TAUSIK

English

Русский

English

Русский

Appearance

English | Русский

What changed in 1.9 ​

A page for whoever is upgrading. Before reading what breaks, it is worth knowing what it was for.

What 1.9 is, in three points ​

📏 A release that stopped taking its own word for things. 1.8 answered the question "does the framework enforce discipline". 1.9 asks no question — it makes two statements about the product and requires each to be measured: a substantial token saving, and higher development quality on any model.

And one of the two is MEASURED ONCE — and on this pair there is no saving; said here rather than in a footnote. Telemetry still produces no figure: of 57,251 telemetry rows, 233 (0%) carry input tokens, and no "without TAUSIK" baseline exists in it at all. In session #225 it was 0 rows out of 4,777 — the instrument moved, and did not reach a number. For telemetry this is the ABSENCE of a quantity, not a measured zero. The figure came from a different instrument — a paired replay on a fixed corpus (protocol §7, in Russian): ten identical questions about the repository, one commit, one model, the harness with its rag-first nudges against the same harness without them. With the nudges it cost more: 198,848 against 195,055 tokens of new context and output (+1.9%), 326,323 against 292,715 bytes of exploration results (+11.5%), and search_code, the tool the nudges recommend, was never called with or without them (0 of 62 and 0 of 76 calls). That is one reading on one corpus, in this pair — not a refutation of saving in general: a repeat of the same condition varied by 13%, more than the delta itself. The promise stays what the release undertook to measure; what is published as fact is only what was measured: no saving on this pair, and the nudges do not change tool choice.

🔢 Zero stopped passing itself off as a measurement. A quantity that cannot be obtained is now ABSENT, not nought. That sounds like a nicety until you look at the numbers: the telemetry asserted 55,471 times that work had cost $0.00, and model pinning had NEVER fired in 231 sessions. Both defects looked exactly like "a feature nobody needed".

🖧 Guarantees are stated by rule, not by host. Five environments were handed the same rules text promising automatic enforcement; on two of the five that was untrue (measured in session #225; Codex became the sixth host later in the release, on the same terms — see below). Each host now reads what is actually deployed for it — and by rule, because "everything here is instructions" is untrue too on a host without hooks: closing a task IS refused there.

Full list of changes: CHANGELOG.md. This page is not a retelling: the 1.9 section of the CHANGELOG holds more than 250 entries (the lower bound is held by tests/test_release_notes_1_9.py), and what is selected here is what changes the experience of UPGRADING.

BREAKING CHANGES ​

This section is the single source of the future tag's text. That is exactly why it exists: in 1.8 the breaking change to trust tiers stayed in the CHANGELOG and never reached the tag's notes, and a published tag cannot be re-cut.

1. A write gate that cannot read the DB now refuses instead of allowing ​

Before. If scope_write_gate could not open .tausik/tausik.db — locked, corrupt, unreachable — it let the write through. Silently.

Now. It refuses. A gate that could not check has no business reporting the same outcome as a gate that checked and allowed.

Does this affect me. Yes, if you work with the database locked or unreachable: a write that used to pass will now be declined, with the reason. Unlock the DB (usually by closing a parallel process) or disable the gate deliberately in the config. TAUSIK_SKIP_HOOKS=1 still works and is still telemetered.

2. The Notion transport is gone ​

Before. The shared knowledge base could be mirrored to a Notion workspace: tausik brain init/status/sync/move/draft/publish, the tausik-brain MCP server with seven tools, the /brain skill and two hooks that searched the brain before a web fetch.

Now. None of that ships. The shared store is local and file-based — ~/.tausik-knowledge — and --global on decide and memory add is the only way a record leaves this project. tausik knowledge import-brain still reads the local mirror file ~/.tausik-brain/brain.db, so nothing already mirrored is lost.

Does this affect me. Only if you ran the brain. Bootstrap over a 1.8 checkout removes the stale tausik-brain entry from every host config it manages, so your IDE stops logging an MCP error for a server that no longer exists. Run tausik knowledge import-brain once if the mirror holds records you want in the shared store.

What else changes on upgrade ​

Database schema: 44 → 62 ​

Eighteen migrations apply automatically on first access (the figures here are held by tests/test_release_notes_1_9.py against what tag v1.9.0 shipped).

Known issue, fixed in 1.10. Upgrading a 1.8 database with 1.9.0 crashes at v53 (duplicate column name: tz_ref) and leaves schema_version at 44 (github#51, gitlab#18). Upgrade straight to 1.10: the chain skips an ADD COLUMN the table already has and stamps the version after every migration, so an interrupted upgrade resumes where it stopped. A database a 1.9.0 upgrade already broke is carried up by the same run. Each is preceded by a backup at .tausik/tausik.db.bak.v<old>; spares are cleared with tausik db prune --keep N.

The most visible is v58: the token and cost columns in usage_events became nullable. NULL means "not measured", and that is not the same as 0. The migration turned 55,307 rows — provably never measured — into NULL and touched no row carrying a real number.

The last ones are small and worth knowing: v59 lets a graph edge say it was OBSERVED rather than inferred; v60 records a test that was never seen red; v62 puts an index on tasks.defect_of — without it tausik status on a database with a history (1,504 done tasks) spent 5.3 s in one EXISTS, and the SessionStart hook hit its 6 s timeout and silently delivered no context; with the index the query takes 5 ms, status 0.3 s, the hook 0.9 s; v61 adds tasks.tracker_refs — the first column added to tasks after v43, which is how the post-migration that rebuilt tasks from a frozen v43 column list was caught erasing any later column. That rebuild is fixed in the same release — it re-applies every later column after itself — so a database upgrading from before v43 keeps the column (tests/test_migration_v43_model_mismatch.py).

Cost and model: what used to stay silent ​

  • Zero no longer passes as a measurement. calculate_cost_usd returns absence for an unpriced model and 0.0 for one priced at zero; both used to return 0.0, and nothing could tell them apart.
  • Per-session spend was inflated twofold on 156 of 160 sessions; the total now says which part of it was computed on the superseded arithmetic.
  • Prices live in the project's config, and its word beats the shipped table. A rate is a PAIR of numbers (input and output), not one.
  • The session's model is recorded. sessions.model_id was empty in all 231 sessions, so model pinning (RENAR 10.13) never fired. The source is a chain: TAUSIK_AGENT_MODEL → host variables → the host's provider. The model is NOT inferred from the host's name: Claude Code pointed at another endpoint runs that endpoint's model.

Gates: silence stopped being an option ​

  • A gate with neither an implementation nor a command blocks instead of staying quiet.
  • A rejected command override blocks; substituting the built-in default made the refusal invisible.
  • ruff runs on the verify trigger too, not only on commit.
  • New gate cross_model_parity: a difference between hosts that share an extension point must be NAMED. It does not demand sameness — Cursor has no extension point — but an undeclared difference blocks. It fires only on host-layer edits.

Running the tests ​

  • The lane is parallel by default (-n auto in addopts). pytest-xdist is required; without it pytest exits on "unrecognized arguments: -n".
  • Selection follows IMPORTS, not names — the cheap run reaches wider, and half the corpus stopped being unreachable.
  • pyproject.toml no longer promises tausik verify --full, which never existed.

New commands ​

CommandWhat it does
tausik coherenceThe repository-level question no gate was asking: what in the tree stopped adding up
tausik audit evidenceClosure-receipt citations no longer rot in silence

Codex is a sixth host — and the claim stops where the measurement stopped ​

bootstrap.py --ide codex (and --ide all) scaffolds Codex CLI: the tausik-project MCP server registered in the PROJECT's .codex/config.toml, the same skill set as Claude under .codex/skills/, sub-agents generated from the canonical Markdown into .codex/agents/*.toml, and .codex/hooks.json built from the SAME hook declaration Claude's profile uses (absolute paths, because Codex has no workspace variable). Support is a release promise only because a real Codex host proved it, not because the files exist: in the live acceptance run Codex called tausik_status and got the structured reply, its catalog listed i-have-adhd and the tausik-reviewer agent, and the forbidden Path('outside.txt').write_text(...) was refused before the file existed.

The boundary, stated once here. That refusal happened under a hook profile the user had TRUSTED in Codex. The same operation with the generated profile present but untrusted ran to completion — no hook fired. Nothing on disk tells the two states apart, so Rule 1 and the write ACL on Codex are hard once you trust the project hooks, and an untrusted profile enforces nothing. The enforcement matrix carries that condition on both rows and a test keeps it there.

The agent's own quickstart ​

agent-quickstart.md (EN/RU) is written for the agent reading it first: connect on its host — every host bootstrap scaffolds, Codex with its trust condition — check, then the cycle as exact calls with the replies and the refusals it will actually see, each quoted from a live run and held to the code by tests/test_agent_quickstart.py.

Context: across projects, across sessions, through compaction ​

The global config no longer carries another project. In 1.8 the user-tier ~/.tausik/config.json could bring a weakening enabled for one project into yours (auto_verify: true from a neighbour). User-tier weakening now lives in a projects section keyed by the project's absolute path; tausik doctor labels every effective weakening MACHINE-WIDE or project-scoped and lists foreign entries separately.

task start brings memory by relevance. On top of the recency tail in CLAUDE.md, start, resume and task show print a Relevant memory (N) block built from the task's own declaration; an empty answer is named with the terms tried.

The rules file gained a compaction section. It says what to carry verbatim through context compaction. The rules file is preserve-if-exists: on an already-bootstrapped project the section appears only after the generated file is deleted and bootstrap re-run.

knowledge export --redacted masks paths, addresses, URLs and project names with typed placeholders before shared memory is published.

Caveman mode is a response contract, not only a length: the shape done → verified by → left → your call, five named exceptions and a pre-send check — inside the same directive, no second mode.

Output economy: two levers, both off on purpose ​

read_ledger (do not pay twice for an unchanged file) and a byte cap on command output both ship, both are off by default, and each has a decision with a number behind it. Turning them on is your choice, not our default.

See also ​