Repository navigation
[docs]: route learning capture to editable knowledge sources #806
Description
Description
The learning-capture destination matrix at z-shell/.github main f7ec572 predates editable knowledge-domain sources. It names root AGENTS.md, scoped instructions and runbooks without distinguishing their generation sources. Following the matrix literally can edit generated consumers despite the organization source-ownership rule. The editable source is knowledge/domains/governance/learning-capture.md; runbooks/learning-capture.md is its generated consumer.
The matrix also does not distinguish provisional local findings from promoted durable owners. Its pilot calls for a review after 30 non-trivial tasks or four weeks but supplies no current evaluation record; absence of a record is not proof that no review happened.
Requested correction: route migrated guidance through knowledge/delivery.json to its editable knowledge/domains source, regenerate consumers, and explain that local notes are provisional capture rather than proof of durable promotion. Retain an evidence-backed disposition for useful candidates without requiring an extra artifact or noisy completion report for trivial work or No durable learning.
Scope: correct the canonical destination map and candidate-disposition guidance, verify generated output and instruction paths, and reconcile only findings whose fixing source is directly verified. Keep any unperformed retrospective pilot evaluation explicit; do not invent outcomes from missing records or report the full learning-capture rollout complete. This is a focused follow-up to #668, not a replacement procedures epic. Publication and project-priority changes remain separate decisions.
Select the area that is associated with this issue.
General
Contact Details
Please use this issue.
Metadata
Metadata
Assignees
Labels
Type
Projects
- Status
Activity
ss-o commentedon Oct 16, 2026
Instruction impact review for #806, before changing the source:
The retrospective pilot evaluation is not performed by this destination correction. Missing records remain an evidence gap, not a measured zero. Local note reconciliation will use directly verified fixing source and test results, not prior completion claims.
ss-o commentedon Oct 16, 2026
The local #806 correction is ready for review, uncommitted. The destination matrix now selects editable knowledge-domain sources, preserves the approval boundary for project-native consumers, distinguishes local notes from durable promotion, and records useful-candidate dispositions without adding a hook or a second learning workflow. Only the canonical source and its generated runbook changed.
Verification: knowledge delivery is fresh for 57 consumers; public policy validation, routing inventory, approved skill/project revisions, repository knowledge coverage, Zsh policy validation and source/consumer link checks pass. All 176 policy/routing tests and 21 knowledge tests pass. Pinned Trunk 1.25.0 checks both changed Markdown files with no issues after formatting only the source and regenerating its consumer. The seven-part impact review is in the preceding comment; routing and runtime configuration are unchanged.
The retrospective pilot evaluation has not been performed. It remains an explicit follow-up rather than an inferred zero. No hosted runtime-discovery claim or full-rollout completion is made. This issue stays open pending separately authorized publication; proposed delivery status is In Review, but the actual Project 28 field remains Triage because the enabled auto-close workflow's trigger configuration cannot be inspected through the API.
docs: route learning capture through editable source owners (#808)
ss-o commentedon Oct 17, 2026
The source correction is landed in #808. Migrated destinations now route through their editable knowledge-domain owners, and the runbook distinguishes provisional local findings from durable promotion.
A bounded retrospective has now reviewed 30 real completed task records with sampled readiness receipts from October 10 through October 16. The retained outcomes are 16 reports of learning applied locally or attached to an existing owner, 3 candidate-only dispositions and 11 unknown dispositions. The sample forms 16 conversation-connected clusters, not 30 independent conversations. These are retained reports, not independent proof of every historical implementation or completed publication.
The runbook's other measures have explicit limits:
The original first-thirty/four-week evaluation record was not located within the inspected evidence. This is a frozen convenience sample, not a reconstruction of that original cohort or the globally latest thirty eligible tasks. Full per-task evidence is retained privately; no raw conversations, private paths or runtime configuration are published. Inspection corrected a candidate that had been incorrectly counted as applied and discarded an unsupported claim about archival retention.
The source repair and the requested bounded evaluation record are complete. This does not establish a zero-omission rate or full rollout effectiveness and does not justify a new reminder hook, mandatory extra artifacts or expanded automatic writes. The historical unknowns remain explicit; other open parser/tooling issues are not closed by this evaluation.