Graph: Skill & Program nodes — query-by-skill + tool sub-identity (LLP 0073–0079)

[philcunliffe]

Extends the structural activity graph with two new first-class node types, turning brittle full-text scans of ai_gateway_messages into cheap graph lookups.

• Session –ran→ Skill (issue Query the activity graph by skill #229): ask the graph skill-shaped questions — "which sessions ran hypaware-ai-improvement-report", "sessions per skill", "which repos a skill runs against" — by joining skill usage to the repos/models/authors already in the graph.
• Session –invoked→ Program (issue Tool nodes stop at the tool name #230): model what a session did with a tool (git, sed, hyp, npm), not just that it touched exec_command. Distinct edge type so the two never collide under one edge_type filter.

Design & settled decisions

• Design: LLP 0073. Plan: LLP 0079.
• LLP 0074 — Claude skill activation = strict three-surface union (Skill tool-call + offset-0 SKILL.md marker + offset-0 slash <command-name>) with a static built-in exclusion list; loose matching rejected on the measured ~23% false-positive rate. (Human-reviewed: the union was chosen over marker-only for coverage, accepting built-in-drift / custom-slash false-positive risk.)
• LLP 0075 — Codex skill activation = path-pattern on the exec_command read of .codex/skills/<name>/SKILL.md; can't share the Claude rule; availability roster ≠ activation.
• LLP 0076 — derive at projection time now (zero gateway/adapter/hook changes); a capture-side skill_activated event is the named future fix, out of scope here.
• LLP 0077 — Program key = validity-gated lowercased basename(argv[0]) of the first command (env-skip, wrapper + bash -lc unwrap, fail-closed); pipeline tails & subcommands deferred to V2.
• LLP 0078 — dispatch_source = additive boolean edge-props (dispatch_tool/dispatch_slash/dispatch_marker/dispatch_shell_read) unioned by mergeRow; kit widened backward/forward-compatibly (edge ids unchanged).

Tasks (all verified-merged)

T1 kit edge-props · T2 Program nodes + invoked (#230) · T3 Claude Skill nodes + ran, 3-surface union (#229) · T4 Codex exec-read Skill derivation · T5 e2e projection + the issue-headline SQL/neighbors query proof. PROJECTOR_VERSION 1→2.

Fixes#229
Fixes#230

Change-Set: graph-skill-tool-nodes

[philcunliffe]

🔍 neutral review — round 1 (Codex + Claude) at 6477026 → request_changes → fixed at e6e307c

Core derivation reviewed sound (134/134 green; the strict Claude false-positive filters — SQL prefix-LIKE + offset-0 regex + role=user/part_type=text — verified to reject mid-text markers, built-ins, and grep-of-SKILL.md). Two real mis-key bugs + three minors found and fixed:

• [major] marker path with a space mis-keyed the Skill node — Base directory for this skill: /Users/John Smith/.claude/skills/hypaware-query minted Skill John. Fixed: marker regex captures the full line (.+, no /m — preserving the leading anchor) then basenames → hypaware-query.
• [major] wrapper flag-args mis-keyed as the Program — sudo -u root git status → Program root, env -C /tmp git → tmp. Fixed: per-wrapper option-with-arg tables (sudo -u, env -C, timeout --signal, stdbuf -o, nice -n, …) consume the value, and any unrecognized -x VALUE shape now fails closed (mints nothing) instead of guessing.
• [minor] Codex skill over-match — echo/rm naming a SKILL.md minted a spurious activation. Fixed: gated on a read-like argv[0] (cat/sed/head/grep/…).
• [minor] nested Codex skill dirs — documented the accepted single-segment-only (fail-closed) scope with a test rather than speculatively widening.
• [minor] cardinality wording — softened LLP 0073/0077 "provably bounded" to "bounded by syntax + observed usage" (hashed basenames like foo.test each mint a node).

After the fix: 150/150 targeted (+16 regression tests) and full npm test 1982/1983 (1 pre-existing skip), typecheck green. Head advanced 6477026 → e6e307c; a round-2 re-review of the delta follows before this is held for merge.

Automated summary of the neutral review (Codex = the independent second family).

[philcunliffe]

🔍 neutral review — round 2 (Codex + Claude) → approve

Re-reviewed the round-1 fix delta at e6e307c: all 5 fixes confirmed correct, 150/150 green, no regression to the round-1-verified behavior (strict Claude offset-0 filters, git commit→git, bash -lc unwrap, dispatch-flag union, edge-id stability, PROJECTOR_VERSION). Adversarial pass found no new mis-key — the fail-closed rewrite strictly mints fewer wrong things.

One minor regression caught and fixed (418c458): the fail-closed wrapper rewrite treated a bare -- (end-of-options) as an unrecognized flag, dropping sudo -- git status to no Program node (a safe-direction miss, not a mis-key). Fixed by consuming -- so the next token is argv[0]; +2 regression tests (sudo -- git→git, env -- duckdb→duckdb), 93/93 facets green.

Verdict: approve. Pending CI green at 418c458, this is mergeable ∧ green ∧ reviewed → will flip out of draft and hold for a human merge (closes #229 + #230).

Automated summary of the neutral review.