v1.0.100 — Unified Persistent Memory & Timeline Search

Context-mode now remembers across compactions and sessions. Skills, roles, decisions, and session history survive context resets — no more re-explaining what you were working on.

Unified Timeline Search

ctx_search gains a new sort="timeline" mode that searches across 3 sources chronologically:

Current session (ContentStore) — tool outputs, indexed content
Prior sessions (SessionDB) — events from the last 7 days
Auto-memory (memory/*.md files) — persistent user preferences

ctx_search(queries: ["what were we working on"], sort: "timeline")

Results are tagged by origin and timestamp:

--- [prior-session | 2026-04-27 14:30 | decision] ---
### [decision] user_correction
Always push to next branch, never feature branches

--- [auto-memory | 2026-04-25 | memory/project_analytics.md] ---
Analytics must be separate project, Datadog model

Default sort="relevance" is unchanged — zero behavior change for existing users.

Auto-Injection on Compaction

When context compacts, context-mode now automatically injects critical state as behavioral directives — not informational summaries the LLM ignores:

<behavioral_directive> — active role ("You are acting as a senior engineer")
<rules> — key decisions ("DO NOT use mocks in integration tests")
<active_skills> — loaded skills ("Re-invoke if relevant: tdd, cloudflare")

500 token hard cap. Fires only on compaction. No mid-session injection.

12 New Event Categories

Phase 1 (session continuity)

Category	What it captures
`user-prompt`	Every user message, ordered — "what was my first prompt?"
`compaction`	Pre-compaction summary + session state
`rejected-approach`	Denied/modified tool calls — "why didn't we do X?"
`session-resume`	Proof that prior session was resumed
`constraint`	Discovered limitations — "FTS5 doesn't support ALTER"

Phase 2 (correlation & analytics)

Category	What it captures
`agent-finding`	Subagent research result summaries
`error-resolution`	Error → diagnosis → fix chains
`external-ref`	Referenced URLs, issues, PRs
`blocked-on`	Pending blockers and their resolution
`iteration-loop`	Stuck loop detection (same tool repeated)
`latency`	Tool call duration (cross-hook timing)

Bug Fixes

Role events silently dropped — snapshot.ts was missing case "role" in the switch statement. Roles now survive compaction.
Skill priority too low — promoted from P3 to P2 so skills aren't evicted during snapshot trimming.
Pi extension filter too aggressive — minPriority: 2 dropped all P3 events (skills, roles, git, errors). Changed to minPriority: 3.

Static Reinforcement

12 adapter configs updated with Memory section — 4 tiers based on hook support
Routing block gains <session_continuity> tag + 0. MEMORY in tool hierarchy
ctx_search tool description includes SESSION STATE reminder

Adapter Support

Tier	Adapters	Memory Features
Full	Claude Code, Qwen Code, OpenClaw	All categories, full timeline search
Partial	Codex, Gemini CLI, VS Code, JetBrains	No user-prompt history
Basic	Cursor, OpenCode, Kilo, Kiro, Pi	Decision + constraint search only

Numbers

39 files changed, +3717/-552 lines
1902 tests pass, 0 failures
8 review rounds, 16+ engineers per round
TDD-first implementation

Full Changelog: v1.0.99...v1.0.100

mksglu/context-mode v1.0.100 on GitHub