Session Observatory — Live Event Logging

Most session retrospectives are written from memory at the end. Memory is lossy. This skill captures the habit of logging events as they happen, so the retrospective has real data to work with.

When to Activate

• At session start — log start <mission-name> as one of the first Bash tool calls, so the started_commit is recorded.
• When friction hits — hook fired repeatedly, tool failed, command took longer than expected, user corrected you.
• When something works unusually well — a technique that you'd recommend to future sessions.
• Before and after irreversible actions — commits, pushes, destructive ops.
• On progress markers — every ~10% through a long job.
• At session close — log end, then run generate.mjs to produce the SESSION-RETRO.md.

Commands

# Start
node tools/session-retro/observe.mjs start "<mission-name>"

# Log an event
node tools/session-retro/observe.mjs event <kind> "<label>" '<json-payload>'

# Check in on the log
node tools/session-retro/observe.mjs status

# End (archives current.jsonl to a dated file)
node tools/session-retro/observe.mjs end

Event Kinds (standard)

Kind	When	Example label
friction	Something took more effort than it should have	"read-before-edit hook fired N times"
win	Something worked notably well	"idempotent pipeline saved 3 re-runs"
correction	User or hook corrected direction	"user reminded: no Co-Authored-By"
tool-use	Notable tool invocation	"installed better-sqlite3 as optional dep"
decision	Judgment call deserving review	"committed 1,881 generated files"
gap	Missing skill/agent/chipset	"no batch-rewriter for N-script cascade"
checkpoint	Progress marker for long-running work	"Pass 2 chapters: 300/602"
tokens	Token consumption (dual-writes to token-budget)	"pass-2-chapter-gen"

Free-form kinds are allowed but less useful for aggregation. Prefer the standard list.

Tokens event — special positional form

tokens is the one kind where positional args are reinterpreted to keep the call site short and to dual-write to current.tokens.jsonl:

# observe.mjs event tokens <in_tokens> <out_tokens> [label]
node tools/session-retro/observe.mjs event tokens 45000 3200 "pass-2-chapter-gen"

The entry lands in both current.jsonl (as a tokens event) and current.tokens.jsonl (as a budget entry), so token-budget.mjs status and the retrospective observations stay in sync without a second call.

Payload Conventions

Payloads are free-form JSON. Common fields:

• count — numeric count of the occurrence
• file / files — paths involved
• duration_min — wall-clock time
• cost / tokens — resource spend
• impact — "workflow" | "output" | "blocked" | "none"
• opportunity — what skill/agent would have helped

Retroactive Logging

If you didn't log live, you can backfill at session close by reviewing what happened and calling observe event for each remembered moment. Better than nothing. The timestamp will be session-end rather than event-time, but the kinds and labels are still useful.

Integration with Retrospective Generator

Events from current.jsonl are consumed by tools/session-retro/generate.mjs and rendered in the SESSION-RETRO.md under Observations, grouped by kind. The human-authored sections (What Worked, What Could Be Better) then reference specific events as evidence.

Example — This Skill's First Use

The release-history feature session logged 33 events retroactively:

• 6 friction (read-before-edit fires, 11-script cascade, no checkpoint for LLM run, ghost-seeding cascade, over-eager regex, hidden drift)
• 4 wins (background Monitor, idempotent pipeline, recursive self-improvement, fresh-repo demo)
• 3 corrections (no Co-Authored-By, dev not main, link chapter files)
• 4 decisions (publish full tree, SQLite default, allowlist gates, Commits column)
• 4 gaps (batch-rewriter, portable-schema, decision-framework, session-token)
• 3 tool-uses (better-sqlite3, claude -p, git tags)
• 9 checkpoints (pass-by-pass progress)

Those 33 events became the concrete evidence for the retro's "What Worked" / "What Could Be Better" / "Recommendations" sections.

Anti-patterns

• Logging every command. Too noisy. Log inflection points.
• Vague labels. "Problem happened" is useless. "Retro extractor missed H3 headings — 57 files hid as a result" is useful.
• Skipping at session close. The log isn't useful if end never runs and the JSONL never archives. Generate the retro before context dies.

Related

• tools/session-retro/observe.mjs — the logger itself
• tools/session-retro/generate.mjs — the retrospective generator
• decision-framework-invoker — often co-fires with decision events