feat(kanban): reconcile asks the board too — the second copy of the list is gone

[LukasWodka]

Closes tracebloc/backend#1846 · epic tracebloc/backend#1646 · twin of #249

Why the class stayed open after #249

#249 replaced the router's hand-maintained list of deploy-state columns with the board's own order. There were two copies that had to agree, and one of them rotted: the router carried the pre-rename Staging (human review) while the board's column is Staging (agent review), so a card hand-closed there lost its deploy state and kanban-archive.yml hid it (#237).

@saadqbal's point on that PR was that fixing one copy leaves the mechanism intact. This removes the other.

Both readers now ask the board

A deploy state is any column at or after On dev and at or before Prod, taken from $PROJ — which both workflows already fetch, so this costs no extra call.

A column inserted between the anchors (exactly how Staging (agent review) arrived) is classified correctly with no edit in either file, and a renamed intermediate column keeps working because its position is what matters.

Unknown leans the opposite way here, deliberately

The router decides whether to overwrite a deploy state, so unknown → refuse to write.

This loop decides whether to assert Done, so unknown → treat as a deploy state and skip. Leaving a card where it is costs nothing; asserting "nothing shipped" over work that did erases the fact — which is what both halves of #1846 are about.

Evidence

Driven by extracting col_index from the YAML and exercising the real comparison, the same way #249's selftest does:

ok  On dev / Staging (agent review) / FR on staging / Prod  -> SKIP_DEPLOY
ok  Backlog / Code review / Done / Cancelled               -> CONSIDER_DONE
ok  Some Unknown Column                                    -> SKIP_DEPLOY
ok  a NEWLY INSERTED column between the anchors            -> SKIP_DEPLOY

house-rules clean; the router's selftest still passes 15/15; every embedded run: block parses under bash -n.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes how weekly reconcile decides Done vs leaving deploy states—wrong classification could strand or mis-terminalize closed issues—but behavior is heavily guarded by extracted selftests and mirrors the already-shipped router approach.

Overview
kanban-reconcile.yml replaces the hard-coded deploy-state column list with the same board-order rule as the closure router: any Status between On dev and Prod is a deploy state. It publishes tab-separated status_order from the ids step (since $PROJ does not survive shells), implements col_index via awk so misses do not abort under set -euo pipefail, and uses shared classify_column() plus selftest:policy-* markers so closed-completed issues are only eligible for Done when classification is no—unknown / noboard / deploy columns are skipped (opposite lean from the router, which refuses to write Done).

kanban-closure-router.yml refactors the Done overwrite guard to the same classify_column() + policy block instead of inline index comparisons.

CI / tests: kanban-deploy-state-selftest now runs when kanban-reconcile.yml changes. kanban-deploy-state-selftest.py extracts logic from both workflows, asserts byte-identical classify_column, runs both col_index preambles, checks cross-workflow agreement, and exercises each file’s policy verbatim (router errors on noboard, reconcile skips on unknown).

^{Reviewed by Cursor Bugbot for commit 61ac077. Bugbot is set up for automated code reviews on this repo. Configure here.}

[LukasWodka]

Both fixed in 45d4413 — and fixing the first exposed a third.

1. col_index was defined in a different shell

Correct, and it would have broken the workflow outright. It lived in the ids step and was called from plan; every run: block is a new shell, so the function and the $PROJ it read were both gone. The first closed-completed issue would have aborted classify with command not found, and the weekly backstop would have applied no moves at all.

bash -n cannot see this — each block parses perfectly on its own — which is exactly why it shipped. The board order is now published as a step output from the step that has $PROJ, and the helper lives beside its caller. Confirmed: step[3] defines=True uses=True.

2. "No status" is not an unplaceable column

Also correct, and the consequence is the sharper half: those are exactly the cards this backstop exists to terminalize, so they would have stayed closed, non-terminal and unarchived forever.

The distinction is now explicit — the absence of a column is evidence nothing deployed; an unrecognised column is not. Only the second is treated as a deploy state.

3. The fix for (1) carried the set -e trap

Worth reporting rather than quietly correcting. My first version used grep -nxF | head | cut, which fails the pipeline when the column is not found — and _ci=$(col_index ...) is a bare assignment, so an unrecognised column aborted the classify step instead of being handled as unplaceable. Verified directly: the line after the assignment never ran.

That is release-train#73's trap, which I documented at length, written again here. A helper whose job is to report "not found" must not do it by failing. awk always exits 0 and prints -1.

Evidence

Driven end to end against a fixture board — 13 cases, all passing:

On dev / Staging (agent review) / FR on staging / Ready for prod / Prod  -> SKIP_DEPLOY
Backlog / Code review / Done / Cancelled                                -> CONSIDER_DONE
No status / <empty>                                                     -> CONSIDER_DONE
Some Unknown Column                                                     -> SKIP_DEPLOY
a NEWLY INSERTED column between the anchors                             -> SKIP_DEPLOY

house-rules clean; the router's selftest still 15/15; every embedded run: block parses.

bugbot run

[LukasWodka]

Fixed in 8b616ec — and this one is the sharpest finding on the PR, because it caught the change reintroducing its own bug one level up.

This PR created a second col_index, and the selftest extracted only the router's — with a paths: filter that omitted kanban-reconcile.yml entirely. A PR touching only the new copy would never have run the check. The guard could stay green while the thing it guards changed underneath it, which is exactly the two-copy rot the change exists to close.

Two copies are forced, so the guarantee had to change rather than the count

kanban-closure-router.yml is a reusable workflow that runs in the caller's checkout, so a shared script in this repo is not on disk for it. kanban-reconcile.yml runs here and works from step outputs rather than $PROJ. They cannot share code, and their inputs differ — jq over a project document, awk over a tab-separated order.

So the selftest now extracts both, runs every case through each, and requires them to agree:

DISAGREE: router=DEPLOY_STATE, reconcile=ANCHORS_MISSING

Verified by breaking only the reconcile copy — every case turns into that line. Restored: 15/15 across both.

And the paths: filter now lists kanban-reconcile.yml on both triggers, so the check actually fires when the copy it now covers changes.

Running total on this PR, since it is worth being straight about

#	finding	mine
1	col_index defined in a different shell — would have broken the workflow outright	yes
2	No status skipped as unplaceable — the cards this backstop exists to terminalize	yes
3	the fix for (1) carried the set -e trap from release-train#73	yes
4	the selftest could not see the copy this PR adds	yes

Four rounds, four real defects, none visible in the diff. ruff and house-rules clean.

bugbot run

[LukasWodka]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 61ac077. Configure here.}