diff options
| -rw-r--r-- | claude-rules/knowledge-base.md | 2 | ||||
| -rw-r--r-- | docs/design/2026-07-13-runtime-portability-inventories.org | 58 | ||||
| -rw-r--r-- | todo.org | 18 |
3 files changed, 71 insertions, 7 deletions
diff --git a/claude-rules/knowledge-base.md b/claude-rules/knowledge-base.md index 5658498..d8aa6a6 100644 --- a/claude-rules/knowledge-base.md +++ b/claude-rules/knowledge-base.md @@ -53,7 +53,7 @@ Pull before writing, commit and push after (`git -C ~/org/roam add -A && git com ## Capture, then promote -Harness memory (`~/.claude/projects/<enc>/memory/`) remains the per-project capture layer: fast, automatic, allowed to be at-risk. At wrap-up (or a task audit, or an explicit prompt), promote facts that meet the inclusion bar into the KB as nodes. The wrap-up workflow asks; answer it honestly — promotion discipline is what keeps the capture layer from silting up. +Harness memory (`~/.claude/projects/<enc>/memory/`) remains the per-project capture layer: fast, automatic, allowed to be at-risk. An agent running without harness memory (a non-Claude runtime) captures into its session log instead and promotes from there — the promotion discipline is the same. At wrap-up (or a task audit, or an explicit prompt), promote facts that meet the inclusion bar into the KB as nodes. The wrap-up workflow asks; answer it honestly — promotion discipline is what keeps the capture layer from silting up. ## Inventory diff --git a/docs/design/2026-07-13-runtime-portability-inventories.org b/docs/design/2026-07-13-runtime-portability-inventories.org new file mode 100644 index 0000000..98fe74d --- /dev/null +++ b/docs/design/2026-07-13-runtime-portability-inventories.org @@ -0,0 +1,58 @@ +#+TITLE: Runtime Portability — Hook, MCP, and Memory Inventories +#+AUTHOR: Craig Jennings +#+DATE: 2026-07-13 + +Decision-prep for the generic-agent-runtime arc (spec: [[file:2026-05-28-generic-agent-runtime-spec.org]]). Three of the eight gap child tasks are inventories; this doc is their deliverable. Each section ends with a verdict and any decision left for Craig. + +* Hook parity inventory + +What's wired today (global =.claude/settings.json=, symlinked to =~/.claude/settings.json=): + +- PreToolUse AskUserQuestion hard-deny — blocks the popup-choice tool, forcing inline numbered options per =interaction.md=. Exists because prose alone failed. +- PreCompact =precompact-priorities.sh= — saves priorities before context compaction. +- SessionStart =session-title.sh= (cosmetic title) and =session-clear-resume.sh= (resume-after-/clear plumbing). +- Stop =ai-wrap-teardown.sh= — sentinel-gated session teardown. + +Shipped but unwired (reference copies in =hooks/= + =settings-snippet.json=): =git-commit-confirm.py=, =gh-pr-create-confirm.py=, =destructive-bash-confirm.py=. Not active under Claude either, so they impose no parity requirement today. + +Language bundles wire PostToolUse validators per edit (=validate-bash.sh=, =validate-el.sh=, =validate-go.sh=, plus python/typescript equivalents) — and every bundle also ships the same validation as a =githooks/pre-commit=. + +Per-hook portability verdicts: + +- AskUserQuestion deny: moot off Claude. The popup tool is Claude Code's; Codex-style and local harnesses have no equivalent tool to deny. No port needed. +- PostToolUse validators: covered off Claude. The bundles' git pre-commit hooks run the same validators, so enforcement survives at commit time instead of edit time. Acceptable downgrade, no work needed. +- SessionStart clear-resume: not portable as-is; /clear semantics are Claude's. This is the same surface as the "session plumbing per runtime" child task — resolve it there, not as a hook port. +- session-title: cosmetic, skip on other runtimes. +- PreCompact priority-save: genuine gap. Codex-style harnesses compact without a hook point. Downgrade is prose (the session-context write triggers in protocols.org already say "save before compaction"), which is weaker but is also the pre-2026 Claude behavior. Accept prose, or shorten sessions on those runtimes. +- Stop wrap-teardown: partial gap. Codex has a turn-completion notify command that could run the same sentinel-gated script; local harnesses vary. Port the script trigger where a hook point exists; where none does, teardown becomes a manual step in the wrap phrase. + +Verdict: only PreCompact and Stop carry real porting work, and both have acceptable degraded modes. Nothing here blocks a pilot. + +* MCP portability check + +Locally-configured servers (user scope, =~/.claude.json=): linear, notion, figma, slack-deepsat, google-calendar, google-docs-personal, google-docs-work, drawio, google-keep. No project-scope servers anywhere. All are ordinary MCP server definitions, portable to any MCP-speaking harness (Codex CLI reads them from its config; most local harnesses that matter speak MCP now). Port = translate the server blocks + carry the auth material. Mechanical. + +claude.ai-managed connectors (session-injected, not in local config): Gmail, and the claude.ai variants of Calendar/Drive. These do not travel. Overlap analysis: calendar and docs are already covered by the locally-configured servers; Gmail's workflow-preferred path is already =cmail-action= (local CLI), so the managed Gmail connector is convenience, not dependency. + +Named gap: signal-mcp. protocols.org "Paging Craig" names it as the only supported away-from-desk page path, and it is not in any local config — it lives claude.ai-side. Off Claude there is currently no page channel at all. This lands squarely on the open Signal-pager task ([#C], todo.org): the signal-cli runbook that task produces would be the runtime-neutral paging path. Recommend noting runtime-portability as an added motivation on that task. + +Verdict: portable except paging; paging's fix is already a filed task. + +* Memory story for non-Claude agents + +Claude auto-memory (=~/.claude/projects/<enc>/memory/=, MEMORY.md index) is harness-owned: written by Claude's memory tooling, loaded by its session start. A non-Claude agent should neither read nor write it. + +The designed cross-agent store already exists: the org-roam KB (=knowledge-base.md= — reading is plain rg over files, writing is one node per fact, both runtime-neutral). Project state lives in file artifacts every agent reads anyway: todo.org, notes.org, session anchors, docs/. + +One wording gap: knowledge-base.md's "Capture, then promote" section names harness memory as the capture layer. For a non-Claude agent the capture layer is the session log itself. Recommended one-sentence addition there: an agent without harness memory captures into its session log and promotes from it at wrap-up. Shared-asset edit, needs approval. + +Verdict: no build. One approved sentence in knowledge-base.md closes it. + +* Decisions for Craig + +All four approved by Craig, 2026-07-13: + +1. PreCompact downgrade off Claude: prose-only "save before compaction" accepted. +2. Stop-teardown off Claude: port via Codex notify where available, manual teardown elsewhere. +3. Runtime-portability motivation note added to the Signal-pager task. +4. knowledge-base.md capture-layer sentence added (non-Claude runtimes capture into the session log and promote from there). @@ -85,6 +85,9 @@ RECONCILE FIRST: =protocols.org= "Paging Craig" already documents an agent-pagin *** 2026-07-13 Mon @ 05:16:37 -0500 Folded home's ownership ack + current-state report into the reconcile scope home confirmed (2026-07-11 reply) rulesets owns this task and the two-paths reconciliation is the right first step. New facts for the reconcile: from a home session on 2026-07-09, signal-mcp was NOT connected, and the local signal-cli is registered as Craig's own number — =send --note-to-self= returns a message id but produces no phone push. So home currently has no working ad-hoc page channel at all; whatever the runbook lands on must give home a live path. +*** 2026-07-13 Mon @ 14:40:00 -0500 Added runtime-portability as a second motivation (Craig approved) +The MCP portability inventory ([[file:docs/design/2026-07-13-runtime-portability-inventories.org]]) found signal-mcp exists only claude.ai-side — no local config anywhere — so a non-Claude agent (Codex-style or local LLM) has no paging path at all. The signal-cli runbook this task produces is therefore also the runtime-neutral page channel, not just home's replacement for ntfy. + ** TODO [#C] KB orphan-node review pass :chore: :PROPERTIES: :CREATED: [2026-07-01 Wed] @@ -263,8 +266,8 @@ An install target that emits the non-Claude entry file (=AGENTS.md= for Codex-st *** TODO Skill parity across runtimes The user-invoked skills (voice, review-code, flush, …) are Claude Code registrations wrapping portable markdown prompts. Decide per skill: re-register in the target harness's custom-command system, or fold into =.ai/workflows/= (already runtime-neutral). Model-invoked skills need a per-harness answer or graceful absence (commits.md already defines the /voice-unavailable fallback pattern). -*** TODO Hook parity inventory -Inventory every settings.json hook (AskUserQuestion hard-deny, PostToolUse validators, wrap-teardown Stop hook, session-clear-resume) and map each to the target harness's mechanism or an accepted prose-only downgrade. The deny hooks exist because prose enforcement failed — a downgrade is a real regression, so decide per hook, not wholesale. +*** 2026-07-13 Mon @ 13:34:17 -0500 Hook parity inventoried — only two hooks carry real porting work +Full mapping in [[file:docs/design/2026-07-13-runtime-portability-inventories.org]]. AskUserQuestion deny is moot off Claude (no popup tool to deny); PostToolUse validators survive via the bundles' git pre-commit hooks; clear-resume folds into the session-plumbing child; session-title is cosmetic. Real gaps: PreCompact priority-save (prose downgrade) and Stop wrap-teardown (Codex notify / manual elsewhere) — decisions in the VERIFY below. *** TODO Launcher runtime flag =ai --runtime <claude|codex|local>=. The =AI_AGENT_ID= shape already carries a runtime segment (=host.project.runtime.epoch=). Keep the change small and contained per the 2026-06-24 helper-task caveat — the helper's =ai --helper= touches the same launcher. @@ -272,11 +275,14 @@ Inventory every settings.json hook (AskUserQuestion hard-deny, PostToolUse valid *** TODO Session plumbing per runtime =/flush auto= self-inject types Claude-specific keystrokes (=/clear= + resume line); clear/resume mechanics differ per harness. Also confirm =session-context-path= and the suspend/startup anchor cycle behave identically when a non-Claude agent drives them. -*** TODO Memory story for non-Claude agents -Claude's auto-memory is harness-owned. The org-roam KB is the designed cross-agent store — document that a non-Claude agent skips harness memory and leans on the KB + session files, and verify knowledge-base.md reads cleanly for a non-Claude reader. +*** 2026-07-13 Mon @ 13:34:17 -0500 Memory story confirmed — KB is already the cross-agent store +Details in [[file:docs/design/2026-07-13-runtime-portability-inventories.org]]. Auto-memory stays Claude-owned; non-Claude agents use the org-roam KB + file artifacts (todo/notes/session anchors), all runtime-neutral already. One wording gap: knowledge-base.md's capture-then-promote names harness memory as the capture layer — a one-sentence addition (session log as the capture layer for agents without harness memory) closes it, pending approval in the VERIFY below. + +*** 2026-07-13 Mon @ 13:34:17 -0500 MCP portability checked — portable except paging +Details in [[file:docs/design/2026-07-13-runtime-portability-inventories.org]]. Nine locally-configured servers (linear, notion, figma, slack-deepsat, google-calendar, google-docs x2, drawio, google-keep) port mechanically to any MCP-speaking harness. claude.ai-managed connectors (Gmail + claude.ai Calendar/Drive) don't travel but are redundant with local servers / cmail-action. The real finding: signal-mcp is NOT locally configured anywhere — the protocols.org paging path is claude.ai-side only, so off Claude there is no page channel. The Signal-pager [#C] task's signal-cli runbook is the fix; add runtime-portability as motivation there (VERIFY below). -*** TODO MCP portability check -MCP is cross-vendor; the self-hosted servers (signal-mcp, google-calendar, linear, slack) should attach to any MCP-speaking harness — verify. The claude.ai-managed connectors (Gmail, Drive, Calendar) don't travel; list self-hosted replacements where a workflow depends on them. +*** 2026-07-13 Mon @ 14:40:00 -0500 Four inventory decisions — Craig approved all four +Recorded in [[file:docs/design/2026-07-13-runtime-portability-inventories.org][the inventories doc]]: prose-only PreCompact downgrade off Claude; Stop-teardown via Codex notify where available, manual elsewhere; runtime-portability note added to the Signal-pager task; capture-layer sentence added to knowledge-base.md (non-Claude runtimes capture into the session log and promote from there). *** TODO Local model floor evaluation The workflows assume long-context instruction-following (startup's multi-file read; the commits.md publish chain). Establish the minimum viable local tier (likely strong-70B+/MoE, 100k+ context), and what compensations a weaker model needs: shortened protocols, more checklist gates, more hook-level enforcement. Feeds the spec's "default local runtime/server" and "first supported local editing CLI" blocker decisions. |