v0.8.0 — PreCompact Hooks + Prompt Audit

Added

• PreCompact hooks — Session learner checkpoints during compaction (async). Session transcript captures pre-compaction dialogue and re-injects it after compaction, preserving conversation context across compact boundaries.
• Configurable stop checklist — Add custom checklist items via stop_checklist_extra in .meridian/config.yaml.
• Configurable instruction reminders — Add custom per-message reminders via instruction_reminders in .meridian/config.yaml.

Changed

• Stop checklist reframed — No more "stopping" language. The checklist now reads as a task list, preventing agents from interpreting it as a signal to wind down and defer work.
• Context injection streamlined — Removed ceremonial header and footer that wasted a turn asking the agent to "acknowledge" the injected context.
• Action counter — Accumulates across stop attempts instead of resetting on skip.

Install / Upgrade

curl -fsSL https://raw.githubusercontent.com/markmdev/meridian/main/install.sh | bash

v0.7.9 — Instruction Reminder Hook

Added

• Instruction reminder hook — injects a short reminder to follow CLAUDE.md instructions on every user message. Prevents instruction drift during long sessions.

Upgrade

curl -fsSL https://raw.githubusercontent.com/markmdev/meridian/main/install.sh | bash

v0.7.8 — Full Transcript Preservation

Changed

• Session transcript — Removed per-entry and total output truncation. Full dialogue content is now preserved in session transcripts.

Upgrade

curl -fsSL https://raw.githubusercontent.com/markmdev/meridian/main/install.sh | bash

v0.7.7 — Session transcript logging, slimmed manual

Changes

Session transcript observability

• session-transcript.py now logs at every decision point — start, skip, extract count, write, and crash — matching session-learner's logging pattern. Silent failures are no longer invisible.

Agent operating manual slimmed

• Manual reduced from 118 to 25 lines. Now covers only Meridian-specific concerns: workspace, doc maintenance, and external tools. Planning, verification, error handling, and code review sections removed — these are covered by CLAUDE.md and skills.

Pebble rules extracted

• Pebble tracking rules moved to pebble-rules.md, injected conditionally only when pebble_enabled is true. Zero context cost for projects that don't use Pebble.

Active plan injection removed

• Plan files are no longer force-injected at session start. The planning skill handles plan state directly.

Dead code cleanup

• Removed get_active_plan_path(), empty file injection loop, unused metadata keys, and fixed broken cross-references in SOUL.md.

Upgrade

curl -fsSL https://raw.githubusercontent.com/markmdev/meridian/main/install.sh | bash

v0.7.6 — Planning Skill Triggering

Changed

• Planning skill description — Optimized for triggering accuracy. Explicit "INSTEAD of EnterPlanMode" directive + methodology details (interviewing, parallel exploration, plan validation) improved eval recall from 0% to 95%.
• Audit skill references — Rewritten to use structural anti-patterns instead of naive grep patterns for more consistent detection.

Upgrade

curl -fsSL https://raw.githubusercontent.com/markmdev/meridian/main/install.sh | bash

v0.7.5

Fixed

• Missing <local-command-stdout> and <task-notification> noise markers in SYSTEM_NOISE_MARKERS — session learner and session transcript were including these as real dialogue.

v0.7.4

Added

• claude_runner shared module — Reusable Python module for claude -p subprocess calls. Ships in both scripts/lib/ (plugin) and .meridian/lib/ (scaffolding).

Changed

• Session learner migrated to use claude_runner

Removed

• build_headless_env() / build_headless_args() from meridian_config.py

v0.7.3 — Gitignore install artifacts

Added

• .meridian/.gitignore — Install artifacts (.version, .manifest, workspace/) are now gitignored by default. No more accidental commits of scaffolding state.

Upgrade

curl -fsSL https://raw.githubusercontent.com/markmdev/meridian/main/install.sh | bash

v0.7.2 — Remove terminal injection summary

Removed

• Terminal injection summary — The systemMessage that showed injection details in the terminal at session start was noisy and showing incorrect info. Removed. Context injection continues to work normally via additionalContext.

Upgrade

# Scaffolding
curl -fsSL https://raw.githubusercontent.com/markmdev/meridian/main/install.sh | bash

# Plugin (inside Claude Code)
/plugin update meridian@markmdev

v0.7.1 — Terminal Injection Summary

Added

• Terminal injection summary — Context injector now shows a one-liner in your terminal at session start via systemMessage. Shows what was injected (workspace, docs count, last session, plan, pebble, manual, soul, nested repos) and error count. Not visible to the agent — just for you.

Example: Meridian: workspace · 42 docs · 1 api-doc · plan · manual · soul · 5 nested repos | No errors

Upgrade

curl -fsSL https://raw.githubusercontent.com/markmdev/meridian/main/install.sh | bash

Then restart Claude Code to pick up the new hook.