Skip to content

ROADMAP.md

Latest commit

 

History

History
100 lines (71 loc) · 4.38 KB

ROADMAP.md

File metadata and controls

100 lines (71 loc) · 4.38 KB

AgentSesh Roadmap

Last updated: 2026-03-14

Vision

v0.x: Diagnostic — "Here's what happened." v1.0: Closed loop — "Here's what happened, and it's in your agent's context for next time." v2.0: Adaptive — "Your agent improved 15% over 20 sessions. Here's what's still trending wrong."

Shipped

v0.10.0 (2026-03-14)

  • sesh analyze --feedback — closed-loop improvement, writes to CLAUDE.md
  • sesh analyze auto-discovers most recent session (no args needed)
  • sesh audit --threshold — CI gate with exit codes
  • Audit metrics: mypy, pyright, pre-commit, tox, nox, justfile, Python docstrings
  • cli.py split into commands/ package

v0.9.0 (2026-03-14)

  • sesh audit — repo legibility scoring, 9 metrics, letter grades

v0.8.0 (2026-03-13)

  • sesh analyze — one-command diagnostic, no database required

v0.7.0 (2026-03-13)

  • sesh debug — prompt debugger (thinking search, action lookup, dotnotes, correlate)

v0.6.0 (2026-03-13)

  • sesh replay, sesh test, sesh fix — playback, regression testing, remediation

v0.1.0–0.5.0 (2026-03-13)

  • Core: parse, grade, detect patterns, store, trend, search, handoff
  • MCP server, web dashboard, watch mode
  • Claude Code + OpenAI Codex parsers

Next (v0.11)

Specific remediations

The --fix output is generic boilerplate ("use Read instead of cat"). It should be session-specific: "you ran wc -l sesh/cli.py at call #47 — that's a Read." Map each detected instance to the concrete alternative.

sesh analyze --previous

Skip the current in-progress session and analyze the most recently completed one. Agents analyzing themselves mid-session get incomplete data.

Cursor/Windsurf support

Parse .cursorrules for --feedback target. Detect Cursor session transcripts if they exist. Audit metric already checks .cursorrules — connect the dots.

v1.0 — The Closed Loop

The minimum viable version shipped in v0.10.0 (--feedback). v1.0 is the polished version:

Automatic analysis

sesh watch + sesh analyze combined: watch for completed sessions, auto-analyze, auto-write feedback. No human intervention. Daemon mode.

Longitudinal feedback

--feedback currently shows last session only. v1.0 shows trends: "bash_overuse in 8 of your last 10 sessions, always with wc and find." Specificity from accumulation.

Multi-target feedback

Write to CLAUDE.md, .cursorrules, custom files, webhooks. CI integration (PR comments with session analysis).

Exit codes everywhere

sesh analyze returns non-zero if grade below threshold. Same CI-gate pattern as sesh audit --threshold.

v2.0 — Adaptive

Improvement tracking

"Your bash_overuse rate dropped from 47% to 12% over 20 sessions." Prove that the feedback loop works with data.

Pattern evolution

Detect when an agent develops new antipatterns as old ones get fixed. "You eliminated bash_overuse but missed_parallelism increased 3x."

Custom pattern definitions

Let users define their own antipatterns. "Flag any session where the agent creates more than 3 new files" or "warn if test coverage drops."

Agent benchmarking

Compare agents against each other or against cohort averages. "Your agent's B average is above median for this codebase size."

Ideas (backlog)

From the 100 Ideas list

  • Voice drift detector — analyze writing/communication patterns across sessions
  • Context budget visualizer — track how agents spend their context window
  • Thread continuity scorer — rate context maintenance across conversations
  • Session cost estimator — predict token usage before a session runs
  • Diff as essay — generate narrative changelog from session tool calls
  • Session archaeology — deep-dive old sessions with full context reconstruction
  • Agent personality tests — structured prompts that reveal behavioral tendencies

Product ideas

  • GitHub Action — sesh audit --threshold 80 as a one-line CI step
  • VS Code extension — show session grade in status bar after completion
  • Leaderboard — opt-in anonymized benchmarking across the community
  • sesh init --guided — interactive setup that asks about your workflow
  • Plugin system — custom parsers, custom patterns, custom audit metrics

Technical debt

  • Split test_debug.py (926 lines)
  • Audit metric weights should be configurable per-project
  • MCP server needs tests
  • Web dashboard needs authentication