TAUSIK

English

Русский

English

Русский

Appearance

A "session" is TWO things ​

Decision #223. The word "session" in TAUSIK fuses two concepts with different fates, and the fusion is precisely what made it impossible to answer which of them could be dropped.

What it isA property of
Work continuityhandoff: what was done, what is in flight, what comes next, what to warn abouta property of the WORK
Agent context hygienethe 180 active-minute limit (SENAR 9.2), the 200-call capacity, the checkpoint counter (SENAR 9.3)a property of the agent's CONTEXT WINDOW

The two halves are no longer coupled. They used to be, in three ways, and every one of them looked harmless:

  • session_handoff refused when the window was closed. It is the other way round: an agent that has hit the 180-minute limit is exactly the one who most needs to write down where it stopped. The refusal destroyed the very document the limit exists for.
  • Saving a handoff zeroed tool_call_count. Resetting the counter is hygiene, and it is now a separate named operation, reset_checkpoint_counter, rather than a side effect of writing a document. /checkpoint still does both — but explicitly, and in that order.
  • The capacity gate silently waved everything through when no session was open. See below.

Why continuity was NOT dropped ​

The task this split grew out of proposed discarding the first half: its role, the argument went, had been taken over by the git projection. That premise was refuted by measurement, not by opinion — four facts:

  1. The projection is off by default. state.auto_export has no default anywhere: DEFAULT_CONFIG carries no state key at bootstrap, and the resolver returns False on any error. A fresh project has no projection.
  2. Sessions are not projected at all. ENTITY_DIRS = (epics, stories, tasks, decisions, memory). Even with the projection enabled, a handoff never reaches the tree.
  3. The handoff fields have no other home. The tree carries "what changed" — task status, its log, decisions, memory. next_steps, warnings and in_progress[].state ("step 3 of 5") have a column nowhere except sessions.handoff.
  4. The projection calls itself incomplete. Its docstring enumerates what it does not cover by name, coverage rests on ~18 manual call sites, and every trigger is fail-open.

The condition for a future drop is named explicitly: the handoff must gain a projected home, i.e. sessions must enter ENTITY_DIRS. The condition is checked by tests/test_session_two_halves.py::TestTheDropConditionIsNotMet — if it goes red, that is not a regression but a signal to revisit #223.

What was refuted in the task's favour: calibration does not depend on sessions at all. calibration_drift and per_tier_metrics read only the tasks table (call_budget/call_actual/tier/completed_at), without a single join to sessions.

No session is NOT "unlimited" ​

The capacity gate used to return early when no session was open: the 200-call check stopped checking and said nothing about it. Worse, it inverted the incentive — the cheapest way to get around a capacity refusal was to end the session and not start a new one.

Now the absence of a session is an explicit refusal for a task that declared a budget. A task without a budget starts as before: the gate has an opinion only about what asked to be accounted for.

What else silently switches off without an open session ​

This matters because every mechanism below fails by writing a zero, not an error — the only way to notice the loss was an empty report:

MechanismWhat happens
Usage telemetry (posttool_usage)no rows written at all ⇒ tasks.cost_actual_usd and tokens_actual stay zero for every task, and the cost budget never fires
Token metrics.tausik/token_metrics.jsonl stops growing; tausik metrics tokens prints "no data yet"
Model pinningstarted_model_id / done_model_id stay NULL, model_mismatch never fires again
The brain's "this session" slicecomputed over all time and passes itself off as per-session — not empty but wrong; the most dangerous case in this list
Audit cadence (SENAR 9.5)"3 sessions since the last audit" never arrives

That is why the capacity gate's refusal names tausik session start — one action restores everything above.

What did not change ​

The sessions table, session_start/session_end, handoffs, and the metrics that slice by session (throughput as "tasks per session", session_hours, the token-metrics window "last N sessions", audit cadence). The split removed the coupling, not the data.

See also ​