Tags

v2.5.0

2.5.0 — the measurement chain lands, and its first job was to correct itself

Rows become an answer. eval/analyze.mjs aggregates per arm and deliberately
stops short of a verdict; eval/verdict.mjs applies six gates whose thresholds
were frozen before the run, including one that reads the operator's journal so a
hand-rejected pass is not silently a zero; eval/report.mjs renders the result as
one self-contained HTML file with no network and no build step. When a gate
withholds the verdict, the arm-comparison chart is not drawn at all — not greyed
out, not captioned — and a test asserts that absence.

The chain's first real finding was about the tool that produced it. `usage`
describes an API RESPONSE, but a response reaches the transcript as several
entries — one per content block — each carrying an identical copy of it. Priced
per entry, the same tokens were charged two to five times over: on 19 frozen
transcripts the run total falls from $3,525.55 to $679.35. The inflation is not
flat across token classes and does not cancel in a ratio of two arms, because
the factor IS the number of content blocks per turn — how many tools the model
reached for, which is the variable the doctrine sets out to move. The
unattributed bucket, which exists to report what falls outside every row, was
itself dropping tool calls and per-request server-tool cost; it no longer does.

So the honest summary of this release is uncomfortable and stays on the record:
EPIC-009's first verdict came out AGAINST the doctrine, it was computed from
figures this release proves were too high, and the skew correlates with the arm.
The chain is ready. The number it produced is not yet an answer.

v1.3.0

rejected/task-090-off-1

Odrzucony przebieg §7: P1 bez podstawionego numeru taska (0XX). Sesja eb076d7c.

v2.4.0

2.4.0 — the measurement can finally have two sides

`tune-context experiment` switches the doctrine and RAG off and back on, so the
same work can be run in both arms: hook registrations, the managed CLAUDE.md
block, doctrine skills, agents and the project .mcp.json. Every touched path is
fingerprinted and copied aside before it is modified, so a `set` interrupted
halfway is still reversible; `restore` puts the bytes back or refuses, naming
what would be lost. `verify` reads the off arm as "off" instead of as a broken
install, and `init` refuses while an arm is set.

Cost stops being a proxy. eval/prompt-metrics.mjs turns transcripts into one row
per prompt — time split into queued and work, tokens split by cache class,
model, effort, tools, delegations — and eval/pricing.mjs prices them at API list
price, cache multipliers included. It is NOT a bill: a Pro/Max plan charges its
flat fee, and the fields say so in their names. The breakdown carries the epic's
thesis in one line: on the local baseline, cache reads and cache writes are
nearly the whole figure while plain input is 0.07% of it.

Rows know which arm they came from. A branch shaped exp/<task>/<arm> says which
arm was intended, the TASK-087 ledger which was set, and a disagreement leaves
the arm unresolved with the reason on the row rather than picking a winner.
pass_order rides along, because the second implementation of a task is cheaper
for a reason that has nothing to do with the arm.

Pre-release review caught two measurement bugs: `restore` now ends a run even
when it parked nothing (the "on" arm's window used to stay open forever, and
only that arm's, which made the contamination one-sided), and a cost that is not
a finite number is refused instead of being counted as priced and turning the
whole run's total into $NaN.

v2.3.1

2.3.1 — third context notice at 1.5x, graded by rung

Middle rung at 1.5x the threshold (180k by default) joins 120k and 240k,
cutting worst-case silence between notices from 120k to 60k. The ladder is
multipliers, so TUNE_CONTEXT_REMIND_TOKENS scales all three.

Each rung is louder than the last: wait for a natural break, raise it at the
next pause anyway, say it this turn. Fixes the second notice calling itself
'second and final', which a three-rung ladder makes false.

Docs describing the reminder as firing once (README, tune-context skill) —
wrong since 2.1.0, corrected here.

v2.3.0

v2.3.0 — the adoption report can tell "RAG installed and ignored" from "no RAG here": a project .mcp.json probe behind a soft `~` verdict, guarded by a temporal check (config must predate the sessions), explicit rejections, observation precedence and the PO-frozen session floor now applied per project; search counted per project so one RAG-using repo no longer reads as coverage for the rest

v2.2.0

v2.2.0 — authored CLAUDE.md sections live below the managed block (never regenerated); verify content-compares deployed hooks/agents/skills against the installed package, plus a read-only `tune-context verify` and a guard for files still in the old layout

v2.1.0

v2.1.0 — context reminder fires a second and final time at 2x the threshold (240k default); systemMessage added best-effort (does not render for UserPromptSubmit)

v2.0.1

v2.0.1 — fix: npm-installed CLI ran nothing (entry-point guard vs extensionless bin symlink)

v2.0.0

v2.0.0 — install integrity, fact fidelity, and a context reminder that measures

EPIC-008: the installer and detector stop producing signal they then measure.
Plugin+CLI dual registration is detected and no longer fakes a churn verdict;
detect writes the package manager it actually found (pnpm/yarn, through to
permissions.allow); CLAUDE_CONFIG_DIR and the allow list have one definition
each; rejected .mcp.json servers are honored; PreCompact is fail-safe with
collision-free state paths and 30-day retention; cli/ gains unit coverage;
docs and CLI UX close the gaps.

TASK-079: new context-size reminder (UserPromptSubmit). The model cannot see
its own context usage, so a hook reads the transcript tail, sums the tokens
actually re-sent, and once past a PO-frozen 120k injects one short note
suggesting /compact — once per crossing, re-armed only after a compaction.

BREAKING: `tune-context` with no subcommand no longer runs `init`; `init` is
now required explicitly. Also adds `tune-context --version`.

See CHANGELOG.md, including "Upgrading from 1.1.0".

v1.1.0

v1.1.0 — CLAUDE_CONFIG_DIR consistency + adoption-report hardening

EPIC-006: CLAUDE_CONFIG_DIR honored across the whole hook lifecycle (registration
+ state paths); verify cross-checks hook wiring; smoke covers both path modes.
EPIC-007: adoption report splits code search from shell plumbing, earns its
verdict (sample-size + availability aware), and stops printing false negatives.

See CHANGELOG.md.

v1.0.1

...

v1.0.0

tune-context v1.0.0 — first public release (npm + Claude Code plugin marketplace)