Sessionstart

Session Start — v2.4

The opening ritual that eliminates the cold-start problem. When you open a new session, the department head is already briefed AND the machine state is verified. Goal: zero re-orientation time and zero stale-memory action.

The counterpart to /sessionend. Where sessionend closes the loop, sessionstart opens it.

---

BRAIN STRUCTURE — Know Before You Read

Master Brain (Chief of Staff):
  {brain_root}/mission-control/
    DASHBOARD.md             ← company-wide status
    awareness.md             ← per-dept running record (primary source for cross-dept flags)
    decisions/               ← company-wide strategic decisions
    state/                   ← live infrastructure state (circuit-breakers.json, etc.)
    NAMING_CONVENTION.md     ← canonical naming rules
    BRAIN_GOVERNANCE.md      ← chain-of-command rules
    skills/                  ← source-of-truth skill copies

Department Brains (kebab-case, lowercase `brain/` at every dept):
  {operator_departments} — configured in brain.config.yml at install time
  # Pattern:
  #   {dept}/brain/sessions/           | decisions/decisions.md | pipeline/current-state.md

Path hygiene: brain/ (lowercase) is canonical at every dept. If a capital-B Brain/ appears anywhere, surface it in Stage 2.5 as a drift finding and normalize.

Department codes: configured in brain.config.yml — set at install time per operator's dept roster.

---

Stage 0 — Session continuity probe (runs first)

Purpose: Avoid re-running a full 5-stage brief when continuing a same-day session within the usual break window (~1h).

Algorithm

1. Read most recent {brain_root}/mission-control/sessions/YYYY-MM-DD-*.md (any dept) by file mtime.
2. If date == today AND status == closed AND (now - close_timestamp) < 1h:
   • Render the 3-line turn-start probe (template in references/stage0_turn_start.md):
     • Previous session: {session_id} (closed {hh:mm})
     • Open loops carried: {len(open_loops)} — see Q2 on demand
     • Delta since close: {top-3 commits from git log --since={close_timestamp}}
   • Exit sessionstart. Do NOT run Stages 1–5.
3. Else (> 1h since close, or no same-day closed session, or status != closed) → proceed to Stage 1 normally.

Guardrails

• Explicit /sessionstart invocation always runs full Stage 1–5. Stage 0 only short-circuits on implicit "pick up where we left off" opens.
• Stage 2.5 recursion-rule untouched. When Stage 0 falls through to Stage 1, the full verify-before-firing protocol runs unchanged.
• Sessionend unaffected. Sessionend writes frontmatter the same way regardless of whether the next session opens via Stage 0 or Stage 1.

See references/stage0_turn_start.md for the probe template, close-timestamp extraction strategy, and window rationale.

---

The 8 Stages

Stage 1 — Detect Department + Session Shape

See references/stage1_detect.md

• Auto-detect department from context, confirm rather than ask when confident.
• Detect session shape: dept-work (working inside one dept) vs. ops-infra (working on skills, infrastructure, the machine itself) vs. mixed. Shape drives Stage 2.5 depth.

Stage 1.5 — Load Dept CLAUDE.md

See references/stage1_detect.md — "Department → CLAUDE.md Path Mapping"

Primary orientation doc is the dept CLAUDE.md, NOT workspace-root CLAUDE.md.

• Read {dept}/CLAUDE.md in full (60-100 lines — well under context budget)
• Use its "Mission", "Current state", "Canonical files", and "Dept non-negotiables" to frame the Stage 3 brief
• If the file doesn't exist, fall back to workspace-root CLAUDE.md and flag the miss as a Stage 2.5 drift finding ("dept CLAUDE.md missing — scaffold next")

Stage 2 — Read Brain + Memory + Infrastructure State

See references/stage2_read.md

Single-shot brief loader (preferred path). Instead of issuing sequential Read calls per source, invoke one Bash script that emits all Stage 2 sections in a structured stdout block. Run the loader from the Brain root (or set CLAUDE_PROJECT_DIR to your {brain_root}) so the script resolves PROJECT_DIR correctly:

cd {brain_root}                     # or: export CLAUDE_PROJECT_DIR={brain_root}
.claude/skills/sessionstart/Scripts/brief_loader.sh --dept <dept> --shape <dept-work|ops-infra|mixed>

Sections emitted (always in this order, demarcated by ===NAME=== sentinels):

• DEPT_CLAUDE — dept-scoped CLAUDE.md (primary orientation doc)
• LATEST_SESSION — most recent session file frontmatter + closing block
• DECISIONS_PENDING — open / pending decisions for the dept
• DASHBOARD_HEADER — top-matter from Master Brain DASHBOARD.md
• AWARENESS_DEPT — current-date block from awareness.md
• MEMORY_INDEX — auto-memory layer MEMORY.md index, with lifecycle filter (terminal-state entries hidden)
• STATE_FILES — circuit-breakers + relevant infra config
• FLEET_PROBE — fleet host reachability (only on ops-infra / mixed)
• MEMORY_SIZE — auto-memory MEMORY.md size probe (warns near 16 KB consolidation cap)
• DASHBOARD_RENDER — agent-side Q1–Q4 render parity (latest-session-per-dept table)

Missing sections emit SKIP — reason instead of failing. Read the stdout block once; compose the brief from its sections. Do NOT re-read individual files after this — the loader output is the canonical Stage 2 snapshot.

Frozen-snapshot invariant: the MEMORY.md state captured by the loader is frozen for the session — do not re-read or re-weight mid-session. Any memory updates written during Stage 5 go to state/memory-corrections.log + the auto-memory layer; they do not retroactively alter this snapshot.

Fallback (sequential reads). If the loader script is missing or returns an error, fall back to sequential reads in this order:

1. Dept Brain: latest session file + decisions/decisions.md + pipeline/current-state.md
2. Master Brain: DASHBOARD.md header + awareness.md current-date block
3. MEMORY.md index (apply lifecycle filter manually — skip entries tagged terminal/closed/superseded)
4. Infrastructure state files (if session shape = ops-infra or mixed): {brain_root}/mission-control/state/circuit-breakers.json + any active process lock markers

Stage 2.5 — Ground-Truth Verify

See references/stage25_verify.md

For every memory or decision being surfaced as "open" or "P1", run the single filesystem check that would falsify it. Mark each as:

• ✅ confirmed — claim matches reality
• ⚠️ drift — claim doesn't match reality, memory needs update
• ❌ contradicted — reality shows the item is already resolved

Also runs infrastructure probes when session shape = ops-infra:

• Local model host alive? (curl -s {OLLAMA_HOST}/api/tags with 2s timeout)
• Any process lock in progress?

Optional probes (run if installed; skip silently if absent):

• Skill discoverability lint — {brain_root}/mission-control/skills/skill-discoverability-lint/Scripts/lint.py (optional). If present, surface any CRITICAL findings in the Stage 3 brief under "⚠️ Skill Drift". WARN/INFO: omit unless asked. Exit 0 always — never blocks session open. Not bundled in V1 profiles; install separately if you want the probe.
• Handoff closure audit — {brain_root}/mission-control/skills/handoff-closure-audit/Scripts/audit.py (optional). If present, surface DRIFT count only in Stage 3 brief if > 0. Full list on request. Not bundled in V1 profiles; install separately if you want the probe.

Recursion guard. A correction is itself a memory claim. Before any correction is queued for Stage 5, re-run the falsification probe against the source claim — including absence claims ("file X does NOT exist", "skill Y is NOT installed", "operator action required: do Z"). Restated claims surviving a context-compaction event are memory claims regardless of where they appear. Verify before firing, verify before restating. See references/stage25_verify.md §Recursion-rule.

Stage 3 — Deliver Brief

See references/stage3_brief.md

🟢 [Department] — Session Brief | [date]

## Infrastructure Snapshot   ← only for ops-infra / mixed sessions
[1-3 lines on machine state: local model host, circuit breakers, key open loops at infra level]

## Incoming State
## Open Loops
## Pending Decisions
## Pipeline Position

Stage 4 — Cross-Dept Flags

See references/stage4_flags.md

Verify-before-surface for HIGH flags. If awareness.md flags something HIGH, confirm against the implied source before surfacing. Stale HIGH flags are worse than no flag.

Stage 5 — Memory Hygiene

See references/stage5_hygiene.md

If Stage 2.5 found drift, surface a "Memory Drift — Auto-Corrected" block at the top of the brief with what changed. Default mode is auto-correct. Every correction:

1. Is logged to {brain_root}/mission-control/state/memory-corrections.log (timestamped, reversible)
2. Shows the old claim, the ground-truth finding, and the new memory body
3. Updates MEMORY.md index if the short hook needs rewording

Stage 6 — Capture Session Intent

See references/stage6_intent.md

Ask one question, write intent header to dept Brain.

Stage 7 — Render HTML

See references/stage7_html_render.md

Render the same synthesized brief to HTML in the operator's house theme. Markdown chat brief stays as-is (Reversibility #5 — additive). HTML is the archivable / shareable artifact.

• Theme: read {house_theme_path} first; drop the :root token block into <style> verbatim; use var(--*) references throughout (no inline hex literals). Default: {brain_root}/mission-control/audits/theme.md if no operator override.
• Output: {brain_root}/mission-control/state/briefs/{YYYY-MM-DD}-{dept}-sessionstart.html (append -{HHMM} if multiple on the same date).
• Write mechanism — prefer a shell heredoc / file write_text over the harness Write tool. Some agent harnesses (e.g. Claude Code's Launch preview) auto-surface Write-tool output in an in-app preview panel. If you intend the browser/file to be the deliverable and do NOT want a duplicate preview-panel copy, author the HTML and write it via cat > <path> <<'HTMLEOF' … HTMLEOF (or python write_text) — the panel surfacing is keyed to the Write tool, not to the file on disk. Harness-agnostic; harmless on harnesses without a preview panel.
• Same content, different render. Reproducibility #8 — render-target separation. Synthesis from Stages 1–6 is target-agnostic; markdown and HTML BOTH render from that state. Drift between them is a render-layer bug.
• Chat delivery: append a one-line link after the markdown brief.
• Conditional sections obey the same hide-when-empty rules as the markdown brief.
• Trust tags from Stage 2.5 (✅ / ⚠️ / ❌) render as badge-green / badge-yellow / badge-red pills per the theme spec.

If house_theme_path is not configured in the operator profile, skip Stage 7 — markdown brief stands alone.

---

Key Rules

1. Verify before surfacing. Memory + decisions + awareness flags all get a 1-line filesystem spot-check before they hit the brief.

2. Auto-correct drift by default. Log the corrections, show them at the top of the brief, offer a rollback command if the operator disagrees.

3. Shape-aware depth. Dept-work sessions get a light infra check. Ops/infra sessions get the full machine probe. Don't run the full probe when opening a sales session — it's noise.

4. Read first, speak second. Never deliver the brief from memory or the dashboard alone. The dashboard can lag reality.

5. Open loops are sacred — if sessionend wrote it under "Open Loops — Next Session," surface it. But verify each one in Stage 2.5 before declaring it still open.

6. Don't pad. Empty sections are omitted. Short brief that's accurate beats long brief that's noise.

7. Cross-dept flags are surgical. Max 3 bullets, verified-before-surface.

8. The brief is oriented to action, not history. Incoming state is 1-2 lines. Open loops + decisions + verified infra state are the main content.

9. Dept CLAUDE.md is the primary orientation doc. Workspace-root CLAUDE.md is fallback.

10. Corrections verify before firing — recursion rule. A Stage 5 auto-correction is itself a memory claim and gets the same falsification probe its source claim got — applied before mutation, before logging, before restatement.

11. House theme governs HTML render. Stage 7 HTML render always reads {house_theme_path} first as the canonical token block — no inline hex literals, all var(--*) references. Same house theme any other operator-facing HTML render uses (audit reports, full-business-audit, command center). Parity #2 — the house style is one of the architectural surfaces skills track.

---

Scope Contract

Dimension	Scope
Read paths	{dept}/brain/sessions/ + decisions/decisions.md + pipeline/current-state.md (per-dept) · {brain_root}/mission-control/DASHBOARD.md + awareness.md + sessions/ + decisions/ + state/*.json · workspace-root + dept CLAUDE.md · auto-memory layer MEMORY.md + referenced memory files · {house_theme_path} (Stage 7 render only)
Write paths	Session intent header → {dept}/brain/sessions/YYYY-MM-DD-{slug}-{dept}.md (Stage 6 only) · {brain_root}/mission-control/state/memory-corrections.log (Stage 5 drift correction, append-only) · auto-memory layer {user,feedback,project,reference}_*.md updates + MEMORY.md index · {brain_root}/mission-control/state/briefs/{date}-{dept}-sessionstart.html (Stage 7 HTML render)
MCP / tool surface	Read, Bash (filesystem probes + brief loader + dashboard render; optional probes: skill-discoverability-lint, handoff-closure-audit — both skipped silently if not installed), Write/Edit (limited to session intent header + memory-corrections.log + auto-memory layer + briefs dir)
Network egress	Local model host probe only (Stage 2.5 infra check; curl -s {OLLAMA_HOST}/api/tags with 2s timeout); no external network
Surface	Claude Code primary
Credentials	None required for reads; no secrets written
Escalation trigger	Stage 2.5 probe reports ⚠️ drift on >3 memory claims → pause for explicit operator ack before running Stage 5 auto-corrections; Stage 6 never auto-captures intent — always asks

<!-- AUTOLINK-START — Obsidian wikilink graph backfill (idempotent; safe to re-run) -->

References (auto-wikilinks)

• [[BRAIN_GOVERNANCE]]
• [[DASHBOARD]]
• [[MEMORY]]
• [[NAMING_CONVENTION]]
• [[awareness]]
• [[current-state]]
• [[decisions]]
• [[stage0_turn_start]]
• [[stage1_detect]]
• [[stage25_verify]]
• [[stage2_read]]
• [[stage3_brief]]
• [[stage4_flags]]
• [[stage5_hygiene]]
• [[stage6_intent]]

<!-- AUTOLINK-END -->