Repository navigation
docs(rfc): Projects — identity at registration, keys land with the PR that reads them - #11828
Conversation
|
Intent: Record the maintainer's two decisions on the thin-Project manifest contract -- |
Opus 4.8 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Verify #11678's head still matches this text at its merge — the RFC now describes a moving PR, and the claims were pinned at [DESIGN-REVIEWED] 12b9d27 |
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
… that reads them The thin-Project manifest carries no id, mcp or memory key in PR A. The registry mints the Project id at add and every install-local key hangs off it; re-adding a registered location returns the existing registration. memory.mode enters the schema with PR B and mcp[] with PR C, because unknown keys are refused by name and admitting a key later is compatible where retracting a parsed-but-unused one is not. Adds open question 9 (identity across installs, decided under PR B), re-audits the front matter and index row at today's main.
ed24c14 to
12b9d27
Compare
|
iamwhatever
left a comment
There was a problem hiding this comment.
Approved after a full read of the diff on the current head: the change matches the body and linked issue, every load-bearing claim was spot-checked against main, the failure paths were traced and fail safe, tests pin the behaviour, and every review-lane item on this head is either PASS or disposed.
… that reads them (kirodotdev#11828) The thin-Project manifest carries no id, mcp or memory key in PR A. The registry mints the Project id at add and every install-local key hangs off it; re-adding a registered location returns the existing registration. memory.mode enters the schema with PR B and mcp[] with PR C, because unknown keys are refused by name and admitting a key later is compatible where retracting a parsed-but-unused one is not. Adds open question 9 (identity across installs, decided under PR B), re-audits the front matter and index row at today's main.
Problem / Motivation
Revision 2 of the Projects RFC (#11564, merged) sketched a
project.yamlcarrying three keys nothing in PR A (#11678) reads: a manifest-mintedid,mcp[]andmemory.mode. The First Principles lane raised the parsed-but-unused keys on three consecutive rounds of PR A, and its retrospective found that the argument the RFC gave for them inverts on inspection: unknown manifest keys are validation errors, so admitting a key later is compatible while retracting a shipped one is breaking. The manifest id had the same shape, justified by the previous attempt's store rather than by any current consumer.Why it matters
project.yamlis committed into users' repositories and read by every install forever. A key that ships before anything reads it becomes a promise the schema cannot take back; an identity that lives in an agent-writable file is one more thing a privileged reader has to distrust. The maintainer decided both: the keys leave PR A's schema and each arrives with the PR that acts on it, and the Project id is assigned at registration. This PR records that decision where PR A's body points to it.What changed (motivation → approach → change)
apiVersion,kind,name,description,sources;memory.modeandmcp[]are shown separately as arriving with PR B and PR C, with the compatibility rule that puts a key in the schema only when something reads it, and its reader side: an install that has not upgraded refuses a newer manifest by name, upgrade is the remedy, andapiVersionmoves only on a retraction or a change of meaningdocs/request-for-change/rfc-crew-projects.md§ What a Project is, revisedidparagraph is replaced: the registry mints the id at add; every install-local key (managed clone, provenance records, digest, member record, PR B's store) hangs off it; re-adding a registered location returns the existing registration; cross-install sameness is not a PR A propertyid:/mcp:/memory:refused by name); PR B and PR C each state that they add their own manifest keymain,implementation-prslists #11678; index row says #7181 is closed and names the three-PR splitdocs/request-for-change/README.md🟩 added · 🟨 changed · 🟥 removed
Tests
Docs only.
scripts/docs_lint.py,scripts/check_comment_history.py(baseorigin/main) andscripts/check_brand_name.pypass; the GPT and Opus local review lanes returned no findings on this diff.Manual verification
Every claim about the implementation was checked against PR A's head
2532e2e8ff:_TOP_LEVEL_KEYSis the five keys named here,ProjectRegistry.addmints the id and matches an existing registration by normalized location, andtest_project_manifest.pypins thatid,mcpandmemoryare each refused by name.Related Issues
Related: #3551 (Projects tracking issue), #11564 (Revision 2, merged), #11678 (PR A, which this text now describes).
no linked issue: the decision is recorded here and implemented in #11678; nothing else tracks it.
Checklist
feat|fix|docs|refactor|perf|test|chore|ci|build|revert: ...)