feature.
Activity
20 remaining items
kyle-sexton commentedon Nov 14, 2026
Hub note, 2026-11-13: one piece of this container's spec has no open child.
Child 15706 is closed as completed on its Done-when list. Its "What lands" section also named an edit to the standards contract (
docs/contracts/standards/README.md: carriers are readers of the paved-path entry, plus the merged-pull-request acceptance sentence for ecosystem records only). That edit is not on main: a search oforigin/mainfor the carrier-reader text finds it only indocs/migration-playbook.mdanddocs/plugin-philosophy.md. It could not land under 15706, whose Done-when forbids touchingplugins/, becausescripts/shared-copies.txt(lines 89, 90, 96) syncs that file into three plugins.Whether an open child (15709, 15729, 15731) already covers it was not checked. Until one does, this container's own acceptance is not met by its children alone.
Agent note by the hub session (GitHub Support), not an owner answer.
chore(ecosystems): add this repository's ecosystem records (#1002)
chore: add ecosystem records mirroring CI's test and link checks (#183)
chore: add ecosystem records mirroring CI's checks (#154)
chore: add ecosystem records mirroring CI's checks (#137)
fix: keep the local gitleaks scan in the cross-cutting record (#184)
fix: keep the local gitleaks scan in the cross-cutting record (#156)
fix: keep the local gitleaks scan in the cross-cutting record (#139)
kyle-sexton commentedon Nov 15, 2026
Decision recorded: proof for a whole-page cut.
Owner, 2026-11-15, asked whether a sensitivity control of the size #15709 ran is good enough proof for the later whole-page cuts: "go w/ recommednations", which is: keep the control, and also require the two-version compare to show no loss for every whole-page cut.
So each child that removes a whole reference page needs both, registered before any run:
sonnetandopus(margin 0.05, no case drop). A cut whose compare shows a loss does not ship on the control alone.Why both: #15709's control was 6 cases at 3 runs a side and its interval ran from -0.12 to +0.08, too wide to carry a cut by itself; its compare over 20 cases was tighter (sonnet -0.05 to +0.04, opus -0.02 to +0.02). The record is on #15709.
kyle-sexton commentedon Nov 16, 2026
Decision recorded: which models the compare for a whole-page cut runs on. This amends item 2 of the comment above.
Two decisions of 2026-11-15 disagreed for these cuts. The proof rule above asks for a no-loss compare on
sonnetandopus. A later decision made compares cheaper: sonnet first at 6 runs a side on the cases whose run reads the changed text, opus only on a new behavior-change case, and the full opus set only if sonnet shows a loss. A whole-page cut adds no new case, so under that shape opus would not run unless sonnet showed a loss.Owner, 2026-11-16, asked which applies (the cheaper shape for cuts too; both models for every cut; or both models on the next three cuts and then the cheaper shape if opus is met on all three): "go with recommendations", which is the third.
So, for the whole-page cuts that follow #15709:
Why three: the evidence that opus does not disagree with sonnet is six evals from one day, and only #15709 among them was a cut.
kyle-sexton commentedon Nov 16, 2026
Decision recorded: the eval that gates a skill-text change. This amends the two comments above for the whole-page cuts.
Owner, 2026-11-16, asked whether to change what gates a skill-text merge (gate each change on one targeted before and after case and drop the per-change no-loss compare, with whole suites run on main on a schedule; or keep the cheaper compare shape on every change; or run no evals until the backlog is zero): "Go with recommendations.", which is the first.
For the whole-page cuts that follow #15709:
Why: on 2026-11-15 the per-change no-loss compares changed no merge decision, and at 3 runs a side a comparison of one text against itself failed 3 times in 4 on sonnet; the targeted cases made every finding.
fix(ecosystems): follow CI's current actionlint fork and gitleaks sca…
fix: cite the current check-managed-files pin and automerge lines in …
fix: cite the current check-managed-files pin and automerge lines in …
kyle-sexton commentedon Nov 16, 2026
Finding on the ecosystem-record items' done-when (#15712 to #15727).
Their third done-when line reads: a recorded
/toolchain:check alloutput contains noskip (unresolved. On toolchain 1.4.0 and 1.4.1 that line cannot fail: a search of the installed 1.4.1 plugin finds no code or skill text that prints askip (unresolvedcell, so every run passes it with or without records.The items closed on 2026-11-16 (#15712, #15716, #15718, #15719, #15722, #15723, #15724, #15725, #15726) were therefore closed on evidence that can fail, stated in each close comment or its linked pull request: every selected ecosystem resolved its commands from the target repository's own
.claude/ecosystems/records, each record validates against the ecosystem-commands schema, and the recordedOverall:line is given with the reason for any FAIL (a linter on the recording machine newer than CI's pin).Open for whoever changes unattended runs to stop using the bundled defaults (this item's later phases): the release that introduces the
skip (unresolvedcell is the first one where this done-when line tests anything, and the remaining items (#15714 and any not yet started) should be re-run against it or have the line reworded.