Skip to content

release: ai 33f312a to production - #197

Merged
jm-rivera merged 3 commits into
mainfrom
release/2026-11-14-2
Nov 14, 2026
Merged

jm-rivera merged 3 commits into
mainfrom
release/2026-11-14-2

Conversation

@jm-rivera

Copy link
Copy Markdown
Contributor

Releases ai at 33f312a to production, cut on release/2026-11-14-2: three pull requests after the release at 0fcbd78.

Release order: production apis already serves catalogue 2026.11.1 with order on line_chart, bar_chart and table (checked on GET /api/presentation/release after DevelopmentFinanceObservatory/apis#302), which #192 needs. Studio reads new_plan once DevelopmentFinanceObservatory/website#145 lands; until then it keeps the earlier card's baseline, as today. No migration, setting or secret change. A chart drawn before this release has no order in its stored intent, so asking for the same chart again draws one new card.

Testing: CI passed on each pull request, and Deploy dev passed on 33f312a. After the production deploy, ask Studio for "a bar chart of the 10 largest recipients of French aid in 2023" and check that the bars run largest first.

…s tolerantly (#189)

* fix(harness): forward the platform's unit as JSON and read facet units tolerantly

The harness refused a stored card whose `fact_summary.unit` carried a member
its wire model did not know, so every new member dfo-apis added to a unit
broke Studio charts until the harness caught up. `FactSummary.unit` now takes
a unit name or any JSON object and forwards it to Studio unchanged; Studio
reads it as display text. An object over 512 bytes is replaced by the cut
marker when the fact summary is over its cap.

The compiler reads a facet unit's currency, prices and base year strictly and
`deflation_basis` as wording, and ignores other members. A facet unit it
cannot read makes the retrieval's unit unverified with `found` left empty and
the unit's text in the detail, logged once per retrieval; readable units that
disagree are named in `found`.

* fix(compiler): read a facet unit through its ref when the grammar rejects it and cap the disagreeing units

* fix(compiler): name the readable units in found when an unreadable unit sits beside them

* fix(compiler): give a retrieval with no rows no unit finding
…p one plan lineage per plan call (#191)

* fix(harness): cut an over-limit chart once and relay dfo-apis' refusal unchanged

dfo-apis judges a chart write and states the advice in its refusals. The harness checks the select's shape, ranks the members of the one dimension a line chart lists or leaves unbound, and writes the largest 8 once: before the first write when the select is certainly over the limit (`over_series_limit`), or after a refusal on MAX_LINE_CHART_SERIES or MAX_TABLE_ROWS. Every other refusal reaches the model as dfo-apis sent it, with `written_select` when the harness wrote a select other than the model's own. The harness codes series_limit_exceeded, member_kind_unread and no_shared_period are deleted.

presentation-mcp takes a refusal's retry mode from the problem body's `retry_mode` when present. Converted constant-price units carry their deflation basis through the compiler, analysis-mcp and presentation-mcp, and a DAC/IDS constant-price pair refuses at run time with both units and bases in `fix`. The spec-kit pin moves to 45e8bc1.

Refs DevelopmentFinanceObservatory/apis#46.

* fix(harness): cut only where a cut shrinks the chart and stop claiming value order

dfo-apis sorts every select list by member id, so `order_by_value` has no
effect on the drawn order; its description says so and the harness no longer
reorders the select or reads the frame for it. A refused write is cut again
only when the one multiplying dimension holds more members than the chart
draws, never on a facet column, and a bar chart refused on its series
estimate is relayed. The cut limits, distinct-member counts and the cut plan
share one path, and the README describes it.

analysis-mcp names one-sided deflation bases in the mismatch refusal, and
presentation-mcp logs a warning when a refusal states a retry mode it does not
know. The dfo-apis wire-pairs fixture and three runner cassette keys follow
the current dfo-apis wording.

* fix(harness): forward the platform's unit as JSON and read facet units tolerantly

The harness refused a stored card whose `fact_summary.unit` carried a member
its wire model did not know, so every new member dfo-apis added to a unit
broke Studio charts until the harness caught up. `FactSummary.unit` now takes
a unit name or any JSON object and forwards it to Studio unchanged; Studio
reads it as display text. An object over 512 bytes is replaced by the cut
marker when the fact summary is over its cap.

The compiler reads a facet unit's currency, prices and base year strictly and
`deflation_basis` as wording, and ignores other members. A facet unit it
cannot read makes the retrieval's unit unverified with `found` left empty and
the unit's text in the detail, logged once per retrieval; readable units that
disagree are named in `found`.

* fix(harness): report an unreadable stored card as the harness's fault and keep one plan lineage per plan call

A write the platform stored but whose card the harness cannot read was
reported as "the platform refused it" with `chart_not_drawn`. The write
result is now checked once in `mcp_client`, where it is built, and comes back
as a `stored_card_unreadable` refusal naming the component, logged once at
ERROR. make_charts stops after the first such card, a Studio edit says the
change was saved but cannot be shown, and the drain shows a failed step.

A `Refused` stop now carries the spec it refused, and a replan holds a source
for the rest of the plan call when its `unit_unverified` finding states no
readable unit, so a replan no longer pins the source that failed.

Each plan call in a turn starts its own plan lineage. `plan.proposed` gains an
optional `new_plan: true` on the first proposal of a later lineage, so Studio
no longer logs a version regression and overwrites the earlier card.

* fix(harness): log an unreadable card the drain drops and pin the held-source rule on real unit evidence

* fix(compiler): read a facet unit through its ref when the grammar rejects it and cap the disagreeing units

* test(harness): rewrite the switch-after-an-entry planner expectation for the moved schema hashes

* fix(compiler): name the readable units in found when an unreadable unit sits beside them

* fix(harness): hold no source for an empty retrieval, catch the canvas up on a readable no-op head, and say later charts were skipped only when they were

* test(harness): record new_plan on the first draft in the switch-after-an-entry planner expectation

* fix(compiler): give a retrieval with no rows no unit finding

* fix(harness): take the caught-up card's title from the stored head

* refactor(harness): rely on the evidence rule for empty retrievals in the held-source check

* docs(harness): tighten the comments and docstrings around unreadable cards and plan lineage

* fix(common): move the frame store layout to v5 so frames decoded without a deflation basis are fetched again

A frame decoded from apis' text unit column before the decoder kept `recipe` and `deflation_basis` holds donor- and recipient-deflated units as one unit, under the same `(spec_id, content_version)` pair a fresh decode uses. A store hit returns it without a fetch, so unit comparisons depended on cache history. Frames, mint records and anchor records now sit under `v5`; a `v4` object is a miss and the frame is fetched and decoded again.

* fix(harness): forward the platform's unit as JSON and read facet units tolerantly

The harness refused a stored card whose `fact_summary.unit` carried a member
its wire model did not know, so every new member dfo-apis added to a unit
broke Studio charts until the harness caught up. `FactSummary.unit` now takes
a unit name or any JSON object and forwards it to Studio unchanged; Studio
reads it as display text. An object over 512 bytes is replaced by the cut
marker when the fact summary is over its cap.

The compiler reads a facet unit's currency, prices and base year strictly and
`deflation_basis` as wording, and ignores other members. A facet unit it
cannot read makes the retrieval's unit unverified with `found` left empty and
the unit's text in the detail, logged once per retrieval; readable units that
disagree are named in `found`.

* fix(compiler): read a facet unit through its ref when the grammar rejects it and cap the disagreeing units

* fix(compiler): name the readable units in found when an unreadable unit sits beside them

* fix(compiler): give a retrieval with no rows no unit finding
…chart intent's order (#192)

* feat(harness): draw ordered and cut charts largest first through the chart intent's order

`make_charts` writes `order: {by: "value", direction: "descending"}` on every chart cut to its largest members and on every line_chart, bar_chart or table the model marks `order_by_value`. A cut line chart or table also names `order.period`, the period the cut ranked at and the title names. A stat card takes no order, and a bar chart no period.

The Studio editor's recreate keeps the stored card's order where the new type takes one, drops its period for a bar chart or a new window, and accepts an `order` the model copies from the stored intent, which before ended the edit turn.

* fix(harness): let a card edit set or clear its order, and drop a stored order period the new series may not hold

The Studio editor wrote the stored card's order whatever the model drafted, so "sort smallest first" or "remove the sorting" rebuilt the card unchanged. A revise patch may now name `order` (null clears it) and a recreate writes the order it drafts, made legal for its type. A patch or recreate that changes the select or window drops the stored card's `order.period`, which a new series may not hold; a period the model names afresh stands. The `order_by_value` description says a table is ordered only when its rows share one unit.
@jm-rivera
jm-rivera merged commit 75cca84 into main Nov 14, 2026
11 checks passed
@jm-rivera
jm-rivera deleted the release/2026-11-14-2 branch November 14, 2026 11:41

This branch was successfully deployed

1 active deployment
dev — 33f312ac Deployed Nov 14, 2026 by jm-rivera via deploy (dev) / deploy #93
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant