This is the single working list for the active externally-owned Pulse findings in Workflow/upwork, Workflow/linkedin, and Workflow/salesoutreach. It is an audit ledger, not a replacement for the platform issue register: individual implementation tickets must retain one narrow repair boundary.

The source of truth for each row is the workflow's db/db.sqlite record. The snapshots contained 32 external_action_required rows: 15 Upwork, 11 LinkedIn, and 6 Sales Outreach.

Platform ticket	Workflow findings	Why it matches	Next action
PLAT-001	Upwork PUL-7B849CC6	run_full_workflow dropped a keyed human-input override.	Run one bounded override case, then close or reopen against current evidence.
PLAT-033	Upwork PUL-F4468936	Both report placeholder/false changelog evidence after a real mutation.	Inspect one new managed mutation and close if its before/after refs are truthful.
PLAT-200	Upwork PUL-DD9EDE3C	Oversized browser snapshots are now staged into an authorized tool-output path rather than discarded.	Reproduce one oversized guarded snapshot and close if its spill artifact is readable.
PLAT-206	Upwork PUL-565F9ED1; Sales Outreach PUL-4719B06	Exact same terminal-module-result collision: reviewer done blocked the later Fixer's disposition write.	Run one review-plus-fix case with a real disposition. Close both only if it persists the repair.

PUL-13197E02 (Upwork tab acquisition) is confirmed not a duplicate of PLAT-028 — see PLAT-231, which also found both of its cited errors are correct input validation in cdp_tabs.go, not a harness defect; reclassified.

Area	Findings	Current conclusion
Review/Fixer receipt and identity flow	LinkedIn PUL-517B83AD; Sales Outreach PUL-E3F22BCC; plus the PLAT-206 pair above	Related, not one root cause. PUL-E3F22BCC is the old child-session-versus-Pulse-run-ID receipt lookup documented by PLAT-196. PLAT-199's one retained Review+Fix flow is the correct direction. The local removal of the redundant receipt gate still needs deployment and one live test.
Pulse finding identity done	Sales Outreach PUL-1E38F625	PLAT-222 scoped target_key-based fingerprint identity by reporting module for every issue kind except harness_issue (a deliberately shared cross-workflow identity, left unchanged). Runtime reverify remains.
Schedule/session lifecycle	Upwork PUL-1EE39C89; LinkedIn PUL-3565D07C; Sales Outreach PUL-5D2B9495; Sales Outreach PUL-985F2597	Same subsystem, four different boundaries: iteration retention/identity, stale schedule projection, delayed terminal-session detection, and an upgrade turn occupying the scheduled message slot. Do not merge them without one shared reproduction. PLAT-194 is related to terminal reconciliation, not proof that all four are fixed.
Evaluation pipeline	Upwork PUL-E67413EC; LinkedIn PUL-E45BE152; LinkedIn PUL-90D1E2C9	Confirmed three different fixes, not one umbrella: PUL-E67413EC is issue_kind: workflow_issue with no platform blocker named — reclassified to Upwork's own Engineering Review, not this queue. PUL-E45BE152 is done — see below. PUL-90D1E2C9 has no structured detail row (only a run_concerns text line survived), so it was left uninvestigated rather than guessed at.
Artifact/change coverage (done)	LinkedIn PUL-17E6F19A; LinkedIn PUL-7607952E	PLAT-230: confirmed both were already settled independently — the changelog gap by PLAT-205's explicit user decision, the dashboard cursor by the mechanism it depends on (builder/improve.html) having been retired platform-wide before this finding could even be actioned.

Priority	Finding(s)	Required repair boundary
Reclassify	Sales Outreach PUL-985F2597	The cited schedule is pulse_review_only and intentionally has no normal workflow message. Its upgrade preflight and later Pulse lifecycle were separate phases; the run became partial because Pulse failed, not because an upgrade displaced a configured job message.
P0	Sales Outreach PUL-5D2B9495	PLAT-219 records the implemented delayed-closure fix: preserve full-run identity at launch and honor its exact durable failure after bounded inactivity. Runtime reverify remains.
P0	LinkedIn PUL-3565D07C	Keep separate: this 2026-08-06 finding concerns stale list_schedules projection after a successful run, not a missing full-run completion callback after durable failure. Reverify against the newer runtime-reconciliation changes before implementing more code.
P0, partially done	LinkedIn PUL-B995BF46; PUL-3BD9F422	PLAT-221 shipped the managed schema-migration route, but a code review caught that the new tool was registered and tested, yet absent from every real agent's allowlist (workflow-step, todo-task, and Workshop/Pulse paths) — fixed as a PLAT-221 follow-up, now genuinely reachable. PUL-B995BF46 itself turned out already resolved independently (stale finding — real action_outcome_bindings/matched_action_outcome_comparisons data since 2026-08-21, filing never re-verified); reclassify/close that workflow row. PUL-3BD9F422 remains genuinely open (post_approval/image_assets unchanged) — its migration SQL still has not been designed. Do not mark this row fully done until that migration exists and applies.
P1 done	LinkedIn PUL-E45BE152	PLAT-227 added the missing pre_validation editable field to update_evaluation_plan, so route gating and pre-run evidence can be realigned atomically.
Reclassify	Upwork PUL-E67413EC	issue_kind: workflow_issue, no platform tool gap named — belongs to Upwork's own Engineering Review, not this repository's platform queue.
P1 (uninvestigated)	LinkedIn PUL-90D1E2C9	No structured detail row survives for this finding — only a run_concerns text line ("evaluation report aggregator dropped a valid current eval-strategy-loop verdict"). Worth checking as a genuine platform-aggregator bug in a future pass; not started here.
P1 done	Sales Outreach PUL-1E38F625	PLAT-222 shipped the module-scoped fix; see above.
P1	Upwork PUL-D7D173FB	PLAT-218 corrects the diagnosis: HTTP preserved the outer object, but nested approved_scope and post_run_proof used the wrong types and the handler hid that fact behind a false outer-object error. Exact field errors and schema/guidance clarification are implemented.
P1 done	Upwork PUL-EDFF0710	PLAT-223 reworded the injected instruction to fall back to query_workflow_db describe rather than auditing every Folder Guard construction site.
Reclassify	Upwork PUL-9CCE9488	PLAT-224 confirmed this is an external agent-browser binary defect (no vendored source, no in-repo output filtering to build on) — same class as PLAT-215. Not fixable as platform code here; needs an upstream fix or documentation correction outside this repository.
P1	Upwork PUL-A8AB0913; PUL-E717A5E1	PLAT-226 confirmed the exact root cause (a shared token-attribution fallback with no signal to disambiguate Pulse/Builder/orchestrator overhead) and ruled out an existing field as a cheap fix, but stopped short of shipping one — it needs new session-kind signaling across three separately-configured call sites plus a live run to verify against, neither safe to guess at statically.
P1 partially done	Sales Outreach PUL-AAC278EF	PLAT-225 fixed the concrete, reproduced half: the bridge no longer discards a chained command's real stdout/stderr when it reports failure, matching the finding's own stated next_check exactly. Left open, and genuinely fuzzier: whether an ordinary non-2xx HTTP response or expected non-zero subcommand embedded in a script should ever be classified as a tool-level failure at all is a shell-scripting-intent judgment call, not addressed here — a script's own exit code is arguably always the correct signal to trust.
P2 done	LinkedIn PUL-61C84987	PLAT-229 fixed the false-pass at its root: the shared JSONPath value extractor now only unwraps genuine wildcard multi-matches, never a definite path's own array value. Code review follow-up fixed a related gap in the same function: a genuine wildcard multi-match now checks every result for value_type, not just the first.
P2 done	LinkedIn PUL-17E6F19A; PUL-7607952E	See above — PLAT-230.
P2 done	Upwork PUL-B0A88D49	PLAT-228 confirmed the specific instance is stale — the whole "bug-review" guidance kind it referenced was retired and replaced by engineering-review, whose reads are genuinely bundled. The systemic read_skill materialization-time validator this row also asked for remains real, unstarted future work.
P2 done	Upwork PUL-13197E02	PLAT-231 — reclassified, not a harness defect.

Finding	Correct destination
LinkedIn PUL-82F50DFF	External account constraint: the LinkedIn personalized-invitation quota is zero. It needs an account-state display/reopen condition, not a platform-code ticket.
LinkedIn PUL-0394F0B4	User-owned Soul/content correction. Route to an answerable decision or an authorized Soul edit; it is not a runtime defect.
Upwork PUL-4BBFBC4B	Accepted data-source/coverage limitation. Keep a bounded evidence note, not a platform repair.
Upwork PUL-D50AD8BC	Obsolete under the current architecture: normal step-summary concern parsing/step-side filing was intentionally retired. A reviewer files canonical Pulse findings; do not restore the old tool merely to close this row.

Recent main changes are relevant, but they do not close the whole list:

• 22ac69c8a / PLAT-206 directly fixes the Upwork/Sales terminal-result pair.
• bb02c1a1a and af580f01d / PLAT-200 directly address the oversized-snapshot path behind Upwork PUL-DD9EDE3C.
• 34d5ab283 / PLAT-190 reduces load-bearing missing-skill guidance, but Upwork PUL-B0A88D49 still needs a materialization-level check and a live reverify.
• 3fb200020 / PLAT-194 is related to the Sales/LinkedIn schedule state findings, but does not prove them fixed.
• cc98b4598 / PLAT-196 only added receipt-ID diagnostics; it did not fix Sales PUL-E3F22BCC.
• d9223aa61 / PLAT-199 simplified review/fix dispatch. The follow-up removal of the redundant receipt gate is still local until committed and deployed.
• 636fb6e0a is a Pulse backlog migration: it improves queue accounting and platform handoff, but intentionally does not repair the underlying platform defects.

1. Deploy the current Pulse receipt-gate removal, then reverify and close the PLAT-206 pair and Sales PUL-E3F22BCC if the live flow succeeds.
2. Run the four already-covered rechecks (PLAT-001, 033, 200, 206) and close stale workflow rows with evidence.
3. Split the P0/P1 rows above into narrow implementation tickets; do not give one agent all 32 rows as one undifferentiated repair task.
4. Reclassify the four non-platform rows so the platform lane means platform work again.