fix(kanban): keep initial blocked tasks sticky

[onfire7777]

Summary

• persist an explicit needs_input block when a task is created as blocked
• preserve sticky behavior for legacy created-blocked rows
• release sticky state on normal unblock or manual promotion while allowing later explicit blocks to re-stick

Verification

• 266 targeted Kanban/state-machine tests passed
• Ruff passed
• manual promotion regression repro returns one ready task
• independent review passed with no remaining findings

This fixes protected gates being incorrectly promoted by DAG recomputation without changing transient circuit-breaker recovery.

[alt-glitch]

This was generated by AI during triage.

Related: #34735 emits a synthetic blocked event and #64830 treats legacy created-blocked payloads as sticky. This PR combines typed creation metadata, legacy handling, and manual-promotion release behavior, so it is a competing lifecycle contract rather than a duplicate.

[onfire7777]

From 2026-07-24 attention report: remains labeled needs-decision + sweeper:risk-session-state. No autonomous action possible. Blocked pending maintainer decision. No changes pushed.

[teknium1]

Thanks for extending the sticky-block lifecycle to creation-time human-ops cards. The current-main premise is real: create_task() creates an initially blocked row at hermes_cli/kanban_db.py:3084-3089, records only created at hermes_cli/kanban_db.py:3172-3190, and recompute_ready() subsequently considers blocked rows at hermes_cli/kanban_db.py:4023-4064.

Problems

• The added initial block_recurrences = 1 conflicts with the existing meaning of that counter. block_task() treats a retained matching kind as a same-cause re-block and escalates at 2 (hermes_cli/kanban_db.py:5571-5618). After an explicit release, the first later needs_input block would therefore enter triage, not re-stick as a normal blocked task.

Suggested changes

• Start creation-time blocked cards at recurrence 0 and add a regression covering release followed by the first same-kind worker block.

Automated hermes-sweeper review.

[teknium1]

Starting this at 1 makes the first later needs_input block after unblock_task() or manual promotion reach the existing recurrence limit of 2 and route to triage, rather than becoming sticky. Initialize it to 0 and cover that release-then-reblock path.

[GottZ]

This was generated by AI during triage.

Summary

Twenty-eight PRs address or reference this complex: most repair the missing create-time blocked lifecycle event that lets recompute_ready() promote an approval gate, while a smaller set broadens authorization, stale-state repair, archival, or unrelated body-update behavior. The focused event-emission approach preserves existing sticky-block and circuit-breaker semantics; broader alternatives introduce distinct review concerns that require splitting or revision.

Related pull requests

• feat(kanban): add initial status for human-ops cards #27526 [closed] related — (+88/-10) — n/a: Introduced initial_status across CLI, database, and tools but did not emit the blocked event required by the sticky guard; closed after its focused feature subset was merged through feat(kanban): add initial-status for human-ops cards (#27526) #28414.
• feat(kanban): add initial-status for human-ops cards (#27526) #28414 [merged] related — (+40/-8) — n/a: Merged reference implementation for initial_status, but its diff emits only created, establishing the missing-event condition behind this issue.
• fix: keep blocked kanban tasks blocked #32581 partial — (+96/-117) — n/a: Stops promotion of every blocked task, covering the symptom but reversing intentional circuit-breaker recovery; the keep_open verdict preserves its regression coverage, not the broad semantic change.
• fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735 best fix — (+123/-0) — n/a: Recorded best-fix for both issues; atomically appends blocked after created and tests event order, repeated recomputation, explicit unblock, and done-parent behavior without changing circuit-breaker recovery.
• fix: keep explicitly blocked Kanban tasks sticky #34756 related — (+190/-10) — n/a: Adds stale sticky-row repair and a claim-boundary guard beyond the root fix, but the contributor review blocks the current form because it undoes supported manual promotion and drops current retry-limit logic.
• fix: keep initial blocked kanban tasks inert #35832 fixes — (+354/-4) — n/a: Contains the same create-time event fix plus archive/worker teardown hardening; contributor review blocks the bundled teardown because Windows may signal an unverified recycled PID.
• fix: keep initially blocked kanban tasks sticky #39085 fixes — (+34/-0) — n/a: Implements the same focused create-time blocked event and one recomputation regression as fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735, but from a substantially older base.
• fix(kanban): keep initial blocked tasks sticky #39832 [closed] partial — (+67/-12) — n/a: Special-cases created(status=blocked) inside _has_sticky_block, bypassing lifecycle-event consumers; closed in favor of fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735 despite its earlier keep_open review.
• fix(kanban): keep initial blocked holds sticky #41497 fixes — (+94/-0) — n/a: Duplicates fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735’s event insertion and adds direct-recompute and dispatcher-dry-run coverage; no changed-line defect was identified, but the production correction is the same.
• fix(kanban): keep initial blocked tasks sticky #43479 [closed] fixes — (+4591/-96) — n/a: Includes the correct create-time event among extensive unrelated gateway, Discord, memory, archival, and repair changes; closed because the bundled work was not ready for public upstream due diligence.
• fix(kanban): preserve initial blocked approval gates #46274 [closed] fixes — (+66/-0) — n/a: Duplicates fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735’s event-based fix with parentless and done-parent regressions; closed but remains a focused reference implementation.
• fix(kanban): keep initial_status=blocked tasks blocked across recompute_ready #46565 fixes — (+26/-0) — n/a: Duplicates fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735 by appending blocked in the creation transaction and testing recomputation, as confirmed by the contributor duplicate ruling.
• fix: keep initial blocked kanban tasks sticky #48253 [closed] partial — (+55/-5) — n/a: Uses created payload inference instead of emitting the established lifecycle event; closed as a duplicate of fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735.
• fix(kanban): emit blocked event for initial_status="blocked" tasks (fixes #47777) #48437 [closed] fixes — (+61/-0) — n/a: Implements the same event insertion with recompute, done-parent, and unblock tests; closed as a duplicate of fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735 despite an APPROVED non-contributor review.
• fix: keep initially blocked kanban cards sticky #60100 [closed] partial — (+45/-0) — n/a: Emits the needed event but records recurrences: 1 without matching persisted state; closed in favor of fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735 after the keep_open review identified that inconsistency.
• fix(kanban): keep initially blocked tasks sticky #61890 [closed] fixes — (+39/-0) — n/a: Duplicates fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735’s atomic event fix with done-parent lifecycle coverage; closed after contributor consolidation.
• fix(kanban): preserve initially blocked human gates #64830 partial — (+37/-12) — n/a: Makes legacy created(status=blocked) rows sticky by changing the predicate, covering the immediate promotion but not emitting the lifecycle event used by notifications and diagnostics.
• fix(kanban): keep initial blocked tasks sticky #66993 fixes — (+89/-0) — n/a: Adds the create-time event plus atomicity and manual-promotion release behavior; contributor triage correctly treats the promotion-event change as competing work rather than a pure duplicate.
• fix(kanban): enforce auditable authorization for human gates #69531 best fix — (+2451/-127) — n/a: Recorded best-fix for [Bug]: Tasks created with --initial-status blocked auto-promote to ready ~1s later with no actor — human approval gate bypassed #39609 and expands the repair into authenticated, fingerprint-bound authorization across claim and spawn paths; the visible diff now includes provider_override and its regression, but the branch still needs current-main reconciliation.
• fix(kanban): keep initial blocked tasks sticky #69679 partial — (+127/-13) — n/a: Adds typed creation metadata, legacy fallback, and manual-promotion release, but contributor review blocks the current recurrence initialization because 1 makes the first later same-kind block escalate instead of re-stick.
• fix(kanban): make creation-time blocks sticky #70733 fixes — (+194/-0) — n/a: Its atomic create-time event addresses the root cause, but the additional annotation branch matches all blocked rows and would make circuit-breaker gave_up rows sticky, contrary to contributor review.
• fix(kanban): keep create-time blocked tasks sticky #71147 partial — (+152/-2) — n/a: Adds the focused event plus a narrowly conditioned idempotent-reuse backfill for legacy blocked rows; its tests require manual placement on current main after the cited test-file pruning.
• Keep initially blocked Kanban tasks sticky #71861 partial — (+39/-0) — n/a: Duplicates fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735’s create-time event and non-promotion regression; the keep_open review identifies only mechanical conflict salvage, not a distinct correction.
• fix(kanban): make --initial-status blocked sticky against recompute_ready #71977 fixes — (+101/-0) — n/a: Duplicates fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735’s event fix while adding dispatcher, unblock, and done-parent regressions; contributor triage identifies fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735 as the canonical implementation.
• fix(kanban): preserve initially blocked task holds #72030 [closed] fixes — (+45/-0) — n/a: Applies the same lifecycle payload as fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735 with narrower regression coverage; closed explicitly as its duplicate.
• [verified] fix(kanban): preserve initial blocked state #74839 [closed] fixes — (+170/-31) — n/a: Combines the duplicate create-time event with independent promote_task() transaction hardening; closed despite the keep_open review after contributor triage required splitting the promotion work.
• fix(kanban): keep initial_status='blocked' tasks sticky #76718 [closed] fixes — (+118/-0) — n/a: Converged to the narrow create-time event fix after removing reviewed block_kind/reclaim conflation, then closed in favor of fix(kanban): emit blocked event when initial_status='blocked' so sticky-block guard fires #34735; it remains a recent minimal reference.
• feat(kanban): add audited update-body command #76765 related — (+203/-0) — n/a: Adds an audited update-body command and does not repair the initial-block promotion cause; contributor review requests stronger hash, length, and no-body-text payload assertions.

Duplicates

#39085, #41497, #46274, #46565, #48437, #61890, #71861, #71977, #72030, and the final narrow form of #76718 implement essentially #34735’s create-time blocked event; #39832, #48253, and #64830 instead infer stickiness from created, while #35832, #66993, #70733, #71147, #74839, and #69679 add separable behavior around the same core repair.

Suggested consolidation

Author action: rebase #69679 onto current main, initialize creation-time block_recurrences to 0, and add the contributor-requested release-then-first-same-kind-reblock regression; do not advance its current diff over that blocking review. Keep #34735 open with the mechanical salvage path recorded by its best-fix verdict, keep #69531 open for its broader recorded best-fix after current-main reconciliation, and close the pure open duplicates #39085, #41497, #46565, #64830, #71861, and #71977 as duplicates of #34735 despite their keep_open reviews because their visible root-cause diffs add no stronger production correction; split salvageable extras from #34756, #35832, #66993, #70733, and #71147, and keep #76765 separate pending its audit/privacy test assertions.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I39609(["issue #39609 (open)"])
    I47777(["issue #47777 (open)"])
    P69679["PR #69679 (open)"]
    P69679 -.->|partial| I39609
    P69679 -.->|partial| I47777
    class I39609 open
    class I47777 open
    class P69679 open
    class P69679 target
    click I39609 "https://github.com/NousResearch/hermes-agent/issues/39609"
    click I47777 "https://github.com/NousResearch/hermes-agent/issues/47777"
    click P69679 "https://github.com/NousResearch/hermes-agent/pull/69679"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 28 pull requests and 2 issues in this complex. Each diff was read against this issue; Assessment working set: 505 kB of PR diffs, 53 kB of issue/PR text, 35 kB of discussion (55 comments), 74 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.