Brainstorming docs on future extensions#748
Open
ajonkisz wants to merge 18 commits into
Open
Conversation
…y, harness gaps - docs/design/mission-control.md — global scope at the top of the sidebar: global sessions/staff, the `bobbit` REST meta-tool, Global System Staff (Caretaker/Archivist/Improver/Observer), flight-recorder audit feed, phased core-vs-pack implementation plan. - docs/design/autoimprovement.md — recursive self-improvement with earned autonomy: Hermes-style per-goal review + curator, Claude-style dreaming gates, confidence-scored trust ladder with shadow-mode calibration, auto-approval with full audit + revert + automatic demotion. - docs/design/client-performance-battery.md — instrumentation-first battery plan: ambient infinite animations, redundant polling, full-tree render cadence; measurement harness, tiered fixes, battery-saver mode. - docs/design/harness-gap-analysis.md — feature/UX gaps vs openclaw, hermes (masters) and the Claude Code mirror, with a prioritized easy-wins roadmap. https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
… docs - docs/design/fable-program-execution-plan.md (new) — single sequencing source of truth across all fable-docs workstreams: binding rules (reuses extension-platform-implementation-plan §0 as universal DoD), doc-is-the-spec deviation protocol, shared-seam coordination, lane dependency graph, full PR slicing with effort + mergeability rationale, and a master checklist for the orchestrating session. - mission-control.md Appendix A — exact constants/storage layout, PersistedStaff.global validation matrix, complete bobbit meta-tool operation catalog v1 (op → tier → REST), ActivityEntry interface + REST/WS contract, mission-control pack layout, owned files per phase. - autoimprovement.md Appendix A — ImprovementProposal/ChangeClass types, JSONL/snapshot/lock storage layout, REST table, tool + rubric placement (pack-owned, never hardcoded), settings keys with ceilings, owned files per phase. - client-performance-battery.md Appendix A — complete classified inventory of every infinite animation (incl. goal-dashboard.css), animation-power.ts contract, perf report shape, poll-site checklist, PR boundaries. - harness-gap-analysis.md — concrete R2 trigger schema (one-shot `at`, gate_failed/session_errored) so trigger work cannot drift. https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
- fable-program-execution-plan.md — expanded to cover ALL eight workstream docs on this branch, not just the four newest: adds the comms-stack (CS) and sidebar/goal-nesting (SN) backlogs as lanes, enumerates every EP and CE sub-goal in the master checklist, adds a shared-seam serialization table (notably: SN-T2/T3 before PB-P2a on api.ts refreshSessions; SN-T5 before PB-P2c on the sidebar render paths; CS merge-map governs session-manager.ts), a Day-1 parallel starter set, and a §6 doc-format contract that defines parity for any future design doc. - extension-platform.md — banner: implementation plan is execution authority, incl. its §0.2 entity-loading refinement superseding §4. - extension-platform-implementation-plan.md, token-cost-efficiency.md, sidebar-goal-nesting-audit.md, comms-stack/04 — Status + workstream pointers to the execution plan; SN gains explicit seam-overlap warnings against the battery plan. https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
…tion plan Brings every workstream to extension-platform-implementation-plan.md grade (the program's gold standard): per-goal Outcome / Owned files / symbol-anchored Steps / tests-authored-first with RED expectations / Acceptance, plus anchor-baseline stop-rules and precision policies. - mission-control-implementation-plan.md (new) — MC-P0..P5; P0–P3 at file/function level (PersistedStaff at staff-store.ts:60, participatesInVisibleOrder pin, renderSidebar insertion, meta-tool registry + catalog pinning tests), P4/P5 contract level. - autoimprovement-implementation-plan.md (new) — AI-P1..P6; proposal store/panel plugin (PROPOSAL_TYPE_REGISTRY), judge + learned-skills pack, Improver with tool-whitelist pins, curator/dreaming/shadow/ autonomy at contract level with ceiling clamps pinned in tests. - client-performance-battery-implementation-plan.md (new) — PB-P0..P3 all at file/function level, with measurement discipline (no behavior-affecting merge without before/after harness numbers) and SN seam guards. - gap-easy-wins-implementation-plan.md (new) — GA-R2/R3/R4/R5/R6/R9, incl. the shared-seam rule that GA-R2's triggers land exactly once. - fable-program-execution-plan.md — §2.1 THE ORDER: wave diagram (W0 prereq → W5 loop closes) with per-wave parallel sets, blocking rules, user-visible milestones per wave, and the orchestrator's rule of thumb; workstream table now has a design vs execution- authority split; lanes point at the new plans. - mission-control.md, autoimprovement.md, client-performance-battery.md, harness-gap-analysis.md — execution-authority banners (the extension-platform.md pattern). https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
…stics) - code-intelligence.md (new) — layer model (lexical/structural/symbol/ spatial/truth/temporal), design stance from owner interview: native first-party tool groups over MCP surfaces; everything behind swappable extension-platform capability names; per-worktree derived state with a relative-paths pin (the graphify cross-worktree lesson); language servers as autodiscovered per-language packs; graphify as optional visualization pack, in-house budgeted code_map for the prompt path; context-services chip answering "what's running, where" from chat. - code-intelligence-implementation-plan.md (new) — CI-1 ast_search/ ast_rewrite (binaries/ shipping, preview-diff rewrites), CI-2 code_check structured diagnostics, CI-3 LSP supervisor + code_* tools with graceful-degradation invariants, CI-4 lsp-<lang> descriptor packs + autodiscovery UX, CI-5 budgeted repo map, CI-6 services chip + graphify pack, CI-7 capability shim → EP G8 swap + BENCH exit gate. - fable-program-execution-plan.md — CI workstream row, wave placement (W3: CI-1/2 · W4: CI-3/5 · W5: CI-4/6/7), checklist incl. 7 language- pack cards, new shared-seam rows (session-setup tool activation, EP G8 shim). https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
…research New code-intelligence-alternatives.md — cited research record (June 2026): market validation (Claude Code native LSP v2.0.74, Copilot CLI LSP, OpenCode; Codex/Gemini still without), LSP engine bake-off (vscode-jsonrpc/-protocol vs Serena/solidlsp vs multilspy vs MCP bridges), per-language server matrix with operational caveats from issue trackers, structural-engine bake-off (ast-grep vs Semgrep/Opengrep, comby, GritQL, srgn, raw tree-sitter), repo-map evidence (Aider algorithm, RepoGraph ICLR'25, stack-graphs archived, embeddings contested), diagnostics formats (per-checker JSON matrix, SARIF-as-input, rdjson, errorformat), and a memory-layer check that validates the EP G2 Hindsight bet with a file-first framing. Applied to the CI docs: - CI-3 engine named (vscode-jsonrpc + vscode-languageserver-protocol; coc.nvim reference; solidlsp cache + multilspy auto-download as ported designs) and invariants expanded to 9 with pinning tests: handshake order, processId + reap, position-encoding negotiation, didOpen/didChange tracker, watched-file forwarding, readiness gating per server, post-edit diagnostics push, gitignore-filtered references, per-worktree CARGO_TARGET_DIR. - Server picks corrected: vtsls (not typescript-language-server), basedpyright (not pyright), csharp-ls before Roslyn LS, kotlin-lsp deferred (alpha). - CI-1 ships @ast-grep/cli as npm dependency (--json=stream) instead of the custom binaries/ route. - CI-2 schema = flattened rdjsonl; per-tool parser matrix; SARIF adapter; pytest via junitxml; bundled errorformat fallback binary. - CI-5 upgraded from flat symbol map to ranked def/ref-graph map with personalized PageRank seeded by session state (Aider mechanism; RepoGraph evidence), tags.scm-vs-napi spike called out. https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
Captures the owner Q&A in the research annex so the decision can't be re-litigated without the evidence: stack comparison diagram, the four-part argument (Serena's valuable layer is the lifecycle we'd rewrite anyway — per-worktree pooling is absent from its architecture; the protocol layer is the cheap part via vscode-jsonrpc; the MCP surface is the observed underuse failure mode a wrapper can't fix; Serena stays as ported designs + the ~one-week escape-hatch pack behind code.symbol-nav), and the one-sentence verdict. Design doc's "why not Serena" bullet now links to it. https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
…ified facts agent-memory.md (new): three-layer model (local transcripts = system of record, stay; markdown staff memory = authoritative/curated; Hindsight = optional episodic recall) with the distill-then-prune Caretaker retention invariant; full Hindsight-vs-alternatives comparison with reasons (mem0, Letta, Zep/Graphiti, cognee, LangMem, file-only); bank-topology decision — one shared tag-scoped bank with pack-side auto-tagging (project:/agent:/goal:/kind:) and scope: project|global|all on recall. Grounded in verified Hindsight docs (2026-06-11): cross-bank search is NOT supported (all ops target one bank), tags ARE the filtering mechanism with strict modes, shared-bank-with-tags is an endorsed pattern. This corrects EP G2's earlier per-project-banks sketch and the bank: current|global|all tool arg, which could not be implemented as written: - extension-platform.md: manifest description + provider flow/tools updated (scope→tag filters, auto-tagging, link to the decision record). - extension-platform-implementation-plan.md G2: two upstream verification checkpoints added (tag-filtered strict recall; delete-by-tag for project offboarding, with per-project-bank fallback if absent). - execution plan: Caretaker retention-consolidation card (after EP G2+MC P3). - CI research annex §7 now points to the full record. https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
… in §2 graph - §2 dependency graph: add the missing CI lane nodes (ast tools/diagnostics → LSP supervisor → packs/map/chip → capability swap, dotted edge to EP G8) and a prose translation naming every goal: the six day-one-parallel lanes, the sidebar-before-performance coupling, the Mission Control chain, the autoimprovement ladder, the CI lane, and the two soft couplings. - §2.1 waves: prose walkthrough of waves 0-5 in goal names rather than IDs — what each wave does, what runs in parallel inside it, and why each cross-wave gate exists (BENCH, SN-T5 rebase, EP G8 shim swap). https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
§2 said the global-session cwd is "~/.bobbit/mission-control/" while A.1 (authoritative storage layout) defines missionControlDir() = path.join(stateDir, "mission-control"), and bobbit-dir.ts resolves the state dir to BOBBIT_DIR env → else <gateway root>/.bobbit — home only when the env points there. §2 now states the A.1 path and resolution rule. https://claude.ai/code/session_01LBW68Z2KvA3EhQMq3c2wrR
…me-and-token-cost-efficiency Renames token-cost-efficiency.md -> time-and-token-cost-efficiency.md (updates all 6 cross-refs) and adds the wall-clock/time dimension the audit lacked: - New §9: latency axis + the inherent process-per-agent root cause. Measured (2026-06-19, 326 sessions): CC runs LARGER contexts per request (377k median) yet turns in ~4.4s; Bobbit's slowness is structural, not per-request. 1.34 verifiers per work session, median 2.8min, serialized on the critical path. - Root cause: every agent/verifier is a separate OS process (rpc-bridge spawn + MCP stdio + optional docker exec) that cold-writes its resident stack and can't inherit the lead's warm cache — vs claude-code's in-process forks + omitClaudeMd. - F5 refinement: verifier cache economics are INVERTED — write-bound (34% of cost is cache writes) not read-bound; re-scopes CE-G3.3. - New findings F13-F17 and CE-G8 latency implementation workstream (risk-proportional gates, slim sub-agent context, warm-cache reuse, latency instrumentation, solo-fast). - CE-G2.1: Hermes exact budget constants + the read_file=inf persist-loop gotcha. Co-authored-by: bobbit-ai <bobbit@bobbit.ai>
Content for the rename in the previous commit (only the bare mv landed there): - New §9: latency axis + inherent process-per-agent root cause. Measured 2026-06-19 (326 sessions): CC runs LARGER contexts per request (377k median) yet turns in ~4.4s; Bobbit slowness is structural. 1.34 verifiers per work session, median 2.8min, serialized on the critical path before team_complete. - Root cause: every agent/verifier is a separate OS process (rpc-bridge spawn + MCP stdio + optional docker exec) that cold-writes its resident stack and cannot inherit the lead's warm cache — vs claude-code in-process forks + omitClaudeMd. - F5 refinement: verifier cache economics INVERTED — write-bound (34% of cost is cache writes) not read-bound; re-scopes CE-G3.3. - F13-F17 + CE-G8 latency workstream (risk-proportional gates, slim sub-agent context, warm-cache reuse, latency instrumentation, solo-fast mode). - CE-G2.1: Hermes exact budget constants + read_file=inf persist-loop gotcha. - Update 6 cross-referencing docs + fable execution-plan CE row. Co-authored-by: bobbit-ai <bobbit@bobbit.ai>
…-clock lever) New exploratory design doc agent-swarm-and-reconciliation.md, the only lever that attacks serialization itself (vs §9's levers that make serial steps cheaper): - Two shapes: fan-out swarm (Claude Code in-process swarm: spawnInProcess + forkContextMessages + omitClaudeMd) and council/reconciliation (karpathy/llm-council, hermes mixture_of_agents) — the latter folds part of the verification tax into the parallel wave as peer cross-review. - Composes on existing platform machinery (capability registry, Lifecycle Hub, model-selector pack EP G9.2, workflow templates) — pack+capability, not core surgery. - Hard dependency: cheap spawning. A swarm of cold OS processes amplifies the process-per-agent tax (§9 F14) by swarm width. Unlock = single-container Bobbit (per-agent sandboxing redundant → in-process/worker_thread members inheriting warm cache, CE-G8.4); later, federation across connected Bobbits via the gateway. - BENCH-gated phased plan SW-G0..G4 (gated on the sandbox change), reusing CE-G0.3 + CE-G8.5 wall-clock metrics. Wires in: time-and-token-cost-efficiency §9.6 (serial→parallel lever + pointer), fable execution-plan SW row. Co-authored-by: bobbit-ai <bobbit@bobbit.ai>
…izing levers Measured 2026-06-20: 51% of gate signals are retries, 70% reviewer-driven, ~40h repeated impl verification. Adds CE-G8.7 (severity floor+round budget), CE-G8.8 (affected-only re-verify), CE-G8.9 (composed/right-sized workflows), CE-G8.10 (plan->cheap-exec->frontier-review + per-role thinking), CE-G8.11 (team-lead diet). Leads with trust-first north star (no lever may raise escaped-defect rate); cross-harness lessons from Claude Code + hermes-agent. Co-authored-by: bobbit-ai <bobbit@bobbit.ai>
1 task
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.