[3.0.0] - 2026年03月13日 — "The Session Never Dies"

Added

• Phoenix Rebirth — Recurring context recycling. When context hits the urgent threshold with Phoenix enabled, state is captured and recycling triggers automatically. After compaction, the cycle rearms — zero clicks, indefinite continuation. Safety-capped at phoenix_max_cycles (default 10)

  • phoenix_mode config (default false) — opt-in toggle, also available via /phoenix
  • phoenix_max_cycles config (default 10) — safety cap per chain
  • Phoenix chain tracking in /tmp/remembrall-phoenix/ — chain ID, cycle count, lineage
  • Extended avadakedavra-capture.sh with --trigger, --cycle, --chain-id args
  • Chain ID restoration on session resume — cycles persist across compactions
  • /phoenix command — toggle on/off, view chain status

• The Pensieve — Compaction-proof session intelligence. Tracks every file read/edit, command, and error throughout a session into structured JSONL. On compaction or handoff, distills into a summary that gets injected into the next session. Claude retains structured knowledge of what it did, even across compactions and session resets. On by default (pensieve: true)

  • hooks/pensieve-track.sh — Background incremental transcript parser (runs on every prompt)
  • hooks/pensieve-distill.sh — Crunches raw JSONL into structured summary JSON
  • hooks/pensieve-inject.sh — Generates compact text for additionalContext (budget-capped)
  • Persisted to ~/.remembrall/pensieve/{project-hash}/
  • Config: pensieve, pensieve_max_sessions (default 3), pensieve_inject_budget (default 2000)

• Time-Turner — Parallel agent at low context. At a configurable threshold (default 30%), spawns a headless claude -p agent in a git worktree with remaining tasks. The agent works independently while the main session compacts. On next resume, offers to merge changes. Opt-in only (time_turner: false by default)

  • hooks/time-turner-spawn.sh — Creates worktree, builds prompt, spawns claude -p
  • hooks/time-turner-check.sh — Checks status, formats report for injection
  • /timeturner skill — status, diff, merge, cancel sub-commands
  • Safety: opt-in, budget-capped (--max-budget-usd), worktree-isolated, never auto-merges
  • Auto-cleanup: stale worktrees >24h removed on SessionStart
  • Config: time_turner, time_turner_model (default sonnet), time_turner_max_budget_usd (default 1.00), threshold_timeturner (default 30)

• The Marauder's Map — /map command for visual session overview. Shows context gauge, files explored with R/E tags, commands with colored exit codes, error counts, burn rate, and Time-Turner status. Built from Pensieve data + bridge + growth tracking

  • scripts/remembrall-map.sh — Terminal visualization
  • commands/map.md — Command definition

• /pensieve skill — Browse and search Pensieve session memories. List sessions, view summaries, search across session history

• Session Lineage (Marauder's Map — Session Ancestry) — Full session DAG tracking. Every session (main, compacted, Time-Turner) is recorded in a lineage index with parent/child relationships. Renders as a text DAG showing session ancestry, branches, and merge status

  • hooks/lineage-record.sh — Records session in lineage index (called by precompact + handoff-create)
  • scripts/remembrall-lineage.sh — Renders text DAG with depth, branch counts, and HP theming
  • commands/lineage.md — /lineage command
  • Library: remembrall_lineage_dir(), remembrall_lineage_record(), remembrall_lineage_depth(), remembrall_lineage_branches()
  • Storage: ~/.remembrall/lineage/{project-hash}/index.json
  • Config: lineage (default true), lineage_max_entries (default 50)
  • HP theme: branches = "Horcrux detected"

• Ambient Learning / Insights (The Pensieve Remembers) — Aggregates Pensieve session data into actionable patterns. Tracks file hotspots, workflow patterns (test-before-commit), error recurrence, and session statistics. Background aggregation on SessionStart

  • hooks/insights-aggregate.sh — Aggregates Pensieve sessions into patterns (background)
  • scripts/remembrall-insights.sh — Renders formatted insights with HP theming
  • commands/insights.md — /insights command
  • Library: remembrall_insights_dir(), remembrall_insights_fresh()
  • Storage: ~/.remembrall/insights/{project-hash}/insights.json
  • Config: insights (default true), insights_inject (default false), insights_min_sessions (default 3)

• Semantic Context Pruning / Obliviate — Memory staleness analyzer that cross-references memory files with Pensieve data. Identifies stale memories not accessed in recent sessions and offers guided pruning with user confirmation. Archives stale memories instead of deleting

  • hooks/obliviate-analyze.sh — Memory staleness analyzer (background, at journal threshold)
  • scripts/obliviate-archive.sh — Moves stale memories to .archive/
  • commands/obliviate.md — /obliviate command
  • skills/obliviate/SKILL.md — Guided pruning skill with user confirmation
  • Library: remembrall_memory_dirs(), remembrall_obliviate_dir(), remembrall_analyze_memory_staleness()
  • Tmp: /tmp/remembrall-obliviate/{session_id}.json
  • Config: obliviate (default true), obliviate_stale_sessions (default 5)
  • HP theme: "Obliviate! N stale memories banished to the archive."

• Context Budget Allocation (The Sorting Hat) — Transcript category breakdown that classifies context consumption into code, conversation, and memory. Compares actuals vs configured budgets and warns when categories are imbalanced. Opt-in

  • hooks/budget-analyze.sh — Transcript category breakdown (background)
  • commands/budget.md — /budget command
  • Library: remembrall_budget_dir(), remembrall_extract_category_bytes(), remembrall_budget_check(), remembrall_budget_validate_total()
  • Tmp: /tmp/remembrall-budget/{session_id}.json
  • Config: budget_enabled (default false), budget_code (default 50), budget_conversation (default 30), budget_memory (default 20)
  • HP theme: "The Sorting Hat detects an imbalance!"

• Patrol Integration (Owl Post) — File-based signal protocol for Patrol plugin interop. Patrol is fully optional — zero behavior change when not installed. Signal types: handoff_trigger, context_alert. Patrol writes signal files, Remembrall reads and consumes them

  • docs/patrol-integration.md — Signal protocol spec (the contract)
  • Library: remembrall_signal_dir(), remembrall_check_patrol_signal(), remembrall_consume_signal(), remembrall_patrol_detected()
  • Tmp: /tmp/remembrall-signals/{session_id}/
  • Config: patrol_integration (default true), patrol_signal_ttl (default 300)
  • HP theme: "Owl Post" / "Ministry of Magic: Patrol (connected)"

• Config validation for all new settings (12 new config keys total)

• remembrall_pensieve_dir() and remembrall_pensieve_tmp() library functions

• Pensieve, Lineage, Insights, Obliviate, Budget, and Patrol data in remembrall-status.sh diagnostic output

• Time-Turner status in remembrall-status.sh diagnostic output

• All new feature cleanup in remembrall-uninstall.sh

• Session Intelligence (Pensieve) section appended to handoff files

• 351 total tests (135+ new tests covering all 5 features)

Changed

• Handoff files now include ## Session Intelligence (Pensieve) section with distilled session data
• Session-resume injects Pensieve memory on ALL session starts (fresh + compact + clear)
• Session-resume spawns insights aggregation in background
• Context-monitor checks for Patrol signals before threshold comparison
• Context-monitor spawns obliviate analyzer at journal threshold
• Context-monitor spawns budget analyzer below journal threshold
• Context thresholds diagram updated: shows Time-Turner trigger at 30%
• Help tables updated with all new commands, skills, and config options
• Marauder's Map (/map) shows budget section when budget is enabled

Full Changelog: v2.7.1...v3.0.0

Fixes

• stop-check.sh: Fixed invalid JSON schema — hookEventName: "Stop" (not in schema) → correct "decision": "block" pattern. Stop hook errors are gone.
• context-monitor.sh: Added dual-channel plan mode enforcement via systemMessage + additionalContext with BLOCKING REQUIREMENT language at warning (35%) and urgent (15%) thresholds. Plan mode now triggers more
  reliably when context is low.

Full Changelog: v2.7.0...v2.7.1

Fixed

• Claude ignores context nudges — wrong JSON format (additionalContext vs hookSpecificOutput) caused Claude to silently discard all nudge messages. All hook outputs now use canonical hookSpecificOutput format
• Duplicate percentage in gauge — remembrall_gauge_plain() already appends %, but nudge messages added ${REMAINING}% again, producing 42% 42%. Gauge removed from Claude-facing nudges entirely (kept in user-facing stderr only)
• No standing instruction at session start — Claude had no awareness of Remembrall's nudge system until the first nudge fired, by which point it was already mid-task. Now emits compliance instruction on every SessionStart

Changed

• Thresholds shifted to 60/35/15 (was 60/30/20) — gives one more turn at 35% warning, makes blocking a true last resort at 15%
• Two-stage urgent escalation — first prompt at ≤15% gets urgent nudge, second consecutive prompt gets hard-blocked with /clear + /replay instructions
• Terse nudge messages — stripped gauge bars, easter eggs (journal only), and verbose instructions. Messages now use REMEMBRALL_WARN: / REMEMBRALL_URGENT: prefixes for reliable Claude detection
• Block guardrails — never block on estimated values (only bridge-confirmed), never block autonomous sessions, only block after two-stage escalation

Added

• Standing instruction emitted on every SessionStart (fresh and resume) via hookSpecificOutput
• Version guard at SessionStart cleans stale temp files when plugin version changes
• session_id included in debug log for easier troubleshooting

Integrate Remembrall gauge installation into the mandatory setup flow and reorder documentation sections.

Full Changelog: v2.6.2...v2.6.3

Bug Fix: Stop hook crash

The Stop hook (stop-check.sh) crashed on every Stop event with:
Stop hook error: Failed with non-blocking status code: No stderr output

Root cause: set -euo pipefail combined with remembrall_find_bridge returning exit code 1 when the bridge file doesn't exist (normal at session start or after /tmp cleanup). The 2>/dev/null was on the wrong side of the $() — it redirected the assignment's stderr instead of the subshell's, so the error was silent.

Impact: Remembrall's stop-time context warning was completely dead — users never got the "context is low, save a handoff" nudge when Claude stopped responding.

Fix: Guarded all three affected hooks:

• stop-check.sh — the active crash (Stop hook)
• context-monitor.sh — same pattern, cosmetic fix (no set -e)
• precompact-handoff.sh — remembrall_calibrate unguarded, could crash on lock/jq failure

Full Changelog: v2.6.1...v2.6.2

v2.6.1 — Persistent Nudges

Bug fix: Claude was ignoring context warnings because nudges fired only once per threshold. If Claude was mid-task and missed the single reminder, context drained silently to 1%.

Fixes

• Persistent nudges at warning (30%) + urgent (20%) — nudges now repeat on every prompt until a handoff file actually exists on disk, instead of firing once and going silent
• Nudge text upgraded to "BLOCKING REQUIREMENT — you MUST invoke the /handoff skill NOW, before responding to the user's request" for stronger compliance
• Preemptive handoff creation only runs on first nudge per threshold (prevents background process spam)
• Plan mode now shows a numbered summary (1. clear context and resume, 2-3. next steps) instead of rambling or going silent after handoff

Unchanged

• Journal threshold (60%) remains fire-once — gentle reminder, no enforcement needed

Install / Update

claude plugin marketplace update cukas

Full Changelog: v2.6.0...v2.6.1

v2.6.0

Features

• Track active skill/workflow in handoff and replay — resuming sessions now know which skill and step to pick up from

Fixes

• Resolve plugin root for skills/commands via /tmp persistence
• Explicit handoff-before-planmode ordering in nudge text
• Use POSIX [0-9] instead of \d in test regex for GNU grep compatibility

Docs

• Add CI badge to README
• Add animated demo GIF below gauge examples in README
• Add Install heading above install commands
• Update README config example and FAQ for v2.5.0

Full Changelog: v2.5.0...v2.6.0

Remembrall v2.5.0

Configurable Thresholds

Context nudge thresholds are no longer hardcoded. Tune when Remembrall speaks up:

{
"threshold_journal": 50,
"threshold_warning": 25,
"threshold_urgent": 15
}

Defaults remain 60/30/20 — adjust to match your workflow.

Debug Logging

New debug system for troubleshooting hook behavior:

remembrall_config_set "debug" "true"

or one-off: REMEMBRALL_DEBUG=1 claude

Logs to ~/.remembrall/debug.log with ISO timestamps, per-hook identification ([context-monitor], [session-resume], etc.), and auto-rotation at 1MB.

Clean Uninstall

New scripts/remembrall-uninstall.sh with --dry-run support. Preview what gets removed before committing:

/remembrall-uninstall # shows dry-run first, then full removal

Removes bridge from settings.json, cleans ~/.remembrall/, and purges temp files.

Other Improvements

• format_version: 2 in handoff frontmatter for forward compatibility
• Configurable recency_window (default: 60s) for handoff-to-session matching
• Background handoff creation — preemptive handoffs no longer risk the 15s hook timeout
• Defensive status line check — won't clobber another plugin's status line
• Fixed jq injection in remembrall_hook_enabled() (now uses --arg)
• DRY refactor — remembrall_default_content_max() extracted, no more duplicate model defaults
• 198 tests, all passing

Full Changelog: v2.3.1...v2.5.0

Full Changelog: v2.3.0...v2.3.1

What's New

Auto-Calibrating Context Estimation

Remembrall now derives your real context limit from bridge data — no compaction events needed. The formula content_max = content_bytes / (used_pct / 100) runs continuously while the bridge is active, auto-calibrating per user and per model. Your specific plugin overhead, skill count, and system prompt size are all accounted for automatically. Calibration
samples are stored once context usage reaches 20%.

Bridge Auto-Creation

Zero setup required. On first session start, session-resume.sh auto-injects the bridge snippet into ~/.claude/settings.json. If you already have a custom status line, the bridge is injected into it. If you don't have one, Remembrall creates a minimal one. No more manual /setup-remembrall needed.

Two-Branch Estimation Strategy

The estimation system now uses a clear two-branch approach:

• Bridge active → use real context % directly (most accurate), calibrate content_max as a side effect
• Bridge unavailable → structural JSONL parser first, raw file-size fallback second — both using calibrated limits when available

Removed Staleness Timeout

The bridge file no longer has a staleness timeout. Context % is read whenever the bridge file exists and invalidated on compact/clear events. This eliminates false fallbacks to estimation when the bridge was simply not refreshed between prompts.

20 Audit Fixes

A comprehensive code audit resolved bugs across all hooks, skills, and config:

• ANSI in JSON — Nudge messages now use plain-text gauges for valid JSON output
• Dead PreCompact hook — Safety net now fires on actual system compaction events
• Calibration race condition — File locking prevents data corruption with concurrent sessions
• Cross-session handoff theft — Recency fallback now verifies frontmatter session ID
• Nudge oscillation — Warning can no longer re-fire after urgent has already triggered
• Stop hook async — Now runs synchronously to block task completion as intended
• Filename safety — Replaced xargs trim with POSIX-safe whitespace handling
• Plus 13 additional fixes across skills, commands, and hook registration

Full Changelog: v2.2.0...v2.3.0