beads/internal/validation/issue.go
Athos Bernardes df90ec784c
fix(bd): canonicalize actor==assignee comparisons in unclaim.go and AuthorizeAssigneeTransferWithPools (#5439)
* fix(bd): canonicalize actor==assignee comparisons in unclaim.go and AuthorizeAssigneeTransferWithPools

Follow-up to ga-wzl83, which canonicalized the two pure-Go validators in
internal/validation/issue.go but explicitly left out two higher-risk sites.

UnclaimIssueInTx / UnclaimIssueIfAssigneeInTx compared actor/expectedAssignee
to the stored assignee verbatim, both in a Go-side precheck AND in the SQL
UPDATE's WHERE clause -- so canonicalizing only the Go side would not have
fixed `bd unclaim` end-to-end. Replaced the `assignee = ?` SQL predicate with
a `row_lock = ?` CAS against the RowVersion read moments earlier: every
ownership/status/started_at mutation always rewrites row_lock (the documented
freshRowLock invariant in lease.go), so this detects the exact same race the
old predicate did -- and more (status/started_at races too) -- without
embedding a spelling-sensitive string comparison in SQL.

AuthorizeAssigneeTransferWithPools (the single fence both the DBTX and
unit-of-work backends evaluate, per its own doc comment) compared
before.Assignee to request.Actor and to request.Patch.Assignee.Value
verbatim, so a caller naming the current holder under a different layer's
identity spelling was wrongly fenced as a stranger.

Both fixes reuse the canonicalActor/actorMatches shape from ga-wzl83, kept as
a small package-local duplicate in issueops (internal/storage/issueops
importing internal/validation would invert the layering).

Verified: new cross-spelling cases in aggregate_authorize_transfer_test.go
and new end-to-end cases in unclaim_embedded_test.go /
unclaim_conditional_embedded_test.go (against a real embedded Dolt backend)
all fail with the exact bug-report symptom against the pre-fix code and pass
after the fix, with no regressions in the existing suites.

ga-5ksp5

* fix(bd): close the sibling-CAS disagreement gate review flagged on #5439

Review verdict on #5439 was "MERGE AFTER ONE FIX": after this branch's
unclaim.go/aggregate.go fix, bd unclaim --if-assignee succeeds under a
respelled identity while bd update --expected-assignee on the same bead
still fails ErrAssigneeMismatch — the same verbatim actor==assignee defect
(ga-wzl83), left in two more places.

- internal/storage/issueops/update_cas.go:46 (CheckExpectedFieldsInTx) now
  compares under actorMatches (package-local, unexported — same package as
  identity.go, no new import needed).
- internal/storage/domain/issue.go:643 (ApplyUpdate, the unit-of-work twin)
  now compares under validation.ActorMatches. domain cannot import
  internal/storage/issueops (issueops already imports domain in several
  files — the reverse would cycle), but domain -> internal/validation is
  safe (validation depends only on internal/types), so this reuses the
  already-shipped validation copy instead of adding a third duplicate.

Two non-blocking asks from the same review, addressed:

- Exported internal/validation's canonicalActor/actorMatches to
  CanonicalActor/ActorMatches so domain/issue.go could reuse them (the
  reviewer's own point: "an issueops -> validation import creates no
  cycle... layering inversion is a style call, not a compile fact" applies
  equally to domain -> validation). Also fixed a garbled mid-sentence doc
  comment on ActorMatches while touching the file (separate nit from the
  same review, on #5438).
- Added a cross-package parity test
  (internal/storage/issueops/identity_parity_test.go) that runs the
  existing CanonicalActor/ActorMatches case tables through both this
  package's local duplicate and validation's exported version, so the two
  copies can't silently drift. Mutation-tested: dropped the hyphen
  separator from issueops's copy, confirmed the two hyphen-related cases
  correctly fail, reverted.

New conformance test (backend/conformance, wired into all three backend
legs) proves the fix end-to-end: an ExpectedAssignee guard now accepts a
different spelling of the current holder and still refuses a genuinely
different identity. Added as a SEPARATE function rather than extending the
existing RunIssueOperationsUpdateConditionalGuardsGateOrdinaryEdits — that
test's later "order-dependent composition" arm depends on the row's
assignee staying exactly "holder" through to the end, which a mid-test
respelling would have broken.

Verification: internal/validation, internal/storage/issueops,
internal/storage/domain(+db,fs,git) all green. New conformance test passes
on embeddeddolt and uow (dolt backend skips locally, no Docker — matches
the existing documented limitation, CI covers it). Broader regression
subset (Update/Claim/Unclaim, 28 embeddeddolt + 22 uow cases) green, no
regressions. go build/vet/gofmt clean.

Filed ga-4ymii for the reviewer's second (also non-blocking) ask: the
remaining verbatim sites at issueops/claim.go (SQL CAS + idempotent
re-claim), its domain/db and domain/issue.go uow twins, and
issueops/lease.go's heartbeat holder check — out of scope for this PR per
the reviewer's own framing ("please file the follow-up").

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* fix(bd): dispatch the respelled-assignee contract case onto the role seam

TestRoleContractCasesCoverEveryContractCase (main, post-dating this branch's
fork point) caught a real gap: the ga-5ksp5 contract case added to this
branch's earlier commit (RunIssueOperationsUpdateConditionalGuardAccepts-
RespelledAssignee) had no roleContractCases dispatch row, so an external
backend running RunRoleContracts would never see it.

The reason isn't a missed row so much as the whole RunIssueOperationsUpdate*
family this case was written beside having been renamed and moved to
RunLifecycleUpdate* in backend/conformance/lifecycle_update_contract.go by
main's "de-stage the hard Lifecycle contracts onto the role seam" refactor,
which landed on main after this branch forked. Ported the case to match:
same logic, adapted to the new fixture's idioms (fixture.Lifecycle.Update
instead of fixture.Operations.Update, lifecycleUpdateRow(...).Priority
instead of a raw SQL assertion, storage.ErrAssigneeMismatch instead of
publicops.ErrAssigneeMismatch, seedLifecycleUpdateIssue instead of
seedClosePolicyIssue) — not a mechanical copy. Also wired it into the three
per-backend TestLifecycleUpdateContract dispatch lists (dolt, embeddeddolt,
uow) beside its ConditionalGuardsGateOrdinaryEdits sibling, so it actually
runs against real backends rather than only satisfying the completeness
guard.

Verified: TestRoleContractCasesCoverEveryContractCase passes; the ported
case passes against embeddeddolt (0.35s) and uow (0.17s) for real, not just
compiles (dolt backend test skips in this sandbox — no Docker); full
backend/conformance suite and repo build clean.

---------

Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
2026-08-09 11:35:01 -07:00

307 lines
11 KiB
Go

package validation
import (
"fmt"
"slices"
"strings"
"github.com/steveyegge/beads/internal/types"
)
// IssueValidator validates an issue and returns an error if validation fails.
// Validators can be composed using Chain() for complex validation logic.
type IssueValidator func(id string, issue *types.Issue) error
// Chain composes multiple validators into a single validator.
// Validators are executed in order and the first error stops the chain.
func Chain(validators ...IssueValidator) IssueValidator {
return func(id string, issue *types.Issue) error {
for _, v := range validators {
if err := v(id, issue); err != nil {
return err
}
}
return nil
}
}
// Exists validates that an issue is not nil.
func Exists() IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil {
return fmt.Errorf("issue %s not found", id)
}
return nil
}
}
// NotTemplate validates that an issue is not a template.
// Templates are read-only and cannot be modified.
func NotTemplate() IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil {
return nil // Let Exists() handle nil check if needed
}
if issue.IsTemplate {
return fmt.Errorf("cannot modify template %s: templates are read-only; use 'bd mol pour' to create a work item", id)
}
return nil
}
}
// NotPinned validates that an issue is not pinned, by either of the two ways
// bd expresses a pin: the Pinned boolean column or the pinned status. Returns
// an error if either trigger fires, unless force is true.
//
// Both triggers are load-bearing because consumers disagree on which one they
// write and read (ga-z3vht). Gas Town pins by status — it enumerates pins with
// List(ListOptions{Status: StatusPinned}) across 21 sites (hook_check.go,
// prime_output.go, molecule_step.go, up.go) — while Gas City reads the boolean
// (compute_awake_set.go). Checking only one strips the other consumer's
// protection outright, so do not "simplify" either clause away.
func NotPinned(force bool) IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil {
return nil // Let Exists() handle nil check if needed
}
if !force && (issue.Pinned || issue.Status == types.StatusPinned) {
return fmt.Errorf("cannot modify pinned issue %s (use --force to override)", id)
}
return nil
}
}
// CanonicalActor normalizes an identity string so two spellings of the same
// Gas Town identity compare equal. The same identity arrives at bd in more
// than one spelling depending on which layer produced the string it was
// handed: a dotted alias like "gastown.mayor" gets its dot replaced wherever
// a dot is unsafe for that context — "__" in a session name, "_" in a Dolt
// table/database name, "-" elsewhere — and bd only ever sees the resulting
// string, never the substitution itself (ga-wzl83). None of ".", "_", "-"
// carries meaning in an identity string: each is always a positional
// separator between a rig and a role/agent name, never part of either name.
// Collapsing any run of them to one canonical separator lets two spellings
// of the same identity compare equal without weakening comparisons between
// genuinely different identities, whose non-separator characters still
// differ (e.g. "gastown.mayor" vs "gastown.dog-3" stay distinct).
//
// Empty stays empty: an actual absence of an actor must never canonicalize
// to the same value as a non-empty one, or AssigneeMatches would start
// accepting an empty actor against an assigned issue.
func CanonicalActor(s string) string {
if s == "" {
return ""
}
var b strings.Builder
b.Grow(len(s))
inSeparator := false
for _, r := range s {
switch r {
case '.', '_', '-':
if !inSeparator {
b.WriteByte('_')
inSeparator = true
}
default:
b.WriteRune(r)
inSeparator = false
}
}
return b.String()
}
// ActorMatches reports whether actor has the same authority as assignee:
// either they're byte-identical, or they canonicalize to the same identity
// (see CanonicalActor). The byte-identical check is not redundant — it
// avoids paying canonicalization on the overwhelmingly common exact-match
// path (ga-5ksp5 gate review, #5438: fixes a garbled doc comment, no
// behavior change).
func ActorMatches(assignee, actor string) bool {
return assignee == actor || CanonicalActor(assignee) == CanonicalActor(actor)
}
// AssigneeMatches validates that the actor has authority to close the issue.
// Authority means the issue is unassigned or the actor matches the current
// assignee. Returns an error on mismatch unless force is true.
//
// Authority is identity-by-string: actor is compared to assignee under
// CanonicalActor, so two principals sharing one canonical actor name both
// pass — including the same principal spelled two different ways by two
// different layers of Gas Town (ga-wzl83). bd has no identity layer, so this
// matches the existing semantics of the actor field — the guard removes
// silent cross-actor closes without adding new identity guarantees.
//
// This guards against the silent-success bug where actor A closes a bead that
// was concurrently re-claimed by actor B: storage accepts the close (id-only
// WHERE clause), so without this check bd reports "✓ Closed" even though A had
// no authority over the bead. (be-035)
func AssigneeMatches(actor string, force bool) IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil || force {
return nil
}
if issue.Assignee == "" || ActorMatches(issue.Assignee, actor) {
return nil
}
return fmt.Errorf("cannot close %s: assignee is %q, actor is %q; reclaim or use --force to override", id, issue.Assignee, actor)
}
}
// AssigneeNotStolen validates that a plain assignee update does not silently
// overwrite another actor's live claim (bd-98s5c). newAssignee is the value
// the write would set (may be "" for an unassign — fenced the same way, since
// stripping a live claim is what bd unclaim refuses without --force).
//
// The refusal fires only when every clause holds; each one is load-bearing:
// - Assignee != "" — unassigned issues are freely
// assignable.
// - Assignee doesn't canonicalize to actor — an actor editing its
// own claim is untouched, under any spelling of its own identity.
// - Assignee doesn't canonicalize to newAssignee — an idempotent re-assert
// of the current holder stays a success (retry/replay safety on the
// proxied path), even when the re-assert names the holder under a
// different spelling.
// - Status == in_progress — reassigning an OPEN bead stays frictionless:
// dispatchers hand-dole open beads and pool-queue takes happen constantly,
// which is why this cannot reuse the unclaim/close ownership rule verbatim.
// - holder is not a claim.pools alias — --claim deliberately treats
// pool-assigned issues as takeable by any actor; without the same
// carve-out here, --claim and -a would give opposite answers on the same
// bead. poolAliases is a thunk so call sites don't pay the config read on
// the overwhelmingly common non-conflicting paths.
//
// Like AssigneeMatches, authority is identity-by-string — compared under
// CanonicalActor rather than verbatim (ga-wzl83), so this removes silent
// cross-actor takeovers without adding new identity guarantees, and without
// falsely treating the current holder as a stranger when actor or
// newAssignee names it under a different layer's spelling of the same
// identity.
func AssigneeNotStolen(actor, newAssignee string, poolAliases func() []string, force bool) IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil || force {
return nil
}
if issue.Assignee == "" || ActorMatches(issue.Assignee, actor) || ActorMatches(issue.Assignee, newAssignee) {
return nil
}
if issue.Status != types.StatusInProgress {
return nil
}
if poolAliases != nil && slices.Contains(poolAliases(), issue.Assignee) {
return nil
}
return fmt.Errorf("cannot reassign %s: held by %q (in_progress); coordinate with the holder (bd mail %s) — pass --force only if their claim is abandoned (crashed agent, expired lease), or use bd reclaim", id, issue.Assignee, issue.Assignee)
}
}
// NotClosed validates that an issue is not already closed.
func NotClosed() IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil {
return nil
}
if issue.Status == types.StatusClosed {
return fmt.Errorf("issue %s is already closed", id)
}
return nil
}
}
// NotHooked validates that an issue is not in hooked status.
func NotHooked(force bool) IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil {
return nil
}
if !force && issue.Status == types.StatusHooked {
return fmt.Errorf("cannot modify hooked issue %s (use --force to override)", id)
}
return nil
}
}
// HasStatus validates that an issue has one of the allowed statuses.
func HasStatus(allowed ...types.Status) IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil {
return nil
}
for _, status := range allowed {
if issue.Status == status {
return nil
}
}
return fmt.Errorf("issue %s has status %s, expected one of: %v", id, issue.Status, allowed)
}
}
// HasType validates that an issue has one of the allowed types.
func HasType(allowed ...types.IssueType) IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil {
return nil
}
for _, t := range allowed {
if issue.IssueType == t {
return nil
}
}
return fmt.Errorf("issue %s has type %s, expected one of: %v", id, issue.IssueType, allowed)
}
}
// EpicHasOpenChildren validates that an epic does not have open children.
// If the issue is an epic with open children, returns an error unless force is true.
// Non-epic issues pass through without validation.
func EpicHasOpenChildren(force bool, openChildCount int) IssueValidator {
return func(id string, issue *types.Issue) error {
if issue == nil || force {
return nil
}
if issue.IssueType != types.TypeEpic {
return nil
}
if openChildCount > 0 {
return fmt.Errorf("epic %s has %d open child issue(s); close children first or use --force to override", id, openChildCount)
}
return nil
}
}
// forUpdate returns a validator chain for update operations.
// Validates: issue exists and is not a template.
func forUpdate() IssueValidator {
return Chain(
Exists(),
NotTemplate(),
)
}
// forClose returns a validator chain for close operations.
// Validates: issue exists, is not a template, and is not pinned (unless force).
func forClose(force bool) IssueValidator {
return Chain(
Exists(),
NotTemplate(),
NotPinned(force),
)
}
// forDelete returns a validator chain for delete operations.
// Validates: issue exists and is not a template.
func forDelete() IssueValidator {
return Chain(
Exists(),
NotTemplate(),
)
}
// forReopen returns a validator chain for reopen operations.
// Validates: issue exists, is not a template, and is closed.
func forReopen() IssueValidator {
return Chain(
Exists(),
NotTemplate(),
HasStatus(types.StatusClosed),
)
}