feat: make self-hosting the standing delivery practice (phases 6–7)

[darko-mijic]

What this is

Phases 6–7 make the Protocol usable as a delivery practice in both directions:

• Phase 6 — outward adoption: ship the graph-first authoring/reader surface, package it for an adopter, and measure it on the origin project.
• Phase 7 — inward standing practice: use that same surface on the Protocol itself, turn forward work into Specs, deliberately bind current realization, add inward expected-outcome oracles, and complete one intent → readiness → implementation → verification → slimming loop.

The origin checkout was not modified during phase 7. Its phase-6 working-copy changes remain reverted and survive only in libar-ai/convex-event-sourcing#181.

Phase 6 — outward usage and adoption

• Ratifies verification posture (MD-23): authored verification mode is intent; realization remains derived.
• Generalizes and wires the records gate into the thirteen-leg repository check.
• Ships both agent skills and eleven executable sdp q recipes in the adopter package.
• Teaches the complete graph-first authoring loop and uses all four readiness rungs honestly.
• Adds the package-surface smoke proof and the first second-caller measurements.
• Records the origin adoption evidence: 57 Specs across all four rungs, zero findings, 73.1% fewer gross completion-event tokens, 38.0% fewer non-cached input-plus-output tokens, and 432.27 ms median fresh query derivation.

Phase 7 — inward self-hosting practice

• Ratifies the example realization posture (MD-24): examples normally carry verification evidence, while implemented stays direct and anchor-derived.
• Makes recipe 1 the operational backlog definition: ready ∧ kind≠example ∧ ¬implemented, with excluded-example counts and verifier-health reporting. The raw-query caveat remains explicit.
• Authors forward intent for runtime observation, impact-graph support, Markdown Pack authoring, and kind-neutral oracle eligibility at the rungs their evidence earns.
• Audits all 13 non-example backlog laws and records a named disposition for each; 20 concrete implementation anchors close them honestly.
• Re-measures and re-argues all eight implemented-but-not-stated-ready rows; each remains individually at defined.
• Adds the first inward oracle over spec:consumers.reader, consuming three existing points with three distinct expected outcomes. The technically eligible duplicate-ID target is explicitly refused as vacuous.
• Changes the shared oracle predicate from behavior-only eligibility to any Spec that owns an example space; validator and reader consumers continue to fail closed through the same predicate.
• Adds two ready executable points plus a self-referential rule-kind oracle for spec:validation.oracle-target-eligibility.
• Mutation-probes the behavior-only predicate, the reader expected branch, and the self-referential expected outcome; each targeted suite turns red before restoration.
• Captures a genuine enrichment datum across 019d063 (implementing Design), 341ead6 (implementation), and 7ae8087 (post-green slimming) without generalizing from one case.
• Ignores user-local .vscode/ state, closes the four-column friction ledger without repeated friction, marks plan 24 EXECUTED, and removes the explicitly authorized superseded phase-6 plan file.

Closing graph

• 128 Specs · 1 Pack · 116 anchors → 245 nodes · 462 edges
• Readiness: ready 86 / defined 37 / scoped 1 / idea 4
• Delivery facts: implemented 41 / has-verifier 87 / observed 0
• Operational backlog: 0
• Excluded ready examples: 53, none missing a verifier
• Drift rows: 8, all freshly argued
• Inward models edges: 2
• Validation: 0 errors / 0 warnings

Verification

• Full thirteen-leg npm run check passes in the normal tree and independently at committed checkpoints in a detached clean worktree.
• Vitest 2→4 is isolated in f8a50bb: installed Vitest 4.1.10, replaced removed singleFork with forks + one worker + isolation disabled, and migrated five deprecated matcher aliases without changing assertions.
• Vitest 4 results: 48 files / 602 tests, followed by the dedicated 55-test CLI pass.
• Dependency advisories moved from 8 (3 moderate / 4 high / 1 critical) to 3 (1 low / 2 high / 0 critical).
• Both derived graph hashes are byte-identical before and after the Vitest migration:
  • self-hosting: de3589892d25f40fe38af3b6be638514bc1d9f263b15811853e9770c86168232
  • checkout example: 17a03acf584ebebdf20051f75ed774a1deac5ba3b7f578c3d59bba9ffa7aa790

Deliberate boundaries

• No graph-schema or exported TypeScript API change.
• No Markdown Pack syntax ruling or implementation.
• No runtime-observation producer or impact graph implementation.
• No batch promotion of the eight drift rows.
• No phase-7 work in the origin checkout and no reapplication of its reverted changes.

Post-review follow-up (454e9e9, 4079825)

A full review pass over the branch diff (engine changes read directly; test, spec, doc, and gate
surfaces swept independently) confirmed the kind-neutral oracle predicate is correct and
fail-closed and found four issues, all fixed:

• Recipe portability (the outstanding review finding): recipes 3, 4, 6, and 9 open with a
  subject const naming this repository's corpus. The catalog intro now states the
  parameterization contract explicitly, and recipe 9 names its id line as the parameter — every
  body still runs verbatim here, and an absent id returns { found: false } rather than failing.
• Tautological skill test: the "no colliding global binary" check collected only
  pnpm-prefixed lines and then asserted the pnpm prefix — true by construction. It now scans
  every shell-fenced line and fails on any bare sdp / npx sdp invocation.
• Latent false alarm in the recipe 10 check: its completeness expectation used
  verifiers.length > 0 while the recipe emits rows only for example-declared or enabled
  bindings; the predicates now mirror each other exactly.
• Gate plan selector: zero-padded primary plans (01-–09-) silently never entered the
  highest-primary-plan selection; they now do, while letter-suffixed plans stay excluded.

A second polish pass (4079825) closed the remaining low-severity items and a real usability
failure: the source checkout now has a working full-CLI entry (pnpm --silent sdp --help via a
new sdp package script) and the README documents the exact pnpm exec sdp failure signature on
macOS (Apple's unrelated /usr/bin/sdp demands Xcode); the temporal guard anchors its sweep to
its own directory instead of the caller's cwd, and skips tracked file symlinks whose canonical
file is independently enumerated — CLAUDE.md → AGENTS.md was being swept twice on every run.

Full thirteen-leg npm run check re-run green after the fixes.

https://claude.ai/code/session_0158eQyLnKGzNZTPYcTehw5M

Greptile Summary

The PR establishes graph-first self-hosting as the standing delivery practice.

• Ships adopter-facing authoring skills and eleven executable query recipes.
• Generalizes self-hosting gates and adds inward Specs, anchors, verifier oracles, and readiness evidence.
• Makes oracle target eligibility kind-neutral while retaining fail-closed validation.
• Migrates the test suite to Vitest 4 and adds source-checkout CLI scripts and documentation.
• Completes the prior portability fix by explicitly identifying repository-specific recipe parameters.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains, and the recipe portability issue is resolved by explicitly documenting each repository-specific subject parameter and the required adopter substitution.

T-Rex Logs

What T-Rex did

• T-Rex attempted to run npm run check and observed an ENOENT error for .agents/skills/sdp-agent-surface/SKILL.md, marking a checkout/setup blocker.
• T-Rex noted that no tracked source files were altered, reinforcing that the issue is a setup blocker rather than a changed-behavior defect.

_{Ran code and verified through T-Rex}

Important Files Changed

Filename	Overview
docs/agent-surface/recipes.md	Adds the expanded recipe catalog and now clearly identifies the subject constants adopters must replace.
src/graph/oracle-bindings.ts	Generalizes oracle eligibility to Specs owning an example space.
src/validate/validators.ts	Applies the shared oracle-eligibility predicate in validation.
check-self-hosting-gates.mjs	Generalizes self-hosting checks and corrects primary-plan selection.
check-temporal.mjs	Anchors temporal checks to the repository root and safely handles tracked symlinks.
package.json	Adds source-checkout CLI access and upgrades the testing dependency configuration.

_{Reviews (8): Last reviewed commit: "fix(cli,temporal): give the source check..." | Re-trigger Greptile}

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.