trajectory-* accepts invented IDs and reports them completed: every start returns ID 0, and '!!!not-an-id!!!' completes with 'EWC++ consolidation applied'; nothing persists

Summary

The trajectory-start / trajectory-step / trajectory-end lifecycle reports success for input it cannot have processed:

• Every trajectory-start returns the same ID: 0.
• trajectory-step and trajectory-end accept arbitrary invented IDs — including !!!not-an-id!!! — and report the step recorded and the trajectory completed.
• trajectory-end --success false reports ✅ Success. So do 0, no, and nonsense. The flag has no effect on the reported outcome.
• Quality: 80% is constant across every invocation.
• Nothing persists: after three starts, a step, and seven ends, intelligence stats still reports Trajectories: 0, Active: 0.

Version: agentic-flow 2.1.2, Node v22.23.0, macOS 15.

Reproduce

# 1. IDs are not unique
for i in 1 2 3; do
  npx agentic-flow hooks intelligence trajectory-start "task $i" --agent tester | grep 'ID:'
done
# 📝 ID: 0
# 📝 ID: 0
# 📝 ID: 0

# 2. Invented IDs are accepted
npx agentic-flow hooks intelligence trajectory-step "traj-TOTAL-FICTION-42" --action "did a thing" --reward 0.9
# 📍 Step Recorded
#    Action: did a thing
#    Reward: 0.9
#    Trajectory: traj-TOTAL-FICTION-42

npx agentic-flow hooks intelligence trajectory-end "!!!not-an-id!!!" --success false
# 🏁 Trajectory Completed
#    ✅ Success
#    Quality: 80%
#    Learning: EWC++ consolidation applied

# 3. --success is ignored
for v in true false 0 no nonsense; do
  npx agentic-flow hooks intelligence trajectory-end "traj-fake-$v" --success $v | grep -E 'Success|Fail'
done
# ✅ Success   (x5)

# 4. Nothing persisted
npx agentic-flow hooks intelligence stats | grep -A2 Learning
# 📈 Learning:
#    Trajectories: 0
#    Active: 0

Why this is worse than a no-op

Each of these prints a specific, plausible claim about work performed:

• 📍 Step Recorded … Trajectory: traj-TOTAL-FICTION-42 — echoes back an ID that was never issued, as if it had been looked up.
• 🏁 Trajectory Completed … Learning: EWC++ consolidation applied — names a specific learning mechanism as having run, for a trajectory that does not exist.
• ✅ Success for --success false — inverts the caller's own input.

A caller driving this API has no way to detect that none of it happened. There is no error, no non-zero exit, and the output is more specific than a generic acknowledgement would be.

Suggested fix

1. Issue unique trajectory IDs, and have trajectory-step / trajectory-end reject IDs that were never issued.
2. Honour --success, and reject values outside true|false.
3. Report a real quality figure or drop the field — 80% for every input, including --success false, is not a measurement.
4. Persist trajectories, or make intelligence stats reflect the fact that they are not persisted.

The round-trip test that catches this class permanently: start a trajectory, capture its ID, end that ID and assert success; then end a fabricated ID and assert failure. The second half is the part that matters — the first half already "passes" today.

Related

• #204 — pattern-store in the same intelligence group reports 💾 Pattern Stored with an HNSW index claim and persists nothing; intelligence stats likewise shows Patterns: 0 beside a 45-pattern store.
• #201 — hooks metrics counters not wired to the commands that should update them.
• ruvnet/ruflo#2908 — sibling package: writes are durable but never indexed, so search cannot find them.

Start with the trajectory-start, trajectory-step, trajectory-end, and intelligence stats command entry points and run the reproduction commands in this issue. Trace how IDs, --success, quality, and persistence are handled. Done means issued IDs are unique, fabricated IDs fail, --success is validated and reflected accurately, and stats reports completed trajectories.