Clock-aligned retros tell you whether you worked well. They don’t tell you whether you worked on the right thing.

Goal drift is the failure mode no one talks about: the agent is busy, productive, generating output — but it left the declared goal behind three turns ago and nobody noticed until the session ended. By then, the context is spent.

This post adds a session goal layer to the clock-aligned retrospection framework. Every automatic retro now opens with a verbatim goal read and a three-way verdict: ON TRACK / DRIFTED / DONE.

See also: Break-time retrospection, Personal Agentic Process.

The problem

The agent-retro extension fires retro prompts at clock boundaries. Those prompts ask: was the work good? They never asked: was it the right work?

The gap: five retros could fire in a two-hour session, each reporting “no backtracking, good engineering” — while the agent silently pivots from the declared goal to an adjacent rabbit hole that looked interesting at the time.

The fix is simple: every retro reads the session goal and states the verdict. If the agent can’t say “ON TRACK” without hesitation, that’s the signal.

Design

Goal file

Each session gets one goal file, co-located with its JSONL session file:

~/.omp/agent/sessions/<cwd>/<uuid>.goal.md

Format:

goal: <one sentence>
started: YYYY-MM-DD HH:MM
constraint: <none | X hours | hard deadline HH:MM>
context: <continuing: <what> | fresh start>

Written once at session start via /session-start or PAP step 4. Also archived to ~/sessions/YYYY/MM-Month/YYYY-MM-DD_HH-MM-SS - <slug>.md for the permanent record.

Why UUID-keyed, not a single shared file

The naive design is one file: ~/.claude/session-goal.md. It works until you run two OMP sessions in parallel — both write to the same file, last writer wins, retros in both sessions read the wrong goal.

OMP session IDs are Bun.randomUUIDv7() — time-ordered, unique, stable. The UUID doesn’t change when you /rename a session (rename only updates the title field in the JSONL header). So <uuid>.goal.md is collision-free and survives renames.

The file accumulates alongside the JSONL. No cleanup needed at close-session — it gets pruned when the JSONL is pruned.

Retro ladder integration

Every level of the retro ladder reads the goal file:

Level Cadence Goal check
retro-mini per break 3-line pulse: goal verbatim + ON TRACK / DRIFTED / DONE
retro per break (full) §0 full alignment: drift cause + next-segment intent
retro-long hourly §2 verdict + time-constraint pacing
close-session §3.5 end of session ACHIEVED / PARTIALLY ACHIEVED / MISSED / SUPERSEDED → archived to Logseq

Extension: content injection

The agent-retro-omp extension already injects retro prompts via sendUserMessage. Adding goal content to the same message costs nothing and saves the agent a file-read tool call on every retro.

At session_start, the extension captures the session file path:

pi.on("session_start", (_event, ctx) => {
  sessionFile = ctx.sessionManager.getSessionFile() ?? undefined;
});

At each fire(), it reads the goal file and embeds the content as HTML comments before the retro prompt:

const goalFile    = sessionFile?.replace(/\.jsonl$/, ".goal.md");
const goalContent = goalFile && existsSync(goalFile)
  ? readFileSync(goalFile, "utf-8").trim()
  : "NOT SET — run /session-start to declare one.";

const sessionCtx = sessionFile
  ? `<!-- session-file: ${sessionFile} -->\n` +
    `<!-- session-goal: ${goalFile} -->\n` +
    `<!-- session-goal-content:\n${goalContent}\n-->\n\n`
  : `<!-- session-goal-content:\n${goalContent}\n-->\n\n`;

The retro message arrives pre-loaded with the goal. The agent reads <!-- session-goal-content --> directly and proceeds — no bash, no file read, one less tool call per 10-minute boundary.

The path is still embedded in <!-- session-goal: PATH --> for cases where the agent needs to write or update the goal.

Two-tier fallback

Auto-fired retros (from the extension) always have the <!-- session-goal-content --> comment. Manual invocations (/retro, /retro-mini) don’t. The retro prompts handle both:

  1. Comment present → read content directly from <!-- session-goal-content -->
  2. Comment absent → ls -t ~/.omp/agent/sessions/*/*.goal.md 2>/dev/null | head -1 then read that file

If both fail: goal is undeclared — flag it as a process gap and prompt /session-start.

What didn’t work

~/.claude/session-goal.md as a live pointer — the original design had the extension point at a fixed path and all retro commands read from there. It works for a single session but fails silently with parallel sessions. The smell: the retro-long for session A reports the goal of session B, which wrote the file more recently. UUID-keyed files eliminate the class of bug.

session-goal-$OMP_SESSION_NAME.md — requires the user to set an env var before launching OMP. Env vars don’t survive across bash tool calls (each call spawns a new MSYS2 shell) unless exported in the persistent shell session. More importantly, they require manual discipline. The UUID file requires no env var — the extension derives the path from ctx.sessionManager.getSessionFile() which is always available.

The verdict in practice

When a retro fires and the goal is present, step 0 produces something like:

Goal: Build the per-session goal tracking framework
Verdict: ON TRACK — currently wiring the extension injection
Next segment: commit agent-retro-omp, update retro commands, push

When the goal is absent (file not written yet):

Goal: NOT SET — run /session-start to declare one.

The “NOT SET” signal embedded in the message means the agent flags it immediately, in the retro turn, without an extra round-trip.

OMP internals used

Source

This post was written during the session that built the framework. The retro fired six times. Goal verdict at each fire: ON TRACK.