100x Workflow Overhaul

⚡ 100x Workflow Overhaul

A single-truth, cross-platform rebuild of the agentic-mercy-10x workbench. The whole point of this release: nothing in the workflow is missed, nothing is faked, and it now runs on Windows as well as Ubuntu — while every user-facing trigger, skill name, and command name that ever worked keeps working.

🎯 Highlights

• Unified prompt router + never-miss trigger floor. One router subprocess per prompt replaces ~15 legacy injectors. hooks/trigger-floor.json is a verbatim superset of every keyword/path/intent rule from all 4 legacy configs — additions allowed, removals impossible (CI floor-check enforces it forever). A shadow-mode parity harness proves zero missed triggers versus the legacy stack. Budget is priority-ordered (~24k), deduped via a session manifest.
• Single model-policy + crash fix. hooks/model-policy.json is the one source of truth for model routing — sonnet by default, opus for UI + genuinely heavy work, fable only on explicit request. The /invoke-impl suite pins implementation to opus. Fixed the subagent model-guard argument-drop crash and added regression tests.
• Active-repo index lifecycle — zero daemons. Code-intel (jcodemunch) + doc-intel (jdocmunch) + graph (graphify) freshness is now event-driven and active-repo-only: a journal triggers a detached, single-shot, time-boxed builder. No always-on systemd watchers (the CPU/RAM problem from the prior design is gone). Structural guarantees keep builds off non-active repos.
• 8 event dispatchers — orchestration, not fusion. settings.json shrinks to 8 dispatch.py <event> registrations, but every original hook stays its own file/module — declared as a link in dispatch.config.json. Guarantees: per-link try/except isolation (one crash never stops the chain), per-link fire telemetry, per-link enable flags, and a doctor that fires synthetic events through every link. Every hook fails open.
• Provenance-aware skill consolidation.
  • 128 upstream-locked skills (the gstack clone + its twins, higgsfield + mmx-cli, all gsd-*, the vendored design-craft skills, vite-react-best-practices, find-skills) are kept byte-identical and hash-verified (validator rule R10) so you can still git pull the author's next version cleanly. Router integration for these is sidecar-only.
  • 24 user-authored merges into canonicals, each preserving the merged-away skill's name as a thin routing alias — old references and keywords keep resolving, permanently.
  • Commands: 139 → 20. One parametric /invoke <acts…> plus delegators, aliases, and utilities. All 139 historic command names still resolve — as a file or via the invoke_compat router translator — and gen-invoke-commands.py --emit-combos re-expands the 120 combo files in seconds.
• Windows + Ubuntu installer + two-OS CI. install.py is a stdlib-only (Python ≥ 3.10), OS-auto-detected, idempotent bootstrap with install / update / doctor verbs. Skills materialize via copy or NTFS junction — never a symlink. A windows-latest + ubuntu-latest CI matrix runs the doctor, R9 floor-check, R10 upstream-intactness, portability grep-gates, and the full test suite on every push.
• Verification report. End-to-end Linux metrics captured against the §2 before/after table: 1 router subprocess/prompt, priority-ordered ≤24k injection, dedup, SessionStart wall time, palette counts (218 skill names / 20 command files), R10 pass, mid-session write→flush→fresh-index round trip, and model-routing spot-checks.

📦 Install

Ubuntu / macOS

git clone https://github.com/AjayIrkal23/agentic-mercy-10x ~/.claude && python3 ~/.claude/install.py

Windows (PowerShell)

git clone https://github.com/AjayIrkal23/agentic-mercy-10x $env:USERPROFILE\.claude ; py -3 $env:USERPROFILE\.claude\install.py

Then: python install.py doctor (health + trigger-surface + model-routing verifier) · python install.py update (pull → deps → re-render → rebuild → doctor). Flags: --dry-run, --ci.

⏳ Time-gated (honestly open, not faked)

These are wired, documented, and evidence-driven — they close on elapsed time / a real second-OS run, not on a checkbox:

• ≥10-session zero-miss shadow-parity window. The router runs in --shadow beside the legacy injectors; cutover to router-only (and retirement of the legacy stack) happens only after ≥10 real sessions with zero missed triggers. Legacy injectors stay one-flip revertible (scripts/flip-router.py --legacy) for 30 days.
• Windows real-VM validation (P7-T2). The windows-latest CI leg covers the automated portion (dispatch on all 8 events, junction materialization, zero symlinks, doctor). A full install.py run into %USERPROFILE%\.claude on a real Windows VM is the remaining manual sign-off.
• 30-day legacy-retention closure + parity sign-off (P7-T4/T5). The legacy prompt-stack and the parity scaffolding are atticked only after the window closes with an empty miss-watchdog log.

🔒 Safety & rollback

• ~/.claude is a git repo. Rollback anchor tag: pre-100x. Every retired artifact is atticked (not deleted) with a per-item justification.
• Secret-audit clean: no credential values in the tracked tree or full history; generated/runtime artifacts and machine-private project paths are gitignored.
• Private planning artifacts remain out of this public repo by design (.gitignore).

Full rollback: git checkout pre-100x.