Skip to content

feat(hermes): self-improvement suite — behavioral profile, memory, guardrails (v1.62.7 → v1.62.17)#272

Open
celstnblacc wants to merge 1 commit into
mainfrom
feat/hermes-self-improvement-v1.62.17
Open

feat(hermes): self-improvement suite — behavioral profile, memory, guardrails (v1.62.7 → v1.62.17)#272
celstnblacc wants to merge 1 commit into
mainfrom
feat/hermes-self-improvement-v1.62.17

Conversation

@celstnblacc

Copy link
Copy Markdown
Owner

Summary

19 releases today adapting Hermes agent self-improvement patterns to superharness.

Hermes Adaptations (Iterations 1-7)

  • Agent-writable memory (two-tier: global + per-project) —

  • Tool-loop guardrails — wired into watcher log analyzer, auto-blocks tasks

  • Auto-promotion — project → global memory after 3 occurrences

  • Behavioral profile engine — zero-touch user adaptation from SQLite patterns

  • Confidence scoring + EWMA decay + hysteresis

  • Verification feedback loop — A/B test every profile change

  • Dashboard profile card — endpoint + UI

  • Memory file caps — 5,000 chars FIFO pruning (Hermes adaptation)

  • Onboarding bootstrap — seed behavioral profile at
    superharness -- non-interactive mode
    Configure using flags:
    shux onboard --non-interactive --git-mode team --autonomy supervised
    Or edit .superharness/profile.yaml directly.
    Run 'shux onboard' in an interactive terminal for the full wizard.

    Available sections:
    shux onboard --section project Project identity
    shux onboard --section agent Agent settings
    shux onboard --section git Git & tracking
    shux onboard --section hooks Hooks
    shux onboard --section watcher Watcher daemon
    shux onboard --section gateway Notifications
    shux onboard --section task First task

[detect] Project stack: Python (git repo) — found at /Users/airm2max/DevOpsSec/superharness
→ superharness will use this to tailor agent instructions and task defaults.
[skip] Step 2 (init): .superharness/ already exists
[skip] Step 2b (global_claude): superharness already in global CLAUDE.md
[git_track] Team mode: .superharness/ will be committed
→ team mode: task state is committed — your whole team shares it.
[git_track] Created/updated .superharness/.gitignore
→ Runtime files (logs, daemon pid, watcher env) excluded from commits.
[doctor] Warning: some checks failed (non-blocking):
superharness doctor
project: /Users/airm2max/DevOpsSec/superharness
PASS dep:python3
PASS dep:claude
PASS dep:codex
PASS project:.superharness present
FAIL file:contract.yaml missing
Re-initialize: superharness init
FAIL file:ledger.md missing
Re-initialize: superharness init
FAIL file:decisions.yaml missing
Re-initialize: superharness init
FAIL file:failures.yaml missing
Re-initialize: superharness init
FAIL dir:handoffs missing
Run: mkdir -p .superharness/handoffs
PASS git:core.hooksPath=/Users/airm2max/DevOpsSec/superharness/.git/hooks
WARN plugin:claude-code superharness not installed
Run: bash adapters/claude-code/install.sh (from superharness repo)
WARN watcher:com.superharness.inbox.superharness not loaded
The background watcher is required — install it with: shux watcher-worker -p .
Or use foreground mode instead: superharness watch --foreground --project .
INFO mcp:no memory server detected (optional — see docs/MCP-MEMORY.md)
INFO modules: none enabled — run 'shux enhance' to add integrations
PASS state-db: state.sqlite3 present and initialised
PASS parity: SQLite is the sole source of truth (no YAML to drift from)
summary: failures=5 warnings=2

→ Fix the failures above, then re-run 'shux doctor'.
→ These won't stop you — proceed and run 'shux doctor' to fix later.
[task] Skipped (no --task-title provided)
→ Add your first task later: shux task create --title "..."
[delegate] Skipped (no --enqueue or no task)
→ When ready: shux delegate to hand work to an agent.

superharness is set up for this project.

Setup status:
✓ detect
✓ init
✓ global_claude
✓ git_track
✓ doctor
✓ task
✓ delegate

Next steps:
shux contract — view all tasks
shux delegate — hand a task to an agent
shux doctor — re-run health checks
shux dashboard — open browser dashboard

Behavioral Profile (optional)
The system learns your patterns automatically. Seed it with 2 quick answers:

  1. Review style? (strict/balanced/lenient) [balanced]:

Code Review Fixes (C1-C6)

  • 392 → 0 bare across 218 files
  • Duplicate and state machine key fixed
  • cli.py 1,022→648 lines (dashboard extracted)
  • delegate.py context hint →

Bugfixes

  • inbox.py enqueue crash for projects (ImportError + undefined logger)

Docs

  • 19 obsolete docs archived, docs/README.md index
  • Health check rules in AGENTS.md, CLAUDE.md, GEMINI.md

Tests

  • 72+ new behavioral tests (I1-I7)
  • 103 CLI tests passing (3 broken files fixed)
  • 7 new dashboard data tests (C6)

New CLI Commands

  • Usage: shux profile <show|edit|reset|lock|unlock> [key]
  • Usage: shux memory roots <list|add|remove> [path]

Append § Validation against 1.62.17 to the 2026-05-21 investigation. The
enqueue crash (engine/inbox.py import + logger), Bug G runaway re-dispatch,
and the deadline-check JSONDecodeError are all fixed — verified end-to-end
(daemon ran 95s clean, 0 err.log lines, stale tombstones GC'd). Document one
residual non-fatal bug: redundant _enqueue_sqlite_shadow double-write
(discuss.py:236) trips the new duplicate guard and logs a traceback on
interactive `shux discuss start`; daemon dispatch path unaffected.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants